Lecture Notes in Financial Econometrics (MSc

course)

Paul Söderlind1

13 June 2013

1 Universityof St. Gallen. Address: s/bf-HSG, Rosenbergstrasse 52, CH-9000 St. Gallen,
Switzerland. E-mail: Paul.Soderlind@unisg.ch. Document name: FinEcmtAll.TeX
Contents

1 Review of Statistics 5
1.1 Random Variables and Distributions . . . . . . . . . . . . . . . . . . 5
1.2 Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Distributions Commonly Used in Tests . . . . . . . . . . . . . . . . . 14
1.4 Normal Distribution of the Sample Mean as an Approximation . . . . 17

A Statistical Tables 19

2 Least Squares Estimation 22

2.1 Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3 Heteroskedasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.4 Autocorrelation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

A A Primer in Matrix Algebra 59

A Statistical Tables 64

3 Regression Diagnostics 67
3.1 Misspecifying the Set of Regressors . . . . . . . . . . . . . . . . . . 67
3.2 Comparing Non-Nested Models . . . . . . . . . . . . . . . . . . . . 68
3.3 Non-Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.4 Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5 Estimation on Subsamples . . . . . . . . . . . . . . . . . . . . . . . 69
3.6 Robust Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

1
4 Asymptotic Results on OLS 80
4.1 Properties of the OLS Estimator when “Gauss-Markov” Is False . . . 80
4.2 Motivation of Asymptotics . . . . . . . . . . . . . . . . . . . . . . . 80
4.3 Asymptotics: Consistency . . . . . . . . . . . . . . . . . . . . . . . 80
4.4 When LS Cannot be Saved . . . . . . . . . . . . . . . . . . . . . . . 82
4.5 Asymptotic Normality . . . . . . . . . . . . . . . . . . . . . . . . . 87

5 Index Models 89
5.1 The Inputs to a MV Analysis . . . . . . . . . . . . . . . . . . . . . . 89
5.2 Single-Index Models . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3 Estimating Beta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.4 Multi-Index Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.5 Principal Component Analysis . . . . . . . . . . . . . . . . . . . . 100
5.6 Estimating Expected Returns . . . . . . . . . . . . . . . . . . . . . . 104

6 Testing CAPM and Multifactor Models 106

6.1 Market Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.2 Calendar Time and Cross Sectional Regression . . . . . . . . . . . . 117
6.3 Several Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.4 Fama-MacBeth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

A Statistical Tables 124

7 Time Series Analysis 127

7.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.2 Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.3 White Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.4 Autoregression (AR) . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.5 Moving Average (MA) . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.6 ARMA(p,q) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.7 VAR(p) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.8 Impulse Response Function . . . . . . . . . . . . . . . . . . . . . . . 142
7.9 Non-stationary Processes . . . . . . . . . . . . . . . . . . . . . . . . 144

2
8 Predicting Asset Returns 155
8.1 Autocorrelations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.2 Other Predictors and Methods . . . . . . . . . . . . . . . . . . . . . 163
8.3 Out-of-Sample Forecasting Performance . . . . . . . . . . . . . . . . 166
8.4 Security Analysts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

9 Maximum Likelihood Estimation 185

9.1 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . 185
9.2 Key Properties of MLE . . . . . . . . . . . . . . . . . . . . . . . . . 191
9.3 Three Test Principles . . . . . . . . . . . . . . . . . . . . . . . . . . 192
9.4 QMLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

10 ARCH and GARCH 194

10.1 Heteroskedasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
10.2 ARCH Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
10.3 GARCH Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
10.4 Non-Linear Extensions . . . . . . . . . . . . . . . . . . . . . . . . . 206
10.5 (G)ARCH-M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
10.6 Multivariate (G)ARCH . . . . . . . . . . . . . . . . . . . . . . . . . 209

11 Risk Measures 214

11.1 Value at Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
11.2 Expected Shortfall . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
11.3 Target Semivariance (Lower Partial 2nd Moment) and Max Drawdown 223

12 Return Distributions (Univariate) 229

12.1 Estimating and Testing Distributions . . . . . . . . . . . . . . . . . . 229
12.2 Tail Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242

13 Return Distributions (Multivariate) 252

13.1 Recap of Univariate Distributions . . . . . . . . . . . . . . . . . . . 252
13.2 Exceedance Correlations . . . . . . . . . . . . . . . . . . . . . . . . 252
13.3 Beyond (Linear) Correlations . . . . . . . . . . . . . . . . . . . . . . 254
13.4 Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
13.5 Joint Tail Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 267

3
14 Option Pricing and Estimation of Continuous Time Processes 274
14.1 The Black-Scholes Model . . . . . . . . . . . . . . . . . . . . . . . . 274
14.2 Estimation of the Volatility of a Random Walk Process . . . . . . . . 282

15 Event Studies 289

15.1 Basic Structure of Event Studies . . . . . . . . . . . . . . . . . . . . 289
15.2 Models of Normal Returns . . . . . . . . . . . . . . . . . . . . . . . 291
15.3 Testing the Abnormal Return . . . . . . . . . . . . . . . . . . . . . . 295
15.4 Quantitative Events . . . . . . . . . . . . . . . . . . . . . . . . . . . 297

16 Kernel Density Estimation and Regression 299

16.1 Non-Parametric Regression . . . . . . . . . . . . . . . . . . . . . . . 299
16.2 Examples of Non-Parametric Estimation . . . . . . . . . . . . . . . . 307

17 Simulating the Finite Sample Properties 312

17.1 Monte Carlo Simulations . . . . . . . . . . . . . . . . . . . . . . . . 313
17.2 Bootstrapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317

18 Panel Data 322

18.1 Introduction to Panel Data . . . . . . . . . . . . . . . . . . . . . . . 322
18.2 Fixed Effects Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
18.3 Random Effects Model . . . . . . . . . . . . . . . . . . . . . . . . . 326

19 Binary Choice Models 329

19.1 Binary Choice Model . . . . . . . . . . . . . . . . . . . . . . . . . . 329
19.2 Truncated Regression Model . . . . . . . . . . . . . . . . . . . . . . 336
19.3 Censored Regression Model (Tobit Model) . . . . . . . . . . . . . . . 340
19.4 Heckit: Sample Selection Model . . . . . . . . . . . . . . . . . . . . 343

4
1 Review of Statistics
More advanced material is denoted by a star ( ). It is not required reading.

1.1 Random Variables and Distributions

1.1.1 Distributions

A univariate distribution of a random variable x describes the probability of different

values. If f .x/ is the probability density function, then the probability that x is between
A and B is calculated as the area under the density function from A to B
Z B
Pr .A  x < B/ D f .x/dx: (1.1)
A

See Figure 1.1 for illustrations of normal (gaussian) distributions.

Remark 1.1 If x  N.;  2 /, then the probability density function is

1 1 2
. x  / :
f .x/ D p e 2

2 2
This is a bell-shaped curve centered on the mean  and where the standard deviation 
determines the “width” of the curve.

A bivariate distribution of the random variables x and y contains the same information
as the two respective univariate distributions, but also information on how x and y are
related. Let h .x; y/ be the joint density function, then the probability that x is between
A and B and y is between C and D is calculated as the volume under the surface of the
density function
Z B Z D
Pr .A  x < B and C  y < D/ D h.x; y/dxdy: (1.2)
A C

5
 N(0,2) distribution Normal distributions: different means
0.4 0.4
Pr(−2 < x ≤ −1) = 16% N(0,2)
0.3 Pr(0 < x ≤ 1) = 26% 0.3 N(1,2)

0.2 0.2

0.1 0.1

0 0
−4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4
x x

Normal distributions: different variances

0.4
N(0,2)
0.3 N(0,1)

0.2

0.1

0
−4 −3 −2 −1 0 1 2 3 4
x

Figure 1.1: A few different normal distributions

A joint normal distributions is completely described by the means and the covariance
matrix " # " # " #!
x x x2 xy
N ; ; (1.3)
y y xy y2
where x and y denote means of x and y, x2 and y2 denote the variances of x and
y and xy denotes their covariance. Some alternative notations are used: E x for the
mean, Std.x/ for the standard deviation, Var.x/ for the variance and Cov.x; y/ for the
covariance.
Clearly, if the covariance xy is zero, then the variables are (linearly) unrelated to each
other. Otherwise, information about x can help us to make a better guess of y. See Figure
1.2 for an example. The correlation of x and y is defined as
xy
xy D : (1.4)
x y

6
 If two random variables happen to be independent of each other, then the joint density
function is just the product of the two univariate densities (here denoted f .x/ and k.y/)

h.x; y/ D f .x/ k .y/ if x and y are independent. (1.5)

This is useful in many cases, for instance, when we construct likelihood functions for
maximum likelihood estimation.

N(0,1) distribution Bivariate normal distribution, corr = 0.1

0.4 0.2

0.1
0.2
2
0
2 0
0 0 x
−2 −1 0 1 2 y −2 −2
x

Bivariate normal distribution, corr = 0.8

0.2

0.1
2
0
2 0
0 x
y −2 −2

Figure 1.2: Density functions of univariate and bivariate normal distributions

1.1.2 Conditional Distributions

If h .x; y/ is the joint density function and f .x/ the (marginal) density function of x, then
the conditional density function is

g.yjx/ D h.x; y/=f .x/: (1.6)

7
For the bivariate normal distribution (1.3) we have the distribution of y conditional on a
given value of x as
 
xy 2 xy xy
yjx  N y C 2 .x x / ; y : (1.7)
x x2

Notice that the conditional mean can be interpreted as the best guess of y given that we
know x. Similarly, the conditional variance can be interpreted as the variance of the
forecast error (using the conditional mean as the forecast). The conditional and marginal
distribution coincide if y is uncorrelated with x. (This follows directly from combining
(1.5) and (1.6)). Otherwise, the mean of the conditional distribution depends on x, and
the variance is smaller than in the marginal distribution (we have more information). See
Figure 1.3 for an illustration.

Bivariate normal distribution, corr = 0.1 Conditional distribution of y, corr = 0.1

0.8
x = −0.8
0.6 x=0
0.2
0.4
0.1
2 0.2
0
2 0
0 x 0
y −2 −2 −2 −1 0 1 2
y

Bivariate normal distribution, corr = 0.8 Conditional distribution of y, corr = 0.8

0.8
x = −0.8
0.6 x=0
0.2
0.4
0.1
2 0.2
0
2 0
0 x 0
y −2 −2 −2 −1 0 1 2
y

Figure 1.3: Density functions of normal distributions

8
1.1.3 Illustrating a Distribution

If we know the type of distribution (uniform, normal, etc) a variable has, then the best way
of illustrating the distribution is to estimate its parameters (mean, variance and whatever
more—see below) and then draw the density function.
In case we are not sure about which distribution to use, the first step is typically to draw
a histogram: it shows the relative frequencies for different bins (intervals). For instance, it
could show the relative frequencies of a variable x t being in each of the follow intervals:
-0.5 to 0, 0 to 0.5 and 0.5 to 1.0. Clearly, the relative frequencies should sum to unity (or
100%), but they are sometimes normalized so the area under the histogram has an area of
unity (as a distribution has).
See Figure 1.4 for an illustration.

Histogram of small growth stocks Histogram of large value stocks

0.1 0.1
mean, std: mean, std:
0.28 7.99 0.61 5.02
skew, kurt, BJ: skew, kurt, BJ:
0.0 5.1 124.5 -0.2 4.0 34.2
0.05 0.05

0 0
−20 −10 0 10 20 −20 −10 0 10 20
Monthly excess return, % Monthly excess return, %

Monthly data on two U.S. indices, 1957:1-2012:12

Sample size: 672

Figure 1.4: Histogram of returns, the curve is a normal distribution with the same mean
and standard deviation as the return series

1.1.4 Confidence Bands and t-tests

Confidence bands are typically only used for symmetric distributions. For instance, a 90%
confidence band is constructed by finding a critical value c such that

Pr . c  x <  C c/ D 0:9: (1.8)

9
Replace 0.9 by 0.95 to get a 95% confidence band—and similarly for other levels. In
particular, if x  N.;  2 /, then

Pr . 1:65  x <  C 1:65 / D 0:9 and

Pr . 1:96  x <  C 1:96/ D 0:95: (1.9)

As an example, suppose x is not a data series but a regression coefficient (denoted

O
ˇ)—and we know that the standard error equals some number  . We could then construct
a 90% confidence band around the point estimate as

ŒˇO 1:65; ˇO C 1:65: (1.10)

In case this band does not include zero, then we would be 90% that the (true) regression
coefficient is different from zero.
Alternatively, suppose we instead construct the 90% confidence band around zero as

Œ0 1:65; 0 C 1:65: (1.11)

If this band does not include the point estimate (ˇ),O then we are also 90% sure that the
(true) regression coefficient is different from zero. This latter approach is virtually the
same as doing a t-test, that, by checking if
ˇ ˇO 0 ˇ
ˇ ˇ
ˇ > 1:65: (1.12)
ˇ ˇ
ˇ  ˇ
ˇ

To see that, notice that if (1.12) holds, then

ˇO < 1:65 or ˇO > 1:65; (1.13)

which is the same as ˇO being outside the confidence band in (1.11).

10
1.2 Moments

1.2.1 Mean and Standard Deviation

The mean and variance of a series are estimated as

PT PT
xN D t D1 x t =T and 
O2 D t D1 .x t N 2 =T:
x/ (1.14)

The standard deviation (here denoted Std.x t /), the square root of the variance, is the most
common measure of volatility. (Sometimes we use T 1 in the denominator of the sample
variance instead T .) See Figure 1.4 for an illustration.
A sample mean is normally distributed if x t is normal distributed, x t  N.;  2 /. The
basic reason is that a linear combination of normally distributed variables is also normally
distributed. However, a sample average is typically approximately normally distributed
even if the variable is not (discussed below). If x t is iid (independently and identically
distributed), then the variance of a sample mean is

Var.x/
N D  2 =T , if x t is iid. (1.15)

A sample average is (typically) unbiased, that is, the expected value of the sample
average equals the population mean, that is,

E xN D E x t D : (1.16)

Since sample averages are typically normally distributed in large samples (according to
the central limit theorem), we thus have

xN  N.;  2 =T /; (1.17)

so we can construct a t-stat as

xN
tD p ; (1.18)
= T
which has an N.0; 1/ distribution.

11
 Proof. (of (1.15)–(1.16)) To prove (1.15), notice that
P 
T
Var.x/
N D Var tD1 x t =T
PT
D t D1 Var .x t =T /
D T Var .x t / =T 2
D  2 =T:

The first equality is just a definition and the second equality follows from the assumption
that x t and xs are independently distributed. This means, for instance, that Var.x2 C
x3 / D Var.x2 / C Var.x3 / since the covariance is zero. The third equality follows from
the assumption that x t and xs are identically distributed (so their variances are the same).
The fourth equality is a trivial simplification.
To prove (1.16)

E xN D E TtD1 x t =T
P

D TtD1 E x t =T
P

D E xt :

The first equality is just a definition and the second equality is always true (the expectation
of a sum is the sum of expectations), and the third equality follows from the assumption
of identical distributions which implies identical expectations.

1.2.2 Skewness and Kurtosis

The skewness, kurtosis and Bera-Jarque test for normality are useful diagnostic tools.
They are

Test statistic Distribution

3
skewness D T1 TtD1 xt 
P
N .0; 6=T /

4 (1.19)
kurtosis D T1 TtD1 xt 
P
N .3; 24=T /
Bera-Jarque D T6 skewness2 C 24
T
.kurtosis 3/2 22 :

This is implemented by using the estimated mean and standard deviation. The distribu-
tions stated on the right hand side of (1.19) are under the null hypothesis that x t is iid
N.;  2 /. The “excess kurtosis” is defined as the kurtosis minus 3. The test statistic for

12
the normality test (Bera-Jarque) can be compared with 4.6 or 6.0, which are the 10% and
5% critical values of a 22 distribution.
Clearly, we can test the skewness and kurtosis by traditional t-stats as in
skewness kurtosis 3
tD p and t D p ; (1.20)
6=T 24=T

which both have N.0; 1/ distribution under the null hypothesis of a normal distribution.
See Figure 1.4 for an illustration.

1.2.3 Covariance and Correlation

The covariance of two variables (here x and y) is typically estimated as

PT
O xy D t D1 .x t N .y t
x/ N =T:
y/ (1.21)

(Sometimes we use T 1 in the denominator of the sample covariance instead of T .)

The correlation of two variables is then estimated as
O xy
Oxy D ; (1.22)
O x O y

where O x and O y are the estimated standard deviations. A correlation must be between
1 and 1. Note that covariance and correlation measure the degree of linear relation only.
This is illustrated in Figure 1.5.
See Figure 1.6 for an empirical illustration.
Under the null hypothesis of no correlation—and if the data is approximately normally
distributed, then
O
p  N.0; 1=T /; (1.23)
1 O2
so we can form a t-stat as
p O
tD Tp ; (1.24)
1 O2
which has an N.0; 1/ distribution (in large samples).

13
 Correlation of 0.9 Correlation of 0
2 2
1 1
0 0
y

y
−1 −1
−2 −2
−5 0 5 −5 0 5
x x

Correlation of -0.9 Correlation of 0

2 2
1 1
0 0
y

−1 −1
−2 −2
−5 0 5 −5 0 5
x x

Figure 1.5: Example of correlations.

1.3 Distributions Commonly Used in Tests

1.3.1 Standard Normal Distribution, N.0; 1/

Suppose the random variable x has a N.;  2 / distribution. Then, the test statistic has a
standard normal distribution
x 
zD  N.0; 1/: (1.25)

To see this, notice that x  has a mean of zero and that x= has a standard deviation of
unity.

14
 25
Monthly data on two U.S. indices 1957:1-2012:12
20
Correlation: 0.58

Large value stocks, montly returns, %

15

10

−5

−10

−15

−20

−25
−20 −10 0 10 20
Small growth stocks, montly returns, %

Figure 1.6: Scatter plot of two different portfolio returns

1.3.2 t-distribution

If we instead need to estimate  to use in (1.25), then the test statistic has tdf -distribution
x 
tD  tn ; (1.26)
O
where n denotes the “degrees of freedom,” that is the number of observations minus the
number of estimated parameters. For instance, if we have a sample with T data points
and only estimate the mean, then n D T 1.
The t-distribution has more probability mass in the tails: gives a more “conservative”
test (harder to reject the null hypothesis), but the difference vanishes as the degrees of
freedom (sample size) increases. See Figure 1.7 for a comparison and Table A.1 for
critical values.

Example 1.2 (t -distribution) If t D 2:0 and n D 50, then this is larger than the10%
critical value (but not the 5% critical value) for a 2-sided test in Table A.1.

15
 N(0,1) and t10 distributions N(0,1) and t50 distributions
0.4 0.4
10% crit val:
0.3 0.3
1.64 1.64
0.2 1.81 0.2 1.68

0.1 N(0,1) 0.1 N(0,1)

t10 t50
0 0
−2 0 2 −2 0 2
x x

Chi-square(n) distributions F(n, 50) distributions

1
n=2 n=2
0.4 n=5 n=5
4.61 2.41
0.3 9.24 1.97
0.5
0.2
0.1
0 0
0 5 10 0 1 2 3 4 5
x x

Figure 1.7: Probability density functions

1.3.3 Chi-square Distribution

If z  N.0; 1/, then z 2  21 , that is, z 2 has a chi-square distribution with one degree of
freedom. This can be generalized in several ways. For instance, if x  N.x ; xx / and
y  N.y ; yy / and they are uncorrelated, then Œ.x x /=x 2 C Œ.y y /=y 2  22 .
More generally, we have

v0˙ 1
v  2n , if the n  1 vector v  N.0; ˙/: (1.27)

See Figure 1.7 for an illustration and Table A.2 for critical values.

Example 1.3 (22 distribution) Suppose x is a 2  1 vector

" # " # " #!
x1 4 5 3
N ; :
x2 2 3 4

16
If x1 D 3 and x2 D 5, then
" #0 " # 1 " #
3 4 5 3 3 4
 6:1
5 2 3 4 5 2

has a  22 distribution. Notice that 6.1 is higher than the 5% critical value (but not the
1% critical value) in Table A.2.

1.3.4 F -distribution

If we instead need to estimate ˙ in (1.27) and let n1 be the number of elements in v

(previously called just n), then

v 0 ˙O 1
v=n1  Fn1 ;n2 (1.28)

where Fn1 ;n2 denotes an F -distribution with (n1 ; n2 ) degrees of freedom. Similar to the t-
distribution, n2 is the number of observations minus the number of estimated parameters.
See Figure 1.7 for an illustration and Tables A.3–A.4 for critical values.

1.4 Normal Distribution of the Sample Mean as an Approximation

In many cases, it is unreasonable to just assume that the variable is normally distributed.
The nice thing with a sample mean (or sample average) is that it will still be normally
distributed—at least approximately (in a reasonably large sample). This section gives
a short summary of what happens to sample means as the sample size increases (often
called “asymptotic theory”)
The law of large numbers (LLN) says that the sample mean converges to the true
population mean as the sample size goes to infinity. This holds for a very large class
of random variables, but there are exceptions. A sufficient (but not necessary) condition
for this convergence is that the sample average is unbiased (as in (1.16)) and that the
variance goes to zero as the sample size goes to infinity (as in (1.15)). (This is also called
convergence in mean square.) To see the LLN in action, see Figure 1.8.
p
The central limit theorem (CLT) says that T xN converges in distribution to a normal
distribution as the sample size increases. See Figure 1.8 for an illustration. This also
holds for a large class of random variables—and it is a very useful result since it allows

17
 √
Distribution of sample avg. Distribution of T × sample avg.
3 0.4
T =5
T = 25
T = 100 0.3
2
0.2
1
0.1

0 0
−2 −1 0 1 2 −5 0 5
Sample average
√
T × sample average

Sample average of z t − 1 where z t has a χ21 distribution

Figure 1.8: Sampling distributions

us to test hypothesis. Most estimators (including least squares and other methods) are
effectively some kind of sample average, so the CLT can be applied.

18
A Statistical Tables

n Critical values
10% 5% 1%
10 1:81 2:23 3:17
20 1:72 2:09 2:85
30 1:70 2:04 2:75
40 1:68 2:02 2:70
50 1:68 2:01 2:68
60 1:67 2:00 2:66
70 1:67 1:99 2:65
80 1:66 1:99 2:64
90 1:66 1:99 2:63
100 1:66 1:98 2:63
Normal 1:64 1:96 2:58

Table A.1: Critical values (two-sided test) of t distribution (different degrees of freedom)
and normal distribution.

19
 n Critical values
10% 5% 1%
1 2:71 3:84 6:63
2 4:61 5:99 9:21
3 6:25 7:81 11:34
4 7:78 9:49 13:28
5 9:24 11:07 15:09
6 10:64 12:59 16:81
7 12:02 14:07 18:48
8 13:36 15:51 20:09
9 14:68 16:92 21:67
10 15:99 18:31 23:21

Table A.2: Critical values of chisquare distribution (different degrees of freedom, n).

n1 n2 2n1 =n1
10 30 50 100 300
1 4:96 4:17 4:03 3:94 3:87 3:84
2 4:10 3:32 3:18 3:09 3:03 3:00
3 3:71 2:92 2:79 2:70 2:63 2:60
4 3:48 2:69 2:56 2:46 2:40 2:37
5 3:33 2:53 2:40 2:31 2:24 2:21
6 3:22 2:42 2:29 2:19 2:13 2:10
7 3:14 2:33 2:20 2:10 2:04 2:01
8 3:07 2:27 2:13 2:03 1:97 1:94
9 3:02 2:21 2:07 1:97 1:91 1:88
10 2:98 2:16 2:03 1:93 1:86 1:83

Table A.3: 5% Critical values of Fn1;n2 distribution (different degrees of freedom).

20
 n1 n2 2n1 =n1
10 30 50 100 300
1 3:29 2:88 2:81 2:76 2:72 2:71
2 2:92 2:49 2:41 2:36 2:32 2:30
3 2:73 2:28 2:20 2:14 2:10 2:08
4 2:61 2:14 2:06 2:00 1:96 1:94
5 2:52 2:05 1:97 1:91 1:87 1:85
6 2:46 1:98 1:90 1:83 1:79 1:77
7 2:41 1:93 1:84 1:78 1:74 1:72
8 2:38 1:88 1:80 1:73 1:69 1:67
9 2:35 1:85 1:76 1:69 1:65 1:63
10 2:32 1:82 1:73 1:66 1:62 1:60

Table A.4: 10% Critical values of Fn1;n2 distribution (different degrees of freedom).

21
2 Least Squares Estimation
Reference: Verbeek (2008) 2 and 4
More advanced material is denoted by a star ( ). It is not required reading.

2.1 Least Squares

2.1.1 Simple Regression: Constant and One Regressor

The simplest regression model is

y t D ˇ0 C ˇ1 x t C u t , where E u t D 0 and Cov.x t ; u t / D 0; (2.1)

where we can observe (have data on) the dependent variable y t and the regressor x t but
not the residual u t . In principle, the residual should account for all the movements in y t
that we cannot explain (by x t ).
Note the two very important assumptions: (i) the mean of the residual is zero; and
(ii) the residual is not correlated with the regressor, x t . If the regressor summarizes all
the useful information we have in order to describe y t , then the assumptions imply that
we have no way of making a more intelligent guess of u t (even after having observed x t )
than that it will be zero.
Suppose you do not know ˇ0 or ˇ1 , and that you have a sample of data: y t and x t for
t D 1; :::; T . The LS estimator of ˇ0 and ˇ1 minimizes the loss function
PT
tD1 .y t b0 b1 x t /2 D .y1 b0 b1 x1 /2 C .y2 b0 b1 x2 /2 C :::: (2.2)

by choosing b0 and b1 to make the loss function value as small as possible. The objective
is thus to pick values of b0 and b1 in order to make the model fit the data as closely
as possible—where close is taken to be a small variance of the unexplained part (the
residual). See Figure 2.1 for an illustration.

Remark 2.1 (First order condition for minimizing a differentiable function). We want
to find the value of b in the interval blow  b  bhigh , which makes the value of the

22
 Regression: y = b0 + b1 x + u
10

8 Intercept (b0 ) and slope (b1 ): 2.0 1.3

2
b0
0
y

−2

−4

−6 b1 is slope of line
−8 Data points
Regression line
−10
−10 −5 0 5 10
x

Figure 2.1: Example of OLS

differentiable function f .b/ as small as possible. The answer is blow , bhigh , or the value
of b where df .b/=db D 0. See Figure 2.2.

The first order conditions for a minimum are that the derivatives of this loss function
with respect to b0 and b1 should be zero. Notice that

@
.y t b0 b1 x t /2 D 2.y t b0 b1 x t /1 (2.3)
@b0
@
.y t b0 b1 x t /2 D 2.y t b0 b1 x t /x t : (2.4)
@b1

Let .ˇO0 ; ˇO1 / be the values of .b0 ; b1 / where that is true

@ PT
ˇO0 ˇO1 x t /2 D 2 TtD1 .y t ˇO0 ˇO1 x t /1 D 0 (2.5)
P
.y t
@ˇ0 tD1
@ PT
ˇO0 ˇO1 x t /2 D 2 TtD1 .y t ˇO0 ˇO1 x t /x t D 0; (2.6)
P
.y t
@ˇ1 tD1

23
 2b2 2b2 + (c − 4)2
20
30
15
20
10 10
5 0
2 6
0 0 4
−2 0 2 b −2 2 c
b

Minimum where Minimum where

d2b2 ∂2b2 ∂(c − 4)2

= 4b = 0 = 4b = 0 and = 2(c − 4) = 0
db ∂b ∂c

Figure 2.2: Quadratic loss function. Subfigure a: 1 coefficient; Subfigure b: 2 coefficients

which are two equations in two unknowns (ˇO0 and ˇO1 ), which must be solved simultane-
ously. These equations show that both the constant and x t should be orthogonal to the fit-
ted residuals, uO t D y t ˇO0 ˇO1 x t . This is indeed a defining feature of LS and can be seen
as the sample analogues of the assumptions in (2.1) that E u t D 0 and Cov.x t ; u t / D 0.
To see this, note that (2.5) says that the sample average of uO t should be zero. Similarly,
(2.6) says that the sample cross moment of uO t and x t should also be zero, which implies
that the sample covariance is zero as well since uO t has a zero sample mean.

Remark 2.2 Note that ˇi is the true (unobservable) value which we estimate to be ˇOi .
Whereas ˇi is an unknown (deterministic) number, ˇOi is a random variable since it is
calculated as a function of the random sample of y t and x t .

Remark 2.3 Least squares is only one of many possible ways to estimate regression co-
efficients. We will discuss other methods later on.

24
 OLS, y = b0 + b1 x + u, b0 = 0 Sum of squared errors
y: -1.5 -0.6 2.1 10
2 x: -1.0 0.0 1.0 b1 : 1.8
R2 : 0.92
1 Std(b1 ): 0.30
0 5
y

Data
−1 2x
−2 OLS
0
−1 −0.5 0 0.5 1 0 1 2 3 4
x b1

OLS Sum of squared errors

y: -1.3 -1.0 2.3 10
2 x: -1.0 0.0 1.0 b1 : 1.8
R2 : 0.81
1
Std(b1 ): 0.50
0 5
y

−1
−2
0
−1 −0.5 0 0.5 1 0 1 2 3 4
x b1

Figure 2.3: Example of OLS estimation

Remark 2.4 (Cross moments and covariance). A covariance is defined as

Cov.x; y/ D EŒ.x E x/.y E y/

D E.xy xEy y E x C E x E y/
D E xy ExEy Ey Ex C Ex Ey
D E xy E x E y:

When x D y, then we get Var.x/ D E x 2 .E x/2 . These results hold for sample moments
too.

When the means of y and x are zero, then we can disregard the constant. In this case,

25
(2.6) with ˇO0 D 0 immediately gives

D ˇO1 TtD1 x t x t or
PT P
t D1 y t x t
PT
O y t x t =T
ˇ1 D PTt D1 : (2.7)
t D1 x t x t =T

In this case, the coefficient estimator is the sample covariance (recall: means are zero) of
y t and x t , divided by the sample variance of the regressor x t (this statement is actually
true even if the means are not zero and a constant is included on the right hand side—just
more tedious to show it).

Example 2.5 (Simple regression) Consider the simple regression model (PSLS1). Sup-
pose we have the following data

Œ y1 y2 y3  D Œ 1:5 0:6 2:1  and Œ x1 x2 x3  D Œ 1 0 1 

To calculate the LS estimate according to (2.7) we note that

XT
x t x t D . 1/2 C 02 C 11 D 2 and
t D1
XT
x t y t D . 1/. 1:5/ C 0. 0:6/ C 1  2:1 D 3:6
t D1

This gives
3:6
ˇO1 D D 1:8:
2
The fitted residuals are
2 3 2 3 2 3 2 3
uO 1 1:5 1 0:3
4uO 2 5 D 4 0:65 1:8 4 0 5 D 4 0:65 :
6 7 6 7 6 7 6 7

uO 3 2:1 1 0:3

The fitted residuals indeed obey the first order condition (2.6) since
XT
x t uO t D . 1/  0:3 C 0. 0:6/ C 1  0:3 D 0:
t D1

See Figure 2.3 for an illustration.

See Table 2.1 and Figure 2.4 for illustrations.

26
 Scatter plot against market return Scatter plot against market return
30 30
US data
Excess return %, HiTec

Excess return %, Utils

20 1970:1-2012:12 20

10 10

0 0

−10 α -0.15
−10 α 0.24
−20 β 1.28 −20 β 0.52

−30 −30
−20 0 20 −20 0 20
Excess return %, market Excess return %, market

Figure 2.4: Scatter plot against market return

2.1.2 Multiple Regression

All the previous results still hold in a multiple regression—with suitable reinterpretations
of the notation.
Consider the linear model

y t D x1t ˇ1 C x2t ˇ2 C


C xk t ˇk C u t
D x t0 ˇ C u t ; (2.8)

where y t and u t are scalars, x t a k 1 vector, and ˇ is a k 1 vector of the true coefficients
(see Appendix A for a summary of matrix algebra). Least squares minimizes the sum of
the squared fitted residuals
PT
O 2t D
PT O 2;
x t0 ˇ/ (2.9)
t D1 u t D1 .y t

by choosing the vector ˇ. The first order conditions are

0kx1 D
PT O or
x t0 ˇ/
PT
D
PT 0 O
(2.10)
t D1 x t .y t t D1 x t y t t D1 x t x t ˇ;

which can be solved as P  1

ˇO D T PT
x x
t D1 t t
0
t D1 x t y t : (2.11)

27
 HiTec Utils
constant 0:15 0:24
. 1:00/ .1:58/
market return 1:28 0:52
.33:58/ .12:77/
R2 0:75 0:34
obs 516:00 516:00
Autocorr (t) 0:73 0:86
White 6:19 20:42
All slopes 386:67 176:89

Table 2.1: CAPM regressions, monthly returns, %, US data 1970:1-2012:12. Numbers

in parentheses are t-stats. Autocorr is a N(0,1) test statistic (autocorrelation); White is a
chi-square test statistic (heteroskedasticity), df = K(K+1)/2 - 1; All slopes is a chi-square
test statistic (of all slope coeffs), df = K-1

Example 2.6 With 2 regressors (k D 2), (2.10) is

x1t .y t x1t ˇO1 x2t ˇO2 /

" # " #
0 PT
D t D1
0 x2t .y t x1t ˇO1 x2t ˇO2 /

and (2.11) is
1
ˇO1
" # " #! " #
PT x1t x1t x1t x2t PT x1t y t
D :
ˇO2
t D1 t D1
x2t x1t x2t x2t x2t y t

Example 2.7 (Regression with an intercept and slope) Suppose we have the following
data:
" #
1 1 1
Œ y1 y2 y3  D Œ 1:5 0:6 2:1  and Œ x1 x2 x3  D :
1 0 1

This is clearly the same as in Example 2.5, except that we allow for an intercept—which
turns out to be zero. The notation we need to solve this problem is the same as for a

28
 HiTec Utils
constant 0:12 0:03
.0:95/ .0:19/
market return 1:11 0:65
.31:08/ .16:67/
SMB 0:23 0:19
.4:37/ . 3:61/
HML 0:58 0:45
. 9:74/ .7:01/
R2 0:83 0:47
obs 516:00 516:00
Autocorr (t) 0:47 1:18
White 70:79 49:06
All slopes 425:95 242:74

Table 2.2: Fama-French regressions, monthly returns, %, US data 1970:1-2012:12. Num-

bers in parentheses are t-stats. Autocorr is a N(0,1) test statistic (autocorrelation); White
is a chi-square test statistic (heteroskedasticity), df = K(K+1)/2 - 1; All slopes is a chi-
square test statistic (of all slope coeffs), df = K-1

general multiple regression. Therefore, calculate the following:

" # " # " #
XT 1 h i 1 h i 1 h i
x t x t0 D 1 1 C 1 0 C 1 1
t D1 1 0 1
" # " # " #
1 1 1 0 1 1
D C C
1 1 0 0 1 1
" #
3 0
D
0 2

" # " # " #

XT 1 1 1
xt yt D . 1:5/ C . 0:6/ C 2:1
t D1 1 0 1
" # " # " #
1:5 0:6 2:1
D C C
1:5 0 2:1
" #
0
D
3:6

29
 PT
To calculate the LS estimate, notice that the inverse of the t D1 x t x t0 is
" # 1 " #
3 0 1=3 0
D ;
0 2 0 1=2

which can be verified by

" #" # " #
1=3 0 3 0 1 0
D :
0 1=2 0 2 0 1

The LS estimate is therefore

X  1
T XT
ˇO D x t x t0 xt yt
t D1 t D1
" #" #
1=3 0 0
D
0 1=2 3:6
" #
0
D :
1:8

2.1.3 Least Squares: Goodness of Fit

The quality of a regression model is often measured in terms of its ability to explain the
movements of the dependent variable.
Let yO t be the fitted (predicted) value of y t . For instance, with (2.1) it would be yO t D
ˇ0 C ˇO1 x t . If a constant is included in the regression (or the means of y and x are zero),
O
then a check of the goodness of fit of the model is given by the fraction of the variation in
y t that is explained by the model

Var.yO t / Var.uO t /
R2 D D1 ; (2.12)
Var.y t / Var.y t /
which can also be rewritten as the squared correlation of the actual and fitted values

R2 D Corr.y t ; yO t /2 : (2.13)

Notice that we must have constant in regression for R2 to make sense—unless all variables
have zero means.

30
Example 2.8 (R2 ) From Example 2.5 we have Var.uO t / D 0:18 and Var.y t / D 2:34, so

R2 D 1 0:18=2:34  0:92:

See Figure 2.3.

Proof. (of (2.12)–(2.13)) Write the regression equation as

y t D yO t C uO t ;

where hats denote fitted values. Since yO t and uO t are uncorrelated (always true in OLS—
provided the regression includes a constant), we have

Var.y t / D Var.yO t / C Var.uO t /:

R2 is defined as the fraction of Var.y t / that is explained by the model

Var.yO t / Var.y t / Var.uO t / Var.uO t /

R2 D D D1 :
Var.y t / Var.y t / Var.y t /

Equivalently, we can rewrite R2 by noting that

Cov .y t ; yO t / D Cov .yO t C uO t ; yO t / D Var .yO t / :

Use this in the denominator of R2 and multiply by Cov .y t ; yO t / = Var .yO t / D 1

Cov .y t ; yO t /2
R2 D D Corr .y t ; yO t /2 :
Var.y t / Var .yO t /

To understand this result, suppose that x t has no explanatory power, so R2 should

be zero. How does that happen? Well, if x t is uncorrelated with y t , then ˇO1 D 0. As
a consequence yO t D ˇO0 , which is a constant. This means that R2 in (2.12) is zero,
since the fitted residual has the same variance as the dependent variable (yO t captures
noting of the movements in y t ). Similarly, R2 in (2.13) is also zero, since a constant is
always uncorrelated with anything else (as correlations measure comovements around the
means). See Figure 2.5 for an example.

31
Remark 2.9 (R2 from simple regression) Suppose yO t D ˇ0 C ˇ1 x t , so (2.13) becomes

2 Cov .y t ; ˇ0 C ˇ1 x t /2 Cov .y t ; x t /2
R D D D Corr.y t ; x t /2 :
Var.y t / Var.ˇ0 C ˇ1 x t / Var.y t / Var.x t /

The R2 can never decrease as we add more regressors, which might make it attractive
to add more and more regressors. To avoid that, some researchers advocate using an ad
hoc punishment for many regressors, RN 2 D 1 .1 R2 /.T 1/=.T k/, where k is the
number of regressors (including the constant). This measure can be negative.

Slope coefficient (b) R2

0.5 Slope with 90% conf band

0.1
0
0.05
−0.5
0
0 20 40 60 0 20 40 60
Return horizon (months) Return horizon (months)

Scatter plot, 36 month returns Monthly US stock returns 1926:1-2012:12

2
Regression: rt = a + brt−1 + ǫt
1
Return

−1

−2
−2 −1 0 1 2
lagged return

Figure 2.5: Prediction equations for US stock returns

2.1.4 Least Squares: Outliers

Since the loss function in (2.2) is quadratic, a few outliers can easily have a very large
influence on the estimated coefficients. For instance, suppose the true model is y t D

32
 OLS: sensitivity to outlier
2
y: -1.125 -0.750 1.750 1.125
x: -1.500 -1.000 1.000 1.500

1
Three data points are on the
line y = 0.75x, while the
fourth has a big error
0
y

−1
Data
OLS (0.25 0.90)
True (0.00 0.75)
−2
−3 −2 −1 0 1 2 3
x

Figure 2.6: Data and regression line from OLS

0:75x t C u t , and that the residual is very large for some time period s. If the regression
coefficient happened to be 0.75 (the true value, actually), the loss function value would be
large due to the u2t term. The loss function value will probably be lower if the coefficient
is changed to pick up the ys observation—even if this means that the errors for the other
observations become larger (the sum of the square of many small errors can very well be
less than the square of a single large error).
There is of course nothing sacred about the quadratic loss function. Instead of (2.2)
one could, for instance, use a loss function in terms of the absolute value of the error
˙ tTD1 jy t ˇ0 ˇ1 x t j. This would produce the Least Absolute Deviation (LAD) estima-
tor. It is typically less sensitive to outliers. This is illustrated in Figure 2.7. However, LS
is by far the most popular choice. There are two main reasons: LS is very easy to compute
and it is fairly straightforward to construct standard errors and confidence intervals for the
estimator. (From an econometric point of view you may want to add that LS coincides
with maximum likelihood when the errors are normally distributed.)

2.1.5 The Distribution of ˇO

Note that the estimated coefficients are random variables since they depend on which par-
ticular sample that has been “drawn.” This means that we cannot be sure that the estimated

33
 OLS vs LAD of y = 0.75x + u
2
y: -1.125 -0.750 1.750 1.125
x: -1.500 -1.000 1.000 1.500

0
y

−1
Data
OLS (0.25 0.90)
LAD (0.00 0.75)
−2
−3 −2 −1 0 1 2 3
x

Figure 2.7: Data and regression line from OLS and LAD

coefficients are equal to the true coefficients (ˇ0 and ˇ1 in (2.1)). We can calculate an es-
timate of this uncertainty in the form of variances and covariances of ˇO0 and ˇO1 . These
can be used for testing hypotheses about the coefficients, for instance, that ˇ1 D 0.
To see where the uncertainty comes from consider the simple case in (2.7). Use (2.1)
to substitute for y t (recall ˇ0 D 0)
PT
tD1 x t .ˇ1 x t C u t / =T
ˇO1 D PT
t D1 x t x t =T
PT
x t u t =T
D ˇ1 C PTt D1 ; (2.14)
x
t D1 t tx =T

so the OLS estimate, ˇO1 , equals the true value, ˇ1 , plus the sample covariance of x t and
u t divided by the sample variance of x t . One of the basic assumptions in (2.1) is that
the covariance of the regressor and the residual is zero. This should hold in a very large
sample (or else OLS cannot be used to estimate ˇ1 ), but in a small sample it may be
different from zero. Since u t is a random variable, ˇO1 is too. Only as the sample gets very
large can we be (almost) sure that the second term in (2.14) vanishes.
Equation (2.14) will give different values of ˇO when we use different samples, that is
different draws of the random variables u t , x t , and y t . Since the true value, ˇ, is a fixed

34
constant, this distribution describes the uncertainty we should have about the true value
after having obtained a specific estimated value.
The first conclusion from (2.14) is that, with u t D 0 the estimate would always
be perfect. In contrast, with large movements in u t we will see large movements in ˇO
(across samples). The second conclusion is that a small sample (small T ) will also lead
to large random movements in ˇO1 —in contrast to a large sample where the randomness
in TtD1 x t u t =T is averaged out more effectively (should be zero in a large sample).
P

There are three main routes to learn more about the distribution of ˇ: O (i) set up a
small “experiment” in the computer and simulate the distribution (Monte Carlo or boot-
strap simulation); (ii) pretend that the regressors can be treated as fixed numbers (or at
least independent of the residuals in all periods) and then assume something about the
distribution of the residuals; or (iii) use the asymptotic (large sample) distribution as an
approximation. The asymptotic distribution can often be derived, in contrast to the exact
distribution in a sample of a given size. If the actual sample is large, then the asymptotic
distribution may be a good approximation.
The simulation approach has the advantage of giving a precise answer—but the dis-
advantage of requiring a very precise question (must write computer code that is tailor
made for the particular model we are looking at, including the specific parameter values).
See Figure 2.11 for an example.
The typical outcome of all three approaches will (under strong assumptions) be that
   1 
ˇO  N ˇ; T 0
2 ; (2.15)
P
t D1 x t x t

which allows for x t to be a vector with k elements. Clearly, with k D 1, x t0 D x t . See

Figure 2.8 for an example.
An alternative way of expressing the distribution (often used in conjunction with
asymptotic) theory is
p
   1 
T .ˇO ˇ/  N 0; T 0
2 : (2.16)
P
t D1 x t x t =T

p
This is the same as (2.15). (To see that, consider dividing the LHS of (2.16) by T .
Then, the variance on the RHS must be divided by T , which gives the same variance as
in (2.15). Then, add ˇ to the LHS, which changes the mean on the RHS to ˇ. We then
have (2.15).)

35
 Distribution of LS slope coefficient
4
Model: yt = 0.9xt + ξt ,
where ξt ∼ N (0, 2) and
T = 200
3
Histogram shows LS
estimate of b in
yt = a + bxt + ut
2 across 25000 simulations

Solid curve is the theo-

retical distribution
1

0
0.4 0.6 0.8 1 1.2 1.4
slope coefficient

Figure 2.8: Distribution of OLS estimate, from simulation and theory

Example 2.10 (Distribution of slope coefficient) From Example 2.5 we have Var.uO t / D
 2 D 0:18 and TtD1 x t x t D 2, so Var.ˇO1 / D 0:18=2 D 0:09, which gives Std.ˇO1 / D
P

0:3.

Example 2.11 (Covariance matrix of b1 and b2 ) From Example 2.7

" #
XT 3 0
x t x t0 D and  2 D 0:18, then
tD1 0 2
ˇO1 Var.ˇO1 / Cov.ˇO1 ; ˇO2 /
" #! " #
Var D
ˇO2 Cov.ˇO1 ; ˇO2 / Var.ˇO2 /
" # " #
1=3 0 0:06 0
D 0:18 D :
0 1=2 0 0:09

The standard deviations (also called standard errors) are therefore

Std.ˇO1 /
" # " #
0:24
D :
Std.ˇO2 / 0:3

36
2.1.6 The Distribution of ˇO with Fixed Regressors

The assumption of fixed regressors makes a lot of sense in controlled experiments, where
we actually can generate different samples with the same values of the regressors (the
heat or whatever). It makes much less sense in econometrics. However, it is easy to
derive results for this case—and those results happen to be very similar to what asymptotic
theory gives.
The results we derive below are based on the Gauss-Markov assumptions: the residu-
als have zero means, have constant variances and are not correlated across observations.
In other words, the residuals are zero mean iid variables. As an alternative to assuming
fixed regressors (as we do here), it is assumed that the residuals and regressors are in-
dependent. This delivers very similar results. We will also assume that the residuals are
normally distributed (not part of the typical Gauss-Markov assumptions).
Write (2.14) as
1
ˇO1 D ˇ1 C PT .x1 u1 C x2 u2 C : : : xT uT / : (2.17)
t D1 x t x t

Since x t are assumed to be constants (not random), the expected value of this expression
is
1
E ˇO1 D ˇ1 C PT .x1 E u1 C x2 E u2 C : : : xT E uT / D ˇ1 (2.18)
tD1 x t x t
since we always assume that the residuals have zero means (see (2.1)). The interpretation
is that we can expected OLS to give (on average) a correct answer. That is, if we could
draw many different samples and estimate the slope coefficient in each of them, then the
average of those estimates would be the correct number (ˇ1 ). Clearly, this is something
we want from an estimation method (a method that was systematically wrong would not
be very attractive).

Remark 2.12 (Linear combination of normally distributed variables.) If the random

variables z t and v t are normally distributed, then a C bz t C cv t is too. To be precise,
a C bz t C cv t  N a C bz C cv ; b 2 z2 C c 2 v2 C 2bczv .

Suppose u t  N 0;  2 , then (2.17) shows that ˇO1 is normally distributed. The rea-

son is that ˇO1 is just a constant (ˇ1 ) plus a linear combination of normally distributed
residuals (with fixed regressors x t = TtD1 x t x t can be treated as constant). It is straight-
P

forward to see that the mean of this normal distribution is ˇ1 (the true value), since the

37
rest is a linear combination of the residuals—and they all have a zero mean. Finding the
variance of ˇO1 is just slightly more complicated. Remember that we treat x t as fixed num-
bers (“constants”) and assume that the residuals are iid: they are uncorrelated with each
other (independently distributed) and have the same variances (identically distributed).
The variance of (2.17) is then
1 1
Var.ˇO1 / D PT Var .x1 u1 C x2 u2 C : : : xT u t / PT
t D1 x t x t t D1 x t x t
1 1
D PT x12 12 C x22 22 C : : : xT2 T PT 2


t D1 x t x t t D1 x t x t
1 1
D PT x12  2 C x22  2 C : : : xT2  2 PT


t D1 x t x t t D1 x t x t
1 P
T
 1
D PT t D1 x t x t  2 PT
t D1 x t x t t D1 x t x t
1
D PT  2: (2.19)
t D1 x t x t

The first line follows directly from (2.17), since ˇ1 is a constant. Notice that the two
PT
t D1 x t x t terms are kept separate in order to facilitate the comparison with the case of
several regressors. The second line follows from assuming that the residuals are uncorre-
lated with each other (Cov.ui ; uj / D 0 if i ¤ j ), so all cross terms (xi xj Cov.ui ; uj /)
are zero. The third line follows from assuming that the variances are the same across ob-
servations (i2 D j2 D  2 ). The fourth and fifth lines are just algebraic simplifications.
Notice that the denominator increases with the sample size while the numerator stays
constant: a larger sample gives a smaller uncertainty about the estimate. Similarly, a lower
volatility of the residuals (lower  2 ) also gives a lower uncertainty about the estimate. See
Figure 2.9.

Example 2.13 When the regressor is just a constant (equal to one) x t D 1, then we have
PT 0
D
PT O D  2 =T:
 10 D T so Var.ˇ/
tD1 x t x t tD1 1

(This is the classical expression for the variance of a sample mean.)

Example 2.14 When the regressor is a zero mean variable, then we have
PT 0 O D  2 = ŒVar.x t /T  :
D Var.x t /T so Var.ˇ/
tD1 x t x t

38
 Regression: y = b0 + b1 x + u Regression: large error variance
10 10
b0 and b1 : 2.0 1.3

5 5

0 0
y

y
−5 −5
Data points
Regression line
−10 −10
−10 −5 0 5 10 −10 −5 0 5 10
x x
Regression: little variation in x Regression: small sample
10 10

5 5

0 0
y

−5 −5

−10 −10
−10 −5 0 5 10 −10 −5 0 5 10
x x

Figure 2.9: Regressions: importance of error variance and variation of regressor

The variance is increasing in  2 , but decreasing in both T and Var.x t /.

Example 2.15 When the regressor is just a constant (equal to one) and one variable
regressor with zero mean, f t , so x t D Œ1; f t 0 , then we have
" # " #
PT 0 T 1 f t 1 0
t D1 x t x t D DT
P
t D1 , so
f t f t2 0 Var.f t /
ˇO1
" #! " #
2
P
T
 1  =T 0
Var D 2 0
t D1 x t x t D :
O
ˇ2 0  2 = ŒVar.f t /T 

A combination of the two previous examples.

39
2.1.7 Multicollinearity

When the regressors in a multiple regression are highly correlated, then we have a practi-
cal problem: the standard errors of individual coefficients tend to be large.
As a simple example, consider the regression

y t D ˇ1 x1t C ˇ2 x2t C u t ; (2.20)

where (for simplicity) the dependent variable and the regressors have zero means. In this
case, the variance of

1 1 2
Var.ˇO2 / D ; (2.21)
1 Corr.x1t ; x2t /2 Var.x2t / T
where the new term is the (squared) correlation. If the regressors are highly correlated,
then the uncertainty about the slope coefficient is high. The basic reason is that we see
that the variables have an effect on y t , but it is hard to tell if that effect is from regressor
one or two.
Proof. (of 2.21). Recall that for a 2  2 matrix we have
" # 1 " #
a b 1 d b
D :
c d ad bc c a

For the regression (2.20) we get

" PT PT # 1
2
x x x
tD1 1t 2t
PT t D1 1t P T 2
D
t D1 x1t x2t t D1 x2t
" PT PT #
1 x2 x x
1t 2t
2 PTt D1 2t PtTD1 2 :
tD1 x1t x2t t D1 x1t
PT PT P
2 2 T
tD1 x1t t D1 x2t t D1 x1t x2t

The variance of the second slope coefficient is  2 time the lower right element of this

40
matrix. Multiply and divide by T to get
PT 2
2 t D1 x1t =T
Var.ˇO2 / D 2
T PT PT P
1 2 1 2 T 1
t D1 T x1t t D1 T x2t t D1 T x1t x2t

2 Var.x1t /
D
T Var.x1t / Var.x2t / Cov.x1t ; x2t /2
 2 1= Var.x2t /
D Cov.x1t ;x2t /2
;
T 1
Var.x1t / Var.x2t /

which is the same as (2.21).

2.1.8 O When the Assumptions Are Wrong

The Distribution of ˇ:

The results on the distribution ˇO have several weak points—which will be briefly dis-
cussed here.
First, the Gauss-Markov assumptions of iid residuals (constant volatility and no cor-
relation across observations) are likely to be false in many cases. These issues (het-
eroskedasticity and autocorrelation) are therefore discussed at length later on.
Second, the idea of fixed regressor is clearly just a simplifying assumptions—and
unlikely to be relevant for financial data. This forces us to rely on asymptotic (“large
sample”) theory (or do simulations). The main results from asymptotic theory (see below)
is that the main result (2.15) is a good approximation in large samples, provided the
Gauss-Markov assumptions are correct (if not, see later sections on heteroskedasticity
and autocorrelation). However, things are more complicated in small samples. Only
simulations can help us there.

Example 2.16 (When OLS is biased, at least in a finite sample) OLS on an AR(1), y t D
ˇ1 C ˇ2 y t 1 C u t , is not unbiased. In this case, the regressors are not fixed and u t is not
independent of the regressor for all periods: u t is correlated with y t (obviously)—which
is the regressor in t C 1. See Figure 2.10.

2.1.9 The Distribution of ˇ: O A Bit of Asymptotic Theory

A law of large numbers would (in most cases) say that both TtD1 x t2 =T and TtD1 x t u t =T
P P

in (2.14) converges to their expected values as T ! 1. The reason is that both are sample

41
 Distribution of LS estimator, T = 25 Distribution of LS estimator, T = 100
Mean and std: Mean and std:
0.1 0.74 0.16 0.1 0.86 0.06

0.05 0.05

0 0
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

True model: yt = 0.9yt−1 + ǫt , ǫt is iid N(0,2)

Estimated model: yt = a + ρyt−1 + ut
Number of simulations: 25000

Figure 2.10: Results from a Monte Carlo experiment of LS estimation of the AR coeffi-
cient.

averages of random variables (clearly, both x t2 and x t u t are random variables). These ex-
pected values are Var .x t / and Cov .x t ; u t /, respectively (recall both x t and u t have zero
means). The key to show that ˇO is consistent is that Cov .x t ; u t / D 0. This highlights the
importance of using good theory to derive not only the systematic part of (2.1), but also
in understanding the properties of the errors. For instance, when economic theory tells
us that y t and x t affect each other (as prices and quantities typically do), then the errors
are likely to be correlated with the regressors—and LS is inconsistent. One common way
to get around that is to use an instrumental variables technique. Consistency is a feature
we want from most estimators, since it says that we would at least get it right if we had
enough data.
Suppose that ˇO is consistent. Can we say anything more about the asymptotic distri-
bution. Well, the distribution of ˇO converges to a spike with all the mass at ˇ, but the
p
distribution of T .ˇO ˇ/, will typically converge to a non-trivial normal distribution.
To see why, note from (2.14) that we can write
p P  1 p PT
T .ˇO ˇ/ D T 2
t D1 x t =T T t D1 x t u t =T: (2.22)

The first term on the right hand side will typically converge to the inverse of Var .x t /, as
p
discussed earlier. The second term is T times a sample average (of the random variable

42
 Distribution of LS t-stat, T = 5 Distribution of LS t-stat, T = 100
t = (b̂ − 0.9)/Std(b̂)
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−4 −2 0 2 4 −4 −2 0 2 4

Model: Rt = 0.9ft + ǫt , ǫt = vt − 2,
where vt has a χ22 distribution Probability density functions

Estimated model: yt = a + bft + ut N(0,1)

0.4
Number of simulations: 25000 χ22 − 2
0.3
T = 5 T = 100 0.2
Kurtosis of t-stat: 46.753 3.049
Frequency of |t-stat| > 1.65 0.294 0.105 0.1
Frequency of |t-stat| > 1.96 0.227 0.054 0
−4 −2 0 2 4

Figure 2.11: Distribution of LS estimator when residuals have a non-normal distribution

x t u t ) with a zero expected value, since we assumed that ˇO is consistent. Under weak
p
conditions, a central limit theorem applies so T times a sample average converges to a
p
normal distribution. This shows that T ˇO has an asymptotic normal distribution. It turns
out that this is a property of many estimators, basically because most estimators are some
kind of sample average. The properties of this distribution are quite similar to those that
we derived by assuming that the regressors were fixed numbers.

2.2 Hypothesis Testing

2.2.1 Testing a Single Coefficient

We are here interested in testing the null hypothesis that ˇ D q, where q is a number of in-
terest. A null hypothesis if often denoted H0 . (Econometric programs often automatically
report results for H0 : ˇ D 0.)

43
 90% confidence band 90% confidence band
0.8 0.4
N(3,0.25) 10% critical N(0,1) 10% critical
0.6 values: √ 0.3 values:
3±1.65 × 0.25 -1.65 and 1.65
0.4 (2.17 and 3.83) 0.2

0.2 0.1

0 0
1 2 3 4 5 −4 −2 0 2 4
(β̂ − 3)/ 0.25
√
β̂

Calculating a p-value
0.4
N(0,1) With β̂ = 1.95,
t = −2.1.
0.3
p-value:
2 × 1.8% = 3.6%
0.2

0.1

0
−4 −2 0 2 4
(β̂ − 3)/ 0.25
√

Figure 2.12: Confidence band around the null hypothesis

We assume that the estimates are normally distributed. To be able to easily compare
with printed tables of probabilities, we transform to a N.0; 1/ variable. In particular, if
the true coefficient is really q, then ˇO q should have a zero mean (recall that E ˇO equals
O we should
the true value) and by further dividing by the standard error (deviation) of ˇ,
have
ˇO q
tD  N.0; 1/ (2.23)
O
Std.ˇ/
In case jt j is very large (say, 1.65 or larger), then ˇO is a very unlikely outcome if E ˇO
(which equals the true coefficient value, ˇ) is indeed q. We therefore draw the conclusion
that the true coefficient is not q, that is, we reject the null hypothesis.
The logic this hypothesis test is perhaps best described by a 90% confidence band

44
around the null hypothesis

Pr.ˇO is in ConfBand/ D 90%, where (2.24)

O
ConfBand D q ˙ 1:65 Std.ˇ/:

See Figure 2.12 for an illustration. The idea is that if the true value of the coefficient is
q, then the estimate ˇO should be inside this band in 90% of the cases (that is, different
samples). Hence, a ˇO outside the band is unlikely to happen if the true coefficient is q:
we will interpret that situation as if the true value is not q.
If the point estimate is outside this confidence band band, then this is the same as
rejecting the null hypothesis. To see that, notice that for ˇO to be outside the band we must
have
jtj > 1:65; (2.25)

See Figure 2.12 for an illustration.

Proof. (of (2.25)) For ˇO to be outside the band we must have

ˇO < q O or ˇO > q C 1:65 Std.ˇ/:

1:65 Std.ˇ/ O

Rearrange this by subtracting q from both sides of the inequalities and then divide both
O
sides by Std.ˇ/
ˇO q ˇO q
< 1:65 or > 1:65:
O
Std.ˇ/ O
Std.ˇ/
This is the same as (2.25).
p
Example 2.17 (t-test) With Std.ˇ/ O D 0:25 and q D 3, the 90% confidence band is
p
3 ˙ 1:65  0:25, that is, Œ2:175; 3:825. Notice that ˇO D 1:95 is outside this band, so
p
we reject the null hypothesis. Equivalently, t D .1:95 3/= 0:25 D 2:1 is outside the
band Œ 1:65; 1:65.

Using a 90% confidence band means that you are using a 10% significance level. If
you want a more conservative test (that is, making it harder to reject the null hypotheis),
then you may change from the critical value 1.65 to 1.96. This gives a 95% confidence
band, so the significance level is 5%. See Figure 2.13 for an illustration and Tables 2.1
and Table 2.2 for examples.

45
 The p-value is a related concept. It is the lowest significance level at which we can
reject the null hypothesis. See Figure 2.12 for an illustration.
p
Example 2.18 (p-value) With Std.ˇ/ O D 0:25, ˇO D 1:95 and q D 3, we have t D 2:1.
According to a N(0,1) distribution, the probability of 2:1 or lower is 1.8%, so the p-value
is 3.6%. We thus reject the null hypothesis at the 10% significance level and also at the
5% significance level.

We sometimes compare with a t -distribution instead of a N.0; 1/, especially when the
sample is short. For instance, with 22 data points and two estimated coefficients (so there
are 20 degrees of freedom), the 10% critical value of a t-distribution is 1.72 (while it is
1.65 for the standard normal distribution). However, for samples of more than 30–40 data
points, the difference is trivial—see Table A.1.

N(0,1) distribution
0.4
5% in each tail: -1.65 1.65
2.5% in each tail: -1.96 1.96
0.3

0.2

0.1

0
−3 −2 −1 0 1 2 3
x

Figure 2.13: Density function of a standard normal distribution

2.2.2 Confidence Band around the Point Estimate

Andther way to construct a confidence band is to center the band on the point estimate

ˇO ˙ 1:65 Std.ˇ/:
O (2.26)

46
In this case, we are 90% sure that the true value will be inside the band. If the value q (say,
q D 3) is not in this band, then this is the same thing as rejecting (on the 10% significance
level) the null hypothesis that coefficient equals q. (As before, change 1.65 to 1.96 to get
a 95% confidence band.)

2.2.3 Power and Size

The size is the probability of rejecting a true H0 . It should be low. Provided you use a
valid test (correct standard error, etc), the size is the significance level you have chosen
(the probability you use to construct critical values). For instance, with a t-test with
critical values . 1:65; 1:65/, the size is 10%. (The size is sometime called the type I
error.)
The power is the probability of rejecting a false H0 . It should be high. Typically, it
cannot be controlled (but some tests are better than others...). This power depends on how
false H0 is, which we will never know. All we we do is to create artificial examples to
get an idea of what the power would be for different tests and for different values of the
true parameter ˇ. For instance, with a t-test using the critical values 1:65 and 1:65, the
power would be
power = Pr.t  1:65/ C Pr.t  1:65/: (2.27)

(1 power is sometimes called the type II error. This is the probability of not rejecting a
false H0 .)
To make this more concrete, suppose we test the null hypothesis that the coefficient is
equal to q, but the true value happens to be ˇ. Since the OLS estimate, ˇO is distributed as
O it must be the case that the t-stat is distributed as
N Œˇ; Std.ˇ/,

ˇO q
" #
ˇ q
tD N ;1 : (2.28)
O
Std.ˇ/ Std.ˇ/O

We can then calculate the power as the probability that t  1:65 or t  1:65, when t
has the distribution on the RHS in (2.28).

O D 1=3, then the power is 0.56. See Figure

Example 2.19 If ˇ D 1:6, q D 1 and Std.ˇ/
2.14.

47
 Distribution of t when true β = 0.6 Distribution of t when true β = 1.01
0.4 0.4
Prob of Prob of
rejection: t = (bLS − 1)/Std(bLS ) rejection:
0.3 0.3
0.33 0.10
0.2 0.2

0.1 0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4
t t

Distribution of t when true β = 1.6 Std(bLS ) = 0.333

0.4
Prob of
rejection: The test is: reject if t ≤ −1.65 or t > 1.65
0.3
0.56
0.2

0.1

0
−4 −2 0 2 4
t

Figure 2.14: Power of t-test, assuming different true parameter values

2.2.4 Joint Test of Several Coefficients: Chi-Square Test

A joint test of several coefficients is different from testing the coefficients one at a time.
For instance, suppose your economic hypothesis is that ˇ1 D 1 and ˇ3 D 0. You could
clearly test each coefficient individually (by a t-test), but that may give conflicting results.
In addition, it does not use the information in the sample as effectively as possible. It
might well be the case that we cannot reject any of the hypotheses (that ˇ1 D 1 and
ˇ3 D 0), but that a joint test might be able to reject it.
Intuitively, a joint test is like exploiting the power of repeated sampling as illustrated
by the following example. My null hypothesis might be that I am a better tennis player
than my friend. After playing (and losing) once, I cannot reject the null—since pure
randomness (wind, humidity,...) might have caused the result. The same is true for the
second game (and loss)—if I treat the games as completely unrelated events. However,
considering both games, the evidence against the null hypothesis is (unfortunately) much

48
stronger.

χ21 distribution χ22 distribution

Pr(x ≥ 2.71) = 0.1 Pr(x ≥ 4.61) = 0.1
1 0.4
0.3
0.5 0.2
0.1
0 0
0 5 10 15 0 5 10 15
x x

χ25 distribution
0.15
Pr(x ≥ 9.24) = 0.1

0.1

0.05

0
0 5 10 15
x

Figure 2.15: Density functions of 2 distributions with different degrees of freedom

A joint test makes use of the following remark.

Remark 2.20 (Chi-square distribution) If v is a zero mean normally distributed vector,

then we have
v 0 ˙ 1 v  2n , if the n  1 vector v  N.0; ˙/:

As a special case, suppose the vector only has one element. Then, the quadratic form can
be written Œv= Std.v/2 , which is the square of a t-statistic.

Example 2.21 (Quadratic form with a chi-square distribution) If the 2  1 vector v has
the following normal distribution
" # " # " #!
v1 0 1 0
N ; ;
v2 0 0 2

49
then the quadratic form
" #0 " #" #
v1 1 0 v1
D v12 C v22 =2
v2 0 1=2 v2

has a 22 distribution.

For instance, suppose we have estimated a model with three coefficients and the null
hypothesis is
H0 W ˇ1 D 1 and ˇ3 D 0: (2.29)

It is convenient to write this on matrix form as

2 3
" # ˇ1 " #
1 0 0 6 7 1
4ˇ2 5 D or more generally (2.30)
0 0 1 0
ˇ3
Rˇ D q; (2.31)

where q has J (here 2) rows. Notice that the covariance matrix of these linear combina-
tions is then
Var.Rˇ/O D RV .ˇ/RO 0; (2.32)
O denotes the covariance matrix of the coefficients, for instance, from (2.19).
where V .ˇ/
Putting together these results we have the test static (a scalar)

.RˇO O 0  1 .RˇO
q/0 ŒRV .ˇ/R q/  2J : (2.33)

This test statistic is compared to the critical values of a 2J distribution—see Table A.2.
(Alternatively, it can put in the form of an F statistics, which is a small sample refine-
ment.)
A particularly important case is the test of the joint hypothesis that all k 1 slope
coefficients in the regression (that is, excluding the intercept) are zero. It can be shown
that the test statistics for this hypothesis is (assuming your regression also contains an
intercept)
TR2 =.1 R2 /  2k 1 : (2.34)

See Tables 2.1 and 2.2 for examples of this test.

50
Example 2.22 (Joint test) Suppose H0 : ˇ1 D 0 and ˇ3 D 0; .ˇO1 ; ˇO2 ; ˇO3 / D .2; 777; 3/
and
2 3
" # 4 0 0
1 0 0 O D6
RD and V .ˇ/ 40 33 05 , so
7
0 0 1
0 0 1
2 32 3
" # 4 0 0 1 0 " #
O 0D 1 0 0 4 0
RV .ˇ/R 40 33 05 40 05 D :
6 7 6 7
0 0 1 0 1
0 0 1 0 1

Then, (2.33) is
0 2 3 10 02 3 1
" # 2 " # " # " 1 # 2 " #
B 1 0 0 6 0 C 4 0 B 1 0 0 6 7 0 C
47775 47775
7
@ A @ A
0 0 1 0 0 1 0 0 1 0
3 3
" #" #
h i 0:25 0 2
2 3 D 10;
0 1 3

which is higher than the 10% critical value of the 22 distribution (which is 4.61).

Proof. (of (2.34)) Recall that R2 D Var .yO t / = Var .y t / D 1 Var .uO t / = Var .y t /,
where yO t D x t0 ˇO and uO t are the fitted value and residual respectively. We therefore get
TR2 =.1 R2 / D T Var .yO t / = Var .uO t / :To simplify the algebra, assume that both y t and
x t are demeaned and that no intercept is used. (We get the same results, but after more
work, if we relax this assumption.) In this case we can rewrite as TR2 =.1 R2 / D
T ˇO 0 Var.x t /ˇ=
O 2 , where  2 D Var .uO t /. If the iid assumptions are correct, then the
variance-covariance matrix of ˇO is V .ˇ/ O D ŒT Var.x t / 1  2 (see (2.15)), so we get

TR2 =.1 R2 / D ˇO 0 T Var.x t /= 2 ˇO

D ˇO 0 V .ˇ/
O 1 ˇ:
O

This has the same form as (2.33) with R D I and q D 0 and J equal to the number of
slope coefficients.

51
2.2.5 A Joint Test of Several Coefficients: F-test

The joint test can also be cast in terms of the F distribution (which may have better small
sample properties).
Divide (2.33) by J and replace V .ˇ/ O by the estimated covariance matrix VO .ˇ/. O This
P  1
O D O 2
is, for instance, from (2.19) VO .ˇ/ N
xi x 0 , but where we (as in reality) have
i D1 i
to estimate the variance of the residuals by the sample variance of the fitted residuals, O 2 .
This gives
 0 h i 1  
RˇO q O 0
RVO .ˇ/R RˇO q
 FJ;T k , where (2.35)
J
X  1
N
O D O
VO .ˇ/ 2 0
xi xi :
i D1

The test of the joint hypothesis that all k 1 slope coefficients in the regression (that is,
excluding the intercept) are zero can be written (assuming your regression also contains
an intercept)
R2 =.k 1/
 Fk 1;T k : (2.36)
.1 R2 /=.T k/
Proof. (of (2.35)) Equation (2.35) can also be written
 0  P  1
 1  
RˇO q R 2 N
i D1 xi xi0 R 0
RˇO q =J
:
O 2 = 2

The numerator is a 2J variable divided by J . If the residuals are normally distributed, then
it can be shown that the denominator is a 2T k variable divided by T k. If the numerator
and denominator are independent (which requires that the residuals are independent of the
regressors), then the ratio has an FJ;T k distribution.

Example 2.23 (Joint F test) Continuing Example 2.22, and assuming that VO .ˇ/ O D
O we have a test statistic of 10=2 D 5. Assume T k D 50, then the 10% criti-
V .ˇ/,
cal value (from an F2;50 distribution) is 2.4, so the null hypothesis is rejected at the 10%
level.

52
 Scatter plot, iid residuals Scatter plot, Var(residual) depends on x2

20 20
10 10
0 0
y

y
−10 −10
−20 −20
−10 −5 0 5 10 −10 −5 0 5 10
x x

y = 0.03 + 1.3x + u
Solid regression lines are based on all data,
dashed lines exclude the crossed out data point

Figure 2.16: Effect of heteroskedasticity on uncertainty about regression line

2.3 Heteroskedasticity

Suppose we have a regression model

y t D x t0 b C u t ; where E u t D 0 and Cov.xi t ; u t / D 0: (2.37)

In the standard case we assume that u t is iid (independently and identically distributed),
which rules out heteroskedasticity.
In case the residuals actually are heteroskedastic, least squares (LS) is nevertheless a
useful estimator: it is still consistent (we get the correct values as the sample becomes
really large)—and it is reasonably efficient (in terms of the variance of the estimates).
However, the standard expression for the standard errors (of the coefficients) is (except in
a special case, see below) not correct. This is illustrated in Figure 2.17.
To test for heteroskedasticity, we can use White’s test of heteroskedasticity. The null
hypothesis is homoskedasticity, and the alternative hypothesis is the kind of heteroskedas-
ticity which can be explained by the levels, squares, and cross products of the regressors—
clearly a special form of heteroskedasticity. The reason for this specification is that if the
squared residual is uncorrelated with these squared regressors, then the usual LS covari-
ance matrix applies—even if the residuals have some other sort of heteroskedasticity (this
is the special case mentioned before).

53
 Std of LS slope coefficient under heteroskedasticity
0.1
OLS formula
White’s
Simulated
0.08

0.06

0.04 Model: yt = 0.9xt + ǫt ,

where ǫt ∼ N (0, ht ), with ht = 0.5exp(αx2t )

bLS is the LS estimate of b in

0.02 yt = a + bxt + ut

Number of simulations: 25000

0
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2
α (effect of regressor on variance)

Figure 2.17: Variance of OLS estimator, heteroskedastic errors

To implement White’s test, let wi be the squares and cross products of the regressors.
The test is then to run a regression of squared fitted residuals on w t

uO 2t D w t0 C v t ; (2.38)

and to test if all the slope coefficients (not the intercept) in are zero. (This can be done
be using the fact that TR2 =.1 R2 /  p2 , p D dim.wi / 1:)

Example 2.24 (White’s test) If the regressors include .1; x1t ; x2t / then w t in (2.38) is the
2 2
vector (1; x1t ; x2t ; x1t ; x1t x2t ; x2t ). (Notice that the cross product of .1; x1t ; x2t / with 1
gives us the regressors in levels, not squares.)

There are two ways to handle heteroskedasticity in the residuals. First, we could use
some other estimation method than LS that incorporates the structure of the heteroskedas-
ticity. For instance, combining the regression model (2.37) with an ARCH structure of
the residuals—and estimate the whole thing with maximum likelihood (MLE) is one way.
As a by-product we get the correct standard errors—provided the assumed distribution

54
(in the likelihood function) is correct. Second, we could stick to OLS, but use another
expression for the variance of the coefficients: a heteroskedasticity consistent covariance
matrix, among which “White’s covariance matrix” is the most common.
To understand the construction of White’s covariance matrix, recall that the variance
O
of ˇ1 is found from
1
ˇO1 D ˇ1 C PT .x1 u1 C x2 u2 C : : : xT uT / : (2.39)
t D1 x t x t

This gives
1 1
Var.ˇO1 / D PT Var .x1 u1 C x2 u2 C : : : xT u t / PT
t D1 x t x t t D1 x t x t
1 1
D PT x12 Var .u1 / C x22 Var .u2 / C : : : xT2 Var .uT / PT


t D1 x t x t t D1 x t x t
1 PT 2 2 1
D PT t D1 x t  t PT ; (2.40)
t D1 x t x t t D1 x t x t

where the second line assumes that the residuals are uncorrelated. This expression cannot
be simplified further since  t is not constant—and also related to x t2 . The idea of White’s
estimator is to estimate TtD1 x t2  t2 by TtD1 x t x t0 uO 2t (which also allows for the case with
P P

several elements in x t , that is, several regressors).

It is straightforward to show that the standard expression for the variance underes-
timates the true variance when there is a positive relation between x t2 and  t2 (and vice
versa). The intuition is that much of the precision (low variance of the estimates) of OLS
comes from data points with extreme values of the regressors: think of a scatter plot and
notice that the slope depends a lot on fitting the data points with very low and very high
values of the regressor. This nice property is destroyed if the data points with extreme
values of the regressor also have lots of noise (high variance of the residual).

Remark 2.25 (Standard OLS vs White’s variance) If x t2 is not related to  t2 , then we

could write the last term in (2.40) as
PT 2 2 1 PT 2
PT 2
t D1 x t  t D t D1  t tD1 x t
T
D  2 TtD1 x t2
P

PT 2
where  2 is the average variance, typically estimated as t D1 u t =T . That is, it is the

55
same as for standard OLS. Notice that
PT 2 2 1 PT 2
PT 2
t D1 x t  t > t D1  t t D1 x t
T
if x t2 is positively related to  t2 (and vice versa). For instance, with .x12 ; x22 / D .10; 1/
and .12 ; 22 / D .5; 2/, TtD1 x t2  t2 D 10  5 C 1  2 D 52 while T1 TtD1  t2 TtD1 x t2 D
P P P
1
2
.5 C 2/.10 C 1/ D 38:5:

2.4 Autocorrelation

Autocorrelation of the residuals (Cov.u t u t s / ¤ 0) is also a violation of the iid assump-

tions underlying the standard expressions for the variance of ˇO1 . In this case, LS is (typi-
cally) still consistent (exceptions: when the lagged dependent variable is a regressor), but
the variances are (again) wrong. In particular, not even the the first line of (2.40) is true,
since the variance of the sum in (2.39) depends also on the covariance terms.
The typical effect of positively autocorrelated residuals is to increase the uncertainty
about the OLS estimates—above what is indicated by the standard error calculated on the
iid assumptions. This is perhaps easiest to understand in the case of estimating the mean
of a data series, that is, when regressing a data series on a constant only. If the residual is
positively autocorrelated (have long swings), then the sample mean can deviate from the
true mean for an extended period of time—the estimate is imprecise. See Figure 2.18 for
an illustration.
There are several straightforward tests of autocorrelation—all based on using the fitted
residuals. The null hypothesis is no autocorrelation. First, estimate the autocorrelations
of the fitted residuals as

s D Corr.uO t ; uO t s /, s D 1; :::; L: (2.41)

p
Second, test the autocorrelation s by using the fact that T Os has a standard normal
distribution (in large samples)
p
T Os  N.0; 1/: (2.42)

An alternative for testing the first autocorrelation coefficient is the Durbin-Watson. The
test statistic is (approximately)
DW  2 2O1 ; (2.43)

56
 Scatter plot, iid residuals Scatter plot, autocorrelated residuals

20 20

0 0
y

y
−20 −20

−10 −5 0 5 10 −10 −5 0 5 10
x x

y = 0.03 + 1.3x + u
Solid regression lines are based on all data,
dashed lines are based on late sample (high x values).
The regressor is (strongly) autocorrelated, since
it is an increasing series (-10,-9.9,...,10).

Figure 2.18: Effect of autocorrelation on uncertainty about regression line

and the null hypothesis is rejected in favour of positive autocorrelation if DW<1.5 or so

(depending on sample size and the number of regressors). To extend (2.42) to higher-order
autocorrelation, use the Box-Pierce test
L
X
QL D T Os2 !d L
2
: (2.44)
sD1

If there is autocorrelation, then we can choose to estimate a fully specified model (in-
cluding how the autocorrelation is generated) by MLE or we can stick to OLS but apply an
autocorrelation consistent covariance matrix—for instance, the “Newey-West covariance
matrix.”
To understand the Newey-West covariance matrix, notice that the first line of (2.40) is
still correct. However, there might be corrrelation across time periods, so the second line
needs to account for terms like Cov.x t u t ; x t s u t s /. For instance, for T D 3 the middle
term of that second line is

Var .x1 u1 C x2 u2 C x3 u3 / D Var.x1 u1 / C Var.x2 u2 / C Var.x3 u3 /C

2 Cov.x2 u2 ; x1 u1 / C 2 Cov.x3 u3 ; x2 u2 / C 2 Cov.x3 u3 ; x1 u1 /:
(2.45)

57
 Autocorrelation of ut Autocorrelation of xt ut

κ = −0.9 κ = −0.9
0.5 κ=0 0.5 κ=0
κ = 0.9 κ = 0.9
0 0

−0.5 −0.5

−0.5 0 0.5 −0.5 0 0.5

ρ ρ

Model: yt = 0.9xt + ǫt ,
where ǫt = ρǫt−1 + ξt , ξt is iid N
xt = κxt−1 + ηt , ηt is iid N

ut is the residual from LS estimate of

yt = a + bxt + ut

Number of simulations: 25000

Figure 2.19: Autocorrelation of x t u t when u t has autocorrelation 

When data is uncorrelated across time (observations), then all the covariance terms are
zero. With autocorrelation, they may not be. For a general T , the middle term becomes
PT
Var .x t u t / C 2 m
P PT
tD1 sD1 t DsC1 Cov .x t u t ; x t s u t s / ; (2.46)

where m denotes the number of covariance terms that might be non-zero (at most, m D
T 1).
The idea of the Newey-West estimator is to estimate (2.46). For instance, with only
one lag (m D 1) the calculation is (with several regressors) TtD1 x t x t0 uO 2t +
P
PT 0 0
t D2 x t x t 1 C x t 1 x t u O t uO t 1 . Notice also that by exclduing all lags (setting m D 0),

the Newey-West estimator concides with White’s estimator. Hence, Newey-West estima-
tor handles also heteroskedasticity.
It is clear from this expression that what really counts is not so much the autocorre-
lation in u t per se, but the autocorrelation of x t u t . If this is positive, then the standard
expression underestimates the true variance of the estimated coefficients (and vice versa).
For instance, the autocorrelation of x t u t is likely to be positive when both the residual and
the regressor are positively autocorrelated. Notice that a constant, x t D 1 is extremely

58
 Std of LS slope under autocorrelation Std of LS slope under autocorrelation
0.1 κ = −0.9 0.1 κ = 0
OLS formula
Newey-West
Simulated
0.05 0.05

0 0
−0.5 0 0.5 −0.5 0 0.5
ρ (autocorrelation of residual) ρ (autocorrelation of residual)

Model: yt = 0.9xt + ǫt ,
Std of LS slope under autocorrelation
where ǫt = ρǫt−1 + ξt , ξt is iid N
0.1 κ = 0.9
xt = κxt−1 + ηt , ηt is iid N

ut is the residual from LS estimate of

yt = a + bxt + ut
0.05
NW uses 15 lags
Number of simulations: 25000
0
−0.5 0 0.5
ρ (autocorrelation of residual)

Figure 2.20: Standard error of OLS slope, autocorrelated errors

positively autocorrelated. In contrast, when the regressor has no autocorrelation, then the
product does not either. This is illustrated in Figures 2.19–2.21.
Figures 2.22–2.23 are empirical examples of the importance of using the Newey-West
method rather than relying of the iid assumptions. In both cases, the residuals have strong
positive autocorrelation.

A A Primer in Matrix Algebra

Let c be a scalar and define the matrices
" # " # " # " #
x1 z1 A11 A12 B11 B12
xD ;z D ;A D , and B D :
x2 z2 A21 A22 B21 B22

Adding/subtracting a scalar to a matrix or multiplying a matrix by a scalar are both

59
 Std of LS intercept under autocorrelation Std of LS intercept under autocorrelation
0.1 κ = −0.9 0.1 κ = 0

0.05 0.05
OLS formula
Newey-West
Simulated
0 0
−0.5 0 0.5 −0.5 0 0.5
ρ (autocorrelation of residual) ρ (autocorrelation of residual)

Model: yt = 0.9xt + ǫt ,
Std of LS intercept under autocorrelation
where ǫt = ρǫt−1 + ξt , ξt is iid N
0.1 κ = 0.9
xt = κxt−1 + ηt , ηt is iid N

ut is the residual from LS estimate of

yt = a + bxt + ut
0.05
NW uses 15 lags
Number of simulations: 25000
0
−0.5 0 0.5
ρ (autocorrelation of residual)

Figure 2.21: Standard error of OLS intercept, autocorrelated errors

element by element
" # " #
A11 A12 A11 C c A12 C c
Cc D
A21 A22 A21 C c A22 C c
" # " #
A11 A12 A11 c A12 c
cD :
A21 A22 A21 c A22 c

Example A.1
" # " #
1 3 11 13
C 10 D
3 4 13 14
" # " #
1 3 10 30
10 D :
3 4 30 40

60
 Slope (b) in rt = a + brt−1 + ǫt Autocorrelation of residual
0.5 1

0 0.5

−0.5 0
0 10 20 0 10 20
Return horizon (months) Return horizon (months)

Slope with two different 90% conf band, OLS and NW std

Monthly US stock returns 1926:1-2012:12, overlapping data

Figure 2.22: Slope coefficient, LS vs Newey-West standard errors

Expectations hypothesis: changes

Overlapping US 12−month interest rates and
Change in avg short rate

5
next−year average federal funds rate: 1970:1−2012:3

Slope coefficient: 0.54

Std (classical and Newey−West): 0.04 0.13
0 Autocorrelation of residual: 0.96

−5
−5 0 5
Change in interest rate

Figure 2.23: US 12-month interest and average federal funds rate (next 12 months)

Matrix addition (or subtraction) is element by element

" # " # " #
A11 A12 B11 B12 A11 C B11 A12 C B12
ACB D C D :
A21 A22 B21 B22 A21 C B21 A22 C B22

Example A.2 (Matrix addition and subtraction/

" # " # " #
10 2 8
D
11 5 6
" # " # " #
1 3 1 2 2 5
C D
3 4 3 2 6 2

61
 To turn a column into a row vector, use the transpose operator like in x 0
" #0
0 x1 h i
x D D x1 x2 :
x2

Similarly, transposing a matrix is like flipping it around the main diagonal

" #0 " #
A 11 A 12 A 11 A 21
A0 D D :
A21 A22 A12 A22

Example A.3 (Matrix transpose)

" #0
10 h i
D 10 11
11
2 3
" #0 1 4
1 2 3
D 42 55
6 7
4 5 6
3 6

Matrix multiplication requires the two matrices to be conformable: the first matrix
has as many columns as the second matrix has rows. Element ij of the result is the
multiplication of the ith row of the first matrix with the j th column of the second matrix
" #" # " #
A11 A12 B11 B12 A11 B11 C A12 B21 A11 B12 C A12 B22
AB D D :
A21 A22 B21 B22 A21 B11 C A22 B21 A21 B12 C A22 B22

Multiplying a square matrix A with a column vector z gives a column vector

" #" # " #
A11 A12 z1 A11 z1 C A12 z2
Az D D :
A21 A22 z2 A21 z1 C A22 z2

Example A.4 (Matrix multiplication)

" #" # " #
1 3 1 2 10 4
D
3 4 3 2 15 2
" #" # " #
1 3 2 17
D
3 4 5 26

62
 For two column vectors x and z, the product x 0 z is called the inner product
" #
h i z
1
x 0 z D x1 x2 D x1 z1 C x2 z2 ;
z2

and xz 0 the outer product

" # " #
x1
h i x z
1 1 x z
1 2
xz 0 D z1 z2 D :
x2 x2 z1 x2 z2

(Notice that xz does not work). If x is a column vector and A a square matrix, then the
product x 0 Ax is a quadratic form.

Example A.5 (Inner product, outer product and quadratic form )

" #0 " # " #
10 2 h i 2
D 10 11 D 75
11 5 5
" # " #0 " # " #
10 2 10 h i 20 50
D 2 5 D
11 5 11 22 55
" #0 " #" #
10 1 3 10
D 1244:
11 3 4 11

A matrix inverse is the closest we get to “dividing” by a matrix. The inverse of a

matrix A, denoted A 1 , is such that

1
AA D I and A 1 A D I;

where I is the identity matrix (ones along the diagonal, and zeroes elsewhere). The matrix
inverse is useful for solving systems of linear equations, y D Ax as x D A 1 y.

Example A.6 (Matrix inverse) We have

" #" # " #
4=5 3=5 1 3 1 0
D , so
3=5 1=5 3 4 0 1
" # 1 " #
1 3 4=5 3=5
D :
3 4 3=5 1=5

63
A Statistical Tables

n Critical values
10% 5% 1%
10 1:81 2:23 3:17
20 1:72 2:09 2:85
30 1:70 2:04 2:75
40 1:68 2:02 2:70
50 1:68 2:01 2:68
60 1:67 2:00 2:66
70 1:67 1:99 2:65
80 1:66 1:99 2:64
90 1:66 1:99 2:63
100 1:66 1:98 2:63
Normal 1:64 1:96 2:58

Table A.1: Critical values (two-sided test) of t distribution (different degrees of freedom)
and normal distribution.

Bibliography
Verbeek, M., 2008, A guide to modern econometrics, Wiley, Chichester, 3rd edn.

64
 n Critical values
10% 5% 1%
1 2:71 3:84 6:63
2 4:61 5:99 9:21
3 6:25 7:81 11:34
4 7:78 9:49 13:28
5 9:24 11:07 15:09
6 10:64 12:59 16:81
7 12:02 14:07 18:48
8 13:36 15:51 20:09
9 14:68 16:92 21:67
10 15:99 18:31 23:21

Table A.2: Critical values of chisquare distribution (different degrees of freedom, n).

n1 n2 2n1 =n1
10 30 50 100 300
1 4:96 4:17 4:03 3:94 3:87 3:84
2 4:10 3:32 3:18 3:09 3:03 3:00
3 3:71 2:92 2:79 2:70 2:63 2:60
4 3:48 2:69 2:56 2:46 2:40 2:37
5 3:33 2:53 2:40 2:31 2:24 2:21
6 3:22 2:42 2:29 2:19 2:13 2:10
7 3:14 2:33 2:20 2:10 2:04 2:01
8 3:07 2:27 2:13 2:03 1:97 1:94
9 3:02 2:21 2:07 1:97 1:91 1:88
10 2:98 2:16 2:03 1:93 1:86 1:83

Table A.3: 5% Critical values of Fn1;n2 distribution (different degrees of freedom).

65
 n1 n2 2n1 =n1
10 30 50 100 300
1 3:29 2:88 2:81 2:76 2:72 2:71
2 2:92 2:49 2:41 2:36 2:32 2:30
3 2:73 2:28 2:20 2:14 2:10 2:08
4 2:61 2:14 2:06 2:00 1:96 1:94
5 2:52 2:05 1:97 1:91 1:87 1:85
6 2:46 1:98 1:90 1:83 1:79 1:77
7 2:41 1:93 1:84 1:78 1:74 1:72
8 2:38 1:88 1:80 1:73 1:69 1:67
9 2:35 1:85 1:76 1:69 1:65 1:63
10 2:32 1:82 1:73 1:66 1:62 1:60

Table A.4: 10% Critical values of Fn1;n2 distribution (different degrees of freedom).

66
3 Regression Diagnostics

3.1 Misspecifying the Set of Regressors

Excluding a relevant regressor will cause a bias of all coefficients (unless those regres-
sors are uncorrelated with the excluded regressor). In contrast, including an irrelevant
regressor is tot really dangerous, but is likely to decrease the precision.
To selecting the regressors, apply the following rules: rule 1: use economic theory;
rule 2: avoid data mining and mechanical searches for the right regressors; rule 3: maybe
use a general-to-specific approach—start with a general and test restrictions,..., keep mak-
ing it simpler until restrictions are rejected.
Remember that R2 can never decrease by adding more regressors—not really a good
guide. To avoid overfitting, “punish” models with to many parameters. Perhaps consider
RN 2 instead
T 1
RN 2 D 1 .1 R2 / ; (3.1)
T k
where k is the number of regressors (including the constant). This measure includes
trade-off between fit and the number of regressors (per data point). Notice that RN 2 can be
negative (while 0  R2  1). Alternatively, apply Akaike’s Information Criterion (AIC)
and the Bayesian information criterion (BIC) instead. They are

k
AIC D ln O 2 C 2 (3.2)
T
k
BIC D ln O 2 C ln T: (3.3)
T
These measure also involve a trade-off between fit (low O 2 ) and number of parameters (k,
including the intercept). Choose the model with the highest RN 2 or lowest AIC or BIC. It
can be shown (by using R2 D 1  2 = Var.y t / so  2 D Var.y t /.1 R2 /) that AIC and

67
BIC can be rewritten as
k
AIC D ln Var.y t / C ln.1 R2 / C 2 (3.4)
T
k
BIC D ln Var.y t / C ln.1 R2 / C ln T: (3.5)
T
This shows that both are decreasing in R2 (which is good), but increasing in the number
of regressors per data point (k=T ). It therefore leads to a similar trade-off as in RN 2 .

3.2 Comparing Non-Nested Models

When one model is not a special case of another

Model A: y t D x t0 ˇ C " t (3.6)

Model B: y t D z t0 C v t (3.7)

Non-nested if z is not a subset of x at the same time as x is not a subset of z. For instance,
these models could represent alternative economic theories of the same phenomenon.
Comparing the fit of these models starts with the usual criteria: R2 , RN 2 , AIC, and BIC.
An alternative approach to compare the fit is to study encompassing. Model B is said
to encompass model A if it can explain all that model A can (and more). To test this, run
the regression
y t D z t0 C x2t
0
ıA C v t ; (3.8)

where x2t are those variables in x t that are not also in z t . Model B encompasses model
A if ıA D 0 (test this restriction). Clearly, we can repeat this to see if A encompasses B.

3.3 Non-Linear Models

Regression analysis typically start with assuming a linear model—which may or may not
be a good approximation.
Notice that models that are non-linear in variables

y t D a C bx t3:4 C " t ; (3.9)

can be handled by OLS: just run OLS using x t3:4 as a regressor.

68
 In contrast, models that are non-linear in parameters

y t D ˇ1 C ˇ2 x tˇ3 C u t (3.10)

cannot be estimated by OLS. Do nonlinear LS (NLS) instead. This requires the use of a
numerical minimization routine to minimize the sum of squared residuals, TtD1 u2t .
P

To test the functional form (...is a linear specification really correct?), estimate non-
linear extension and test if they are significant. Alternatively, do a RESET test

y t D x t0 ˇ C ˛2 yO t2 C v t ; (3.11)

where yO t D x t0 bO (from linear estimation). Test if ˛2 D 0.

3.4 Outliers

LS is sensitive to extreme data points. Maybe we need to understand if there are outliers,
by plotting data and some regression results.
The starting point (as always in empirical work) is to plot the data: time series plots
and histograms—to see if there are extreme data points.
As complement, it is a good idea to try to identify outliers from the regression results.
First, estimate on whole sample to get the estimates of the coefficients ˇO and the fitted
values yO t . Second, estimate on the whole sample, except observation s: and record the
estimates ˇO .s/ and the fitted value for period s (the one that was not used in the estimation)
yOs.s/ D xs0 ˇO .s/ . Repeat this for all data points (s). Third, plot ˇO .s/ ˇ,
O yOs.s/ yO t or uO .s/ O
s =.

If these series make sudden jumps, then that data point is driving the results for the full
sample. It then remains to determine whether this is good (a very informative data point)
or bad (unrepresentative or even wrong data point).

3.5 Estimation on Subsamples

To test for a structural break of (one or more) coefficients, add a dummy for a subsample
and interact it with the those regressors that we suspect have structural breaks (denoted

69
 OLS: sensitivity to outlier
2
y: -1.125 -0.750 1.750 1.125
x: -1.500 -1.000 1.000 1.500

1
Three data points are on the
line y = 0.75x, while the
fourth has a big error
0
y

−1
Data
OLS (0.25 0.90)
True (0.00 0.75)
−2
−3 −2 −1 0 1 2 3
x

Figure 3.1: Data and regression line from OLS

zt )

y t D x t0 ˇ C g t z t0 C " t , where (3.12)

(
1 for some subsample
gt D (3.13)
0 else

and test D 0 (a “Chow test”). Notice that measures the change of the coefficients
(from one sub sample to another)..
To capture time-variation in the regression coefficients, it is fairly common to run the
regression
y t D x t0 ˇ C " t (3.14)

on a longer and longer data set (“recursive estimation”). In the standard recursive es-
timation, the first estimation is done on the sample t D 1; 2; : : : ; ; while the second
estimation is done on t D 1; 2; : : : ; ;  C 1; and so forth until we use the entire sample
t D 1 : : : ; T . In the “backwards recursive estimate” we instead keep the end-point fixed
and use more and more of old data. That is, the first sample could be T ; : : : ; T ; the
second T  1; : : : ; T ; and so forth.
Alternatively, a moving data window (“rolling samples”) could be used. In this case,

70
the first sample is t D 1; 2; : : : ;  ; but the second is on t D 2; : : : ; ;  C 1, that is, by
dropping one observation at the start when the sample is extended at the end. See Figure
3.2 for an illustration.
An alternative is to apply an exponentially weighted moving average (EMA) esti-
mator, which uses all data points since the beginning of the sample—but where recent
observations carry larger weights. The weight for data in period t is T t where T is the
latest observation and 0 <  < 1, where a smaller value of  means that old data carries
low weights. In practice, this means that we define

xQ t D x t T t
and yQ t D y t T t
(3.15)

and then estimate

yQ t D xQ t0 ˇ C " t : (3.16)

Notice that also the constant (in x t ) should be scaled in the same way. (Clearly, this
method is strongly related to the GLS approach used when residuals are heteroskedastic.
Also, the idea of down weighting old data is commonly used to estimate time-varying
volatility of returns as in the RISK metrics method.)
Estimation on subsamples is not only a way of getting a more recent/modern estimate,
but also a way to gauge the historical range and volatility in the betas—which may be
important for putting some discipline on judgemental forecasts.
See Figures 3.2–3.3 for an illustration.
From the estimations on subsamples (irrespective of method), it might be informative
to study plots of (a) residuals with confidence band (0˙2 standard errors) or standardized
residuals with confidence band (0 ˙ 2) and (b) coefficients with confidence band (˙2
standard errors). In these plots, the standard errors are typically from the subsamples.
The recursive estimates can be used to construct another formal test of structural
breaks, the CUSUM test (see, for instance, Enders (2004)). First, do a regression on
the sample t D 1; 2; : : : ;  and use the estimated coefficients (denoted ˇ . / ) to calculate a
“forecast” and “forecast error” for  C 1 as

yO C1 D x0 C1 ˇ . / and v C1 D y C1 yO C1 : (3.17)

71
 β of HiTech sector, recursive β of HiTech sector, backwards recursive
2 2

1.5 1.5

1 1
1960 1980 2000 1960 1980 2000
end of sample start of sample

β of HiTech sector, 5−year data window β of HiTech sector, EWMA estimate

2 2

1.5 1.5

1 1
1960 1980 2000 0.9 0.95 1
end of 5−year sample λ

Figure 3.2: Betas of US industry portfolios

Second, do a second estimation on the sample t D 1; 2; : : : ;  C 1 and calculate

yO C2 D x0 C2 ˇ . C1/ and v C2 D yC2 yO C2 : (3.18)

Third, do the same for all other samples (observation 1 to  C 2, observation 1 to  C 3,

etc). Forth, calculate the standard deviation of those forecast errors (denoted  below).
Fifth, calculate a corresponding sequence of cumulative sums of standardized residuals
v C1
W D

v C1 C v C2
W C1 D

v C1 C v C2 C v C3
W C2 D (3.19)


72
 Distribution of betas estimated on moving 5−year data windows

U.S. monthly data 1947:1−2012:12

Monthly returns, %

NoDur Durbl Manuf Enrgy HiTec Telcm Shops Hlth Utils Other

Figure 3.3: Distribution of betas of US industry portfolios (estimated on 5-year data win-
dows)

and so forth. More generally we have the sequence

t
X vsC1
Wt D , for t D ; :::; T 1: (3.20)
sD

p p 
Sixth and finally, plot W t along with a 95% confidence interval: ˙0:948 T  C 2 .t / = T  .
Reject stability if any observation is outside.

3.6 Robust Estimation

3.6.1 Robust Means, Variances and Correlations

Outliers and other extreme observations can have very decisive influence on the estimates
of the key statistics needed for financial analysis, including mean returns, variances, co-
variances and also regression coefficients.
The perhaps best way to solve these problems is to carefully analyse the data—and

73
 Recursive estimate of slope Recursive residuals
5
2.6
2.4
2.2 0
2
1.8
−5
0 50 100 150 200 0 50 100 150 200
period period

CUSUM statistics and 95% confidence band

Model: yt = βt xt + ut ,
βt = 2 for t = 1 − 150 and βt = 6 for t = 151−
50 where xt ∼ N (2, 1) and ǫt ∼ N (0, 1)

0 Estimated: yt = a + bxt + ut

−50

0 50 100 150 200

period

Figure 3.4: Stability test

then decide which data points to exclude. Alternatively, robust estimators can be applied
instead of the traditional ones.
To estimate the mean, the sample average can be replaced by the median or a trimmed
mean (where the x% lowest and highest observations are excluded).
Similarly, to estimate the variance, the sample standard deviation can be replaced by
the interquartile range (the difference between the 75th and the 25th percentiles), divided
by 1:35
StdRobust D Œquantile.0:75/ quantile.0:25/=1:35; (3.21)

or by the median absolute deviation

StdRobust D median.jx t j/=0:675: (3.22)

Both these would coincide with the standard deviation if data was indeed drawn from a

74
 β of HiTec Recursive std residuals
2 moving 4-year sample 2
recursive
1
beta

1.5 0
−1
1
−2
1970 1980 1990 2000 2010 1970 1980 1990 2000 2010

CUSUM statistics and 90% conf band

50

−50
1970 1980 1990 2000 2010

Figure 3.5: CAPM regression on a US industry index

normal distribution without outliers.

A robust covariance can be calculated by using the identity

Cov.x; y/ D ŒVar.x C y/ Var.x y/=4 (3.23)

and using a robust estimator of the variances—like the square of (3.21). A robust cor-
relation is then created by dividing the robust covariance with the two robust standard
deviations.
See Figures 3.6–3.7 for empirical examples.

3.6.2 Robust Regression Coefficients

Reference: Amemiya (1985) 4.6

The least absolute deviations (LAD) estimator miminizes the sum of absolute residu-

75
 US industry portfolios, ERe
Monthly data 1947:1−2012:12
mean
0.11 median
mean excess return

0.1

0.09

0.08

0.07

0.06
A B C D E F G H I J

Figure 3.6: Mean excess returns of US industry portfolios

als (rather than the squared residuals)

T
ˇOLAD D arg min
X
x t0 b ˇ (3.24)
ˇ ˇ
ˇy t
b
t D1

This estimator involve non-linearities, but a simple iteration works nicely. It is typically
less sensitive to outliers. (There are also other ways to estimate robust regression coeffi-
cients.) This is illustrated in Figure 3.8.
See Figure 3.9 for an empirical example.
If we assume that the median of the true residual, u t , is zero, then we (typically) have
p XT
T .ˇOLAD ˇ0 / !d N 0; f .0/ 2 ˙xx1 =4 , where ˙xx D plim x t x t0 =T; (3.25)
 
tD1

where f .0/ is the value of the pdf of the residual at zero. Unless we know this density
function (or else we would probably have used MLE instead of LAD), we need to estimate
it—for instance with a kernel density method.
p
Example 3.1 (N.0;  2 /) When u t  N.0;  2 ), then f .0/ D 1= 2 2 , so the covari-
ance matrix in (3.25) becomes  2 ˙xx1 =2. This is =2 times larger than when using
LS.

76
 US industry portfolios, Std
Monthly data 1947:1−2012:12
std
0.2 iqr/1.35

0.18
std

0.16

0.14

0.12
A B C D E F G H I J

Figure 3.7: Volatility of US industry portfolios

Remark 3.2 (Algorithm for LAD) The LAD estimator can be written
T
ˇOLAD D arg min x t0 bO
X
w t uO t .b/2 , w t D 1= juO t .b/j ; with uO t .b/ D y t
ˇ
t D1

so it is a weighted least squares where both y t and x t are multiplied by 1= juO t .b/j. It can
be shown that iterating on LS with the weights given by 1= juO t .b/j, where the residuals
are from the previous iteration, converges very quickly to the LAD estimator.

Some alternatives to LAD: least median squares (LMS), and least trimmed squares
(LTS) estimators which solve

ˇOLM S D arg min median uO 2t , with uO t D y t x t0 bO (3.26)



ˇ
h
ˇOLT S D arg min
X
uO 2i , uO 21  uO 22  ::: and h  T: (3.27)
ˇ
i D1

Note that the LTS estimator in (3.27) minimizes the sum of the h smallest squared resid-
uals.

77
 OLS vs LAD of y = 0.75x + u
2
y: -1.125 -0.750 1.750 1.125
x: -1.500 -1.000 1.000 1.500

0
y

−1
Data
OLS (0.25 0.90)
LAD (0.00 0.75)
−2
−3 −2 −1 0 1 2 3
x

Figure 3.8: Data and regression line from OLS and LAD

Bibliography
Amemiya, T., 1985, Advanced econometrics, Harvard University Press, Cambridge, Mas-
sachusetts.

Enders, W., 2004, Applied econometric time series, John Wiley and Sons, New York, 2nd
edn.

78
 US industry portfolios, β
1.5
Monthly data 1947:1−2012:12
OLS
LAD

1
β

0.5
A B C D E F G H I J

Figure 3.9: Betas of US industry portfolios

79
4 Asymptotic Results on OLS

4.1 Properties of the OLS Estimator when “Gauss-Markov” Is False

There are severakl problems when the Gauss-Marlov assumptions are wrong. First, the
result that E ˇO D ˇ (unbiased) relied on the assumption that the regressors are fixed or
(altermatively) that fu1 ; :::; uT g and fx1 ; :::; xT g are independent. Otherwise not true (in
a finite sample). Second, theresult that ˇO is normally distributed relied on residuals being
normally distributed. Otherwise not true (in a finite sample).
What is true when these assumptions are not satisfied? How should we test hypothe-
ses? Two ways to find answers: (a) do computer (Monte Carlo) simulations; (b) find
results for T ! 1 (“asymptotic properties”) and use as approximation.

4.2 Motivation of Asymptotics

The results from asymptotoc theiry are more general (and prettier) than simulations—
and can be used as approximation if sample is large. The basic reasons for this is that
most estimators are sample averages and sample averages often have nice properties as
T ! 1. In particular, we can make use of the law of large numbers (LLN) and the
central limit theorem (CLT). See Figure 4.2

4.3 Asymptotics: Consistency

Issue: will our estimator come closer to the truth as the sample size increases? If not, use
another estimator (method).
Consistency: if Prob.ˇO deviates much from ˇ/ ! 0 as T ! 1. (Notation: plim ˇO D
ˇ)
LLN: (simple version...) plim.x/ N D E.x/. OLS (and most other estimators) are
sample averages of some sort.
As a simple special case, suppose there is only one regressor and that both the depen-
dent variable and the regressor have zero means. The OLS estimate of the slope coefficient

80
 Distribution of LS estimator, T = 25 Distribution of LS estimator, T = 100
Mean and std: Mean and std:
0.1 0.74 0.16 0.1 0.86 0.06

0.05 0.05

0 0
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

True model: yt = 0.9yt−1 + ǫt , ǫt is iid N(0,2)

Estimated model: yt = a + ρyt−1 + ut
Number of simulations: 25000

Figure 4.1: Distribution of LS estimator of autoregressive parameter

is then
X  T
1X
T
ˇO D ˇ C xt xt xt ut (4.1)
t D1
t D1
  1 T
1 XT 1X
DˇC xt xt xt ut ; (4.2)
T tD1 T t D1

where u t are the residuals we could calculate if we knew the true slope coefficient, that
is, the true residuals.
This estimate has the probability limit

plim ˇO D ˇ C ˙xx1 E.x t u t /; (4.3)

where ˙xx1 is a matrix of constants. The key point: is E.x t u t / D 0. If not, OLS is not
consistent.
Some observations:

1. We can not (easily) test this. OLS creates ˇO and the fitted residuals uO t such that
PT
O t D 0.
t D1 x t u

2. The Gauss-Markov assumption that u t and xj are independent implies that E.x t u t / D
0, so the Gauss-Markov assumptions basically disregards the issue of consistency.

81
 √
Distribution of sample avg. Distribution of T × sample avg.
3 0.4
T =5
T = 25
T = 100 0.3
2
0.2
1
0.1

0 0
−2 −1 0 1 2 −5 0 5
Sample average
√
T × sample average

Sample average of z t − 1 where z t has a χ21 distribution

Figure 4.2: Distribution of sample averages

(Assumes that it does not exist.)

3. OLS can be biased, but still be consistent—so it is ok if sample is large. See Fig-
ures 4.1 and 4.3. Notice Cov.u t 1 ; x t / ¤ 0 so biased, but Cov.u t ; x t / D 0 so
consistent)

4. There are cases when E.x t u t / D 0 doesn’t make sense. More on this later.

5. See Figures 4.1 and 4.3 for examples where OLS is consistent, and Figure 4.4 when
it is not.

What have we learned? Well...under what conditions (E.x t u t / D 0) OLS comes closer
to the truth as T increases.

4.4 When LS Cannot be Saved

...not even in large samples (since it’s inconsistent)

Q. When do we have Corr.x; u/ ¤ 0?
A. Need to think hard...
But the usual suspects are (i) excluded variables; (ii) autorrelated errors combined with
lagged dependent variable; (iii) measurement errors in regressors; and (iv) endogenous
regressors.

82
 Distribution of LS estimate, T = 100 Distribution of LS estimate, T = 1000
30 30
mean 0.860 mean 0.896
std 0.057 std 0.014
20 20

10 10

0 0
0.8 0.9 1 0.8 0.9 1
slope estimate slope estimate

True model: yt = ρyt−1 + ǫt ,

where ρ = 0.9 and ǫt is iid N (0, 2)

Estimated model: yt = ρyt−1 + ut

Figure 4.3: Distribution of LS estimator of autoregressive parameter

4.4.1 When LS Cannot be Saved: Excluded Variables

Correct model for log wages

0
y t D x1t ˇ1 C x2t ˇ2 C a t C  t (4.4)

where x1t measure individual characteristics, x2t years of schooling and u t ability. Assume that
> 0 and Cov.x2t ; a t / > 0 (people with more ability tend to have longer schooling).
Suppose we cann measure ability and therefore estimate

0
y t D x1t ˇ1 C x2t ˇ2 C ut
„ƒ‚…
a t C t

D x t0 ˇ C ut : (4.5)

From (4.3) plim ˇO D ˇ C ˙xx1 E.x t u t /, so assuming E.x t  t / D 0 (that the residual in
(4.4) is not correlated with the regressors) gives

plim ˇO D ˇ C ˙xx1 E.x t u t /

D ˇ C ˙xx1 E.x t a t / : (4.6)

If x t and a t are related so E.x t a t / ¤ 0, then OLS in not consistent. For instance, if x t

83
 Distribution of LS estimate, T = 100 Distribution of LS estimate, T = 1000

40 mean 0.921 40 mean 0.942

std 0.035 std 0.008
30 30
20 20
10 10
0 0
0.8 0.9 1 0.8 0.9 1
slope estimate slope estimate

True model: yt = ρyt−1 + ǫt with ǫt = νt + θνt−1 ,

where ρ = 0.9, θ = 0.5 and νt is iid N (0, 2)

Estimated model: yt = ρyt−1 + ut

Figure 4.4: Distribution of LS estimator of autoregressive parameter

are zero mean variables, then E.x t a t / D Cov.x t ; a t /. In our example of how scholling
affects wages, Cov.x t ; a t / > 0 seems reasonable, so ˇO2 > ˇs . That is, the OLS estimate
is likely to overestimate the returns to schooling ˇ2 , sinnce the estimate ˇO2 captures also
the effect of the excluded ability. In contrast, excluding something that is uncorrelated
with other regressors does not create a problem.
Notice the following:

 ˇO is the right number to use if we want to predict: “given x t , what is the best guess
of y t ?” The reason is that ˇO factors in also how x t predicts u t (which clearly also
has an effect on y t ).

 ˇO2 is not the right number to use if we want to understand an economic mechanism:
“if we increase schooling, x2t , by one unit (but holding all other variables constant),
what is the likely effect on y t ?” The reason is that we here need ˇ2 (or at least a
consistent estimate of it).

4.4.2 When LS Cannot be Saved: Autocorrelated Errors Combined with Lagged

Dependent Variable

(macroeconomics)

84
 Suppose y t depends on lags of itself (inertia in the economy?), but the residual is
autoregressive

y t D ˇ1 C ˇ2 x t C ˇ3 y t 1 Cu t and (4.7)
„ƒ‚…
u t D  t C  t 1,  t iid. (4.8)

This leads to Cov.y t 1 ; u t / ¤ 0. To see why: Cov.y t 1 ; u t / D Cov.y t 1 ;  t C  t 1 / >

0 (if  > 0): a positive bias.
As a special case, ˇ2 D 0 gives an ARMA(1,1) model, which cannot be estimated by
OLS. See Figure 4.4.

4.4.3 When LS Cannot be Saved: Measurement Errors in a Regressor

(microeconomics)
Suppose the correct model

y t D ˇ1 C ˇ2 w t C  t ; (4.9)

but we estimate with proxy x t for w t

y t D ˇ1 C ˇ2 x t C u t , with (4.10)
xt D wt C et : (4.11)
„ƒ‚…
measurement error

In this equation e t is the (zero mean) measurement error—and we typically assume that
it is uncorrelated with the true value (w t )
This leads to Cov.x t ; u t / ¤ 0, so OLS is inconsistent. See Figure 4.5.
To see why, solve for w t D x t e t , use in correct model (4.9)

y t D ˇ1 C ˇ2 .x t et / C t
D ˇ1 C ˇ2 x t ˇ2 e t C  t (4.12)
„ ƒ‚ …
ut

and from (4.11) we know that x t is correlated with e t . In fact, it can be shown that

Var.e t /
 
O
plim ˇ2 D ˇ2 1 (4.13)
Var.w t / C Var.e t /

85
 Distribution of LS estimate, h = 0 Distribution of LS estimate, h = 0.5
Mean
std
1 2.0 1 1.3
0.4 0.4
0.5 0.5

0 0
−1 0 1 2 3 4 −1 0 1 2 3 4
slope estimate slope estimate

True model: yt = 1 + 2w t + vt ,
Distribution of LS estimate, h = 1 where w t is iid N (0, 0.5) and vt is iid N (0, 4)

Regression: yt = β1 + β2 xt + ut , with T = 50
1 where xt = w t + e t and e t is iid N (0, h)
0.7
0.3
0.5

0
−1 0 1 2 3 4
slope estimate

Figure 4.5: Effect of measurement error in regressor,h is the variance of the errors

Notice that ˇO2 ! 0 as measurement dominates (Var.e t / ! 1): y t is not related to the
measurement error. In contrast, ˇO2 ! ˇ2 as measurement vanishes (Var.e t / ! 0): no
measurement error.
Proof. (of (4.13)) To simplify, assume that x t has a zero mean. From (4.3), we
then have plim ˇO2 D ˇ2 C ˙xx1 E.x t u t /. Here, ˙xx1 D 1= Var.x t /, but notice from
(4.11) that Var.x t / D Var.w t / C Var.e t / if w t and e t are uncorrelated. We also have
E.x t u t / D Cov.x t ; u t /, which from the definition of x t in (4.11) and of u t in (4.12) gives

Cov.x t ; u t / D Cov.w t C e t ; ˇ2 e t C  t / D ˇ2 Var.e t /:

Together we get

Var.e t /
plim ˇO2 D ˇ2 C ˙xx1 E.x t u t / D ˇ2 ˇ2 ;
Var.w t / C Var.e t /

86
which is (4.13).

4.4.4 When LS Cannot be Saved: Endogenous Regressors (System of Simultaneous

Equations)

(micro and macro) A simplistic macro model

C t D ˇ1 C ˇ2 Y t C u t (4.14)
Yt D Ct C It (4.15)
I t and " t are independent (exogenous) (4.16)

Could be generalized to let Y t have more action.

Key point: u t ! C t ! Y t ! C t ) Cov.Y t ; u t / > 0: a positive bias if we try to
estimate the consumption equation (4.14).

4.5 Asymptotic Normality

Issue: what is the distribution of your estimator in large samples?

p
CLT: (simple version...) T xN  N./ when T becomes really large. Holds for most
random variables. Notice: the distribution of xN converges to a spike as T increases, but
p
the distribution of T xN converges to a nice normal. See Figure 4.2.
p
Subtract ˇ from both sides of (4.2), multiply both sides by T

p p 1 XT
  1
1 XT
T .ˇO ˇ/ D 0
xt xt T xt ut (4.17)
T t D1
„ T ƒ‚t D1 …
„ ƒ‚ … p
!˙xx1 T sample average

p
The first term converges (by a LLN) to a constant, while the second term is T sample
average (of x t u t ). We should therefore expect that ˇO is normally distributed in large
samples—even if the residual doesn’t have a normal distribution. See Figure 4.6 for an
example (expressed in terms of a t-stat).
If an estimator is consistent and asymptotically normal, then use the results as an
approximation in large samples
p
T .ˇO ˇ/ ! N 0;  2 ˙xx1 or “ˇO ! N ˇ;  2 ˙xx1 =T ” (4.18)

87
 Distribution of LS t-stat, T = 5 Distribution of LS t-stat, T = 100
t = (b̂ − 0.9)/Std(b̂)
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−4 −2 0 2 4 −4 −2 0 2 4

Model: Rt = 0.9ft + ǫt , ǫt = vt − 2,
where vt has a χ22 distribution Probability density functions

Estimated model: yt = a + bft + ut N(0,1)

0.4
Number of simulations: 25000 χ22 − 2
0.3
T = 5 T = 100 0.2
Kurtosis of t-stat: 46.753 3.049
Frequency of |t-stat| > 1.65 0.294 0.105 0.1
Frequency of |t-stat| > 1.96 0.227 0.054 0
−4 −2 0 2 4

Figure 4.6: Results from a Monte Carlo experiment with thick-tailed errors.

p p p
Remark 4.1 Step 1: If VarŒ T .ˇO ˇ/ D  2 ˙xx1 , then VarŒ T .ˇO ˇ/= T  D
 2 ˙xx1 =T ; step 2: if E.ˇO ˇ/ D 0, then E.ˇ/
O D ˇ.

Bibliography

88
5 Index Models
Reference: Elton, Gruber, Brown, and Goetzmann (2010) 7–8, 11

5.1 The Inputs to a MV Analysis

To calculate the mean variance frontier we need to calculate both the expected return and
variance of different portfolios (based on n assets). With two assets (n D 2) the expected
return and the variance of the portfolio are
" #
h i 
1
E.Rp / D w1 w2
2
" #" #
h i 2  w1
1 12
P2 D w1 w2 : (5.1)
12 22 w2

In this case we need information on 2 mean returns and 3 elements of the covariance
matrix. Clearly, the covariance matrix can alternatively be expressed as
" # " #
12 12 12 12 1 2
D ; (5.2)
12 22 12 1 2 22

which involves two variances and one correlation (as before, 3 elements).
There are two main problems in estimating these parameters: the number of parame-
ters increase very quickly as the number of assets increases and historical estimates have
proved to be somewhat unreliable for future periods.
To illustrate the first problem, notice that with n assets we need the following number
of parameters

Required number of estimates With 100 assets

i n 100
i i n 100
ij n.n 1/=2 4950

89
The numerics is not the problem as it is a matter of seconds to estimate a covariance
matrix of 100 return series. Instead, the problem is that most portfolio analysis uses
lots of judgemental “estimates.” These are necessary since there might be new assets
(no historical returns series are available) or there might be good reasons to believe that
old estimates are not valid anymore. To cut down on the number of parameters, it is
often assumed that returns follow some simple model. These notes will discuss so-called
single- and multi-index models.
The second problem comes from the empirical observations that estimates from his-
torical data are sometimes poor “forecasts” of future periods (which is what matters for
portfolio choice). As an example, the correlation between two asset returns tends to be
more “average” than the historical estimate would suggest.
A simple (and often used) way to deal with this is to replace the historical correla-
tion with an average historical correlation. For instance, suppose there are three assets.
Then, estimate ij on historical data, but use the average estimate as the “forecast” of all
correlations:
2 3 2 3
1 12 13 1 N N
estimate 4 1 23 5 , calculate N D .O12 C O13 C O23 /=3, and use 4 1 N5 :
6 7 6 7

1 1

5.2 Single-Index Models

The single-index model is a way to cut down on the number of parameters that we need
to estimate in order to construct the covariance matrix of assets. The model assumes that
the co-movement between assets is due to a single common influence (here denoted Rm )

Ri D ˛i C ˇi Rm C ei , where (5.3)
E.ei / D 0, Cov .ei ; Rm / D 0, and Cov.ei ; ej / D 0:

The first two assumptions are the standard assumptions for using Least Squares: the resid-
ual has a zero mean and is uncorrelated with the non-constant regressor. (Together they
imply that the residuals are orthogonal to both regressors, which is the standard assump-
tion in econometrics.) Hence, these two properties will be automatically satisfied if (5.3)
is estimated by Least Squares.
See Figures 5.1 – 5.3 for illustrations.

90
 CAPM regression: Ri − Rf = α + β(Rm − Rf ) + ei
10
Intercept (α) and slope (β):
8
2.0 1.3

Excess return asset i, % 6

2
α
0

−2

−4

−6 β is slope of line
−8 Data points
Regression line
−10
−10 −5 0 5 10
Market excess return, %

Figure 5.1: CAPM regression

The key point of the model, however, is the third assumption: the residuals for dif-
ferent assets are uncorrelated. This means that all comovements of two assets (Ri and
Rj , say) are due to movements in the common “index” Rm . This is not at all guaranteed
by running LS regressions—just an assumption. It is likely to be false—but may be a
reasonable approximation in many cases. In any case, it simplifies the construction of the
covariance matrix of the assets enormously—as demonstrated below.

Remark 5.1 (The market model) The market model is (5.3) without the assumption that
Cov.ei ; ej / D 0. This model does not simplify the calculation of a portfolio variance—but
will turn out to be important when we want to test CAPM.

If (5.3) is true, then the variance of asset i and the covariance of assets i and j are

i i D ˇi2 Var .Rm / C Var .ei / (5.4)

ij D ˇi ˇj Var .Rm / : (5.5)

91
 Scatter plot against market return Scatter plot against market return
30 30
US data
Excess return %, HiTec

Excess return %, Utils

20 1970:1-2012:12 20

10 10

0 0

−10 α -0.15
−10 α 0.24
−20 β 1.28 −20 β 0.52

−30 −30
−20 0 20 −20 0 20
Excess return %, market Excess return %, market

Figure 5.2: Scatter plot against market return

Together, these equations show that we can calculate the whole covariance matrix by
having just the variance of the index (to get Var .Rm /) and the output from n regressions
(to get ˇi and Var .ei / for each asset). This is, in many cases, much easier to obtain than
direct estimates of the covariance matrix. For instance, a new asset does not have a return
history, but it may be possible to make intelligent guesses about its beta and residual
variance (for instance, from knowing the industry and size of the firm).
This gives the covariance matrix (for two assets)
" #! " # " #
Ri ˇi2 ˇi ˇj Var.ei / 0
Cov D Var .Rm / C , or (5.6)
Rj ˇi ˇj ˇj2 0 Var.ej /
" # " #
ˇi h i Var.ei / 0
D ˇi ˇj Var .Rm / C (5.7)
ˇj 0 Var.ej /

More generally, with n assets we can define ˇ to be an n  1 vector of all the betas and ˙
to be an n  n matrix with the variances of the residuals along the diagonal. We can then
write the covariance matrix of the n  1 vector of the returns as

Cov.R/ D ˇˇ 0 Var .Rm / C ˙: (5.8)

See Figure 5.4 for an example based on the Fama-French portfolios detailed in Table
5.2.

92
 HiTec Utils
constant 0:15 0:24
. 1:00/ .1:58/
market return 1:28 0:52
.33:58/ .12:77/
R2 0:75 0:34
obs 516:00 516:00
Autocorr (t) 0:73 0:86
White 6:19 20:42
All slopes 386:67 176:89

Table 5.1: CAPM regressions, monthly returns, %, US data 1970:1-2012:12. Numbers

in parentheses are t-stats. Autocorr is a N(0,1) test statistic (autocorrelation); White is a
chi-square test statistic (heteroskedasticity), df = K(K+1)/2 - 1; All slopes is a chi-square
test statistic (of all slope coeffs), df = K-1

Remark 5.2 (Fama-French portfolios) The portfolios in Table 5.2 are calculated by an-
nual rebalancing (June/July). The US stock market is divided into 5  5 portfolios as
follows. First, split up the stock market into 5 groups based on the book value/market
value: put the lowest 20% in the first group, the next 20% in the second group etc. Sec-
ond, split up the stock market into 5 groups based on size: put the smallest 20% in the first
group etc. Then, form portfolios based on the intersections of these groups. For instance,
in Table 5.2 the portfolio in row 2, column 3 (portfolio 8) belong to the 20%-40% largest
firms and the 40%-60% firms with the highest book value/market value.

Book value/Market value

1 2 3 4 5
Size 1 1 2 3 4 5
2 6 7 8 9 10
3 11 12 13 14 15
4 16 17 18 19 20
5 21 22 23 24 25

Table 5.2: Numbering of the FF indices in the figures.

Proof. (of (5.4)–(5.5) By using (5.3) and recalling that Cov.Rm ; ei / D 0 direct calcu-

93
 US industry portfolios, β (against the market), 1970:1-2012:12
beta 1.5

0.5
NoDur Durbl Manuf Enrgy HiTec Telcm Shops Hlth Utils Other

Figure 5.3: ˇs of US industry portfolios

lations give

i i D Var .Ri /
D Var .˛i C ˇi Rm C ei /
D Var .ˇi Rm / C Var .ei / C 2  0
D ˇi2 Var .Rm / C Var .ei / :

Similarly, the covariance of assets i and j is (recalling also that Cov ei ; ej D 0)

ij D Cov Ri ; Rj

D Cov ˛i C ˇi Rm C ei ; ˛j C ˇj Rm C ej

D ˇi ˇj Var .Rm / C 0
D ˇi ˇj Var .Rm / :

94
 Correlations, data Difference in correlations: data − model

1 0.5

0.5 0

0 −0.5
25 20 25 20
15 10 20 25 15 10 20 25
5 5 10 15 5 5 10 15
Portfolio Portfolio

25 FF US portfolios, 1957:1−2012:12
Index (factor): US market

Figure 5.4: Correlations of US portfolios

5.3 Estimating Beta

5.3.1 Estimating Historical Beta: OLS and Other Approaches

Least Squares (LS) is typically used to estimate ˛i , ˇi and Std.ei / in (5.3)—and the R2
is used to assess the quality of the regression.

Remark 5.3 (R2 of market model) R2 of (5.3) measures the fraction of the variance (of
Ri ) that is due to the systematic part of the regression, that is, relative importance of mar-
ket risk as compared to idiosyncratic noise (1 R2 is the fraction due to the idiosyncratic
noise)
2 Var.˛i C ˇi Rm / ˇi2 m2
R D D 2 2 :
Var.Ri / ˇi m C ei 2

To assess the accuracy of historical betas, Blume (1971) and others estimate betas for
non-overlapping samples (periods)—and then compare the betas across samples. They
find that the correlation of betas across samples is moderate for individual assets, but rel-
atively high for diversified portfolios. It is also found that betas tend to “regress” towards
one: an extreme (high or low) historical beta is likely to be followed by a beta that is
closer to one. There are several suggestions for how to deal with this problem.

95
 To use Blume’s ad-hoc technique, let ˇOi1 be the estimate of ˇi from an early sample,
and ˇOi 2 the estimate from a later sample. Then regress

ˇOi 2 D 0 C 1 ˇOi1 C i (5.9)

and use it for forecasting the beta for yet another sample. Blume found . O0 ; O1 / D
.0:343; 0:677/ in his sample.
Other authors have suggested averaging the OLS estimate (ˇOi1 ) with some average
beta. For instance, .ˇOi1 C1/=2 (since the average beta must be unity) or .ˇOi1 C˙inD1 ˇOi1 =n/=2
(which will typically be similar since ˙inD1 ˇOi1 =n is likely to be close to one).
The Bayesian approach is another (more formal) way of adjusting the OLS estimate.
It also uses a weighted average of the OLS estimate, ˇOi1 , and some other number, ˇ0 ,
.1 F /ˇOi1 C Fˇ0 where F depends on the precision of the OLS estimator. The general
idea of a Bayesian approach (Greene (2003) 16) is to treat both Ri and ˇi as random. In
this case a Bayesian analysis could go as follows. First, suppose our prior beliefs (before
having data) about ˇi is that it is normally distributed, N.ˇ0 ; 02 /, where (ˇ0 ; 02 ) are some
numbers . Second, run a LS regression of (5.3). If the residuals are normally distributed,
so is the estimator—it is N.ˇOi1 ; ˇ12
/, where we have taken the point estimate to be the
mean. If we treat the variance of the LS estimator (ˇ1 2
) as known, then the Bayesian
estimator of beta is

b D .1 F /ˇOi1 C Fˇ0 , where

2
1=02 ˇ1
F D D 2 : (5.10)
1=02 C 1=ˇ1
2 2
0 C ˇ1

When the prior beliefs are very precise (02 ! 0), then F ! 1 so the Bayesian
estimator is the same as the prior mean. Effectively, when the prior beliefs are so precise,
there is no room for data to add any information. In contrast, when the prior beliefs are
very imprecise (02 ! 1), then F ! 0, so the Bayesian estimator is the same as OLS.
Effectively, the prior beliefs do not add any information. In the current setting, ˇ0 D 1
and 02 taken from a previous (econometric) study might make sense.

96
5.3.2 Fundamental Betas

Another way to improve the forecasts of the beta over a future period is to bring in infor-
mation about fundamental firm variables. This is particularly useful when there is little
historical data on returns (for instance, because the asset was not traded before).
It is often found that betas are related to fundamental variables as follows (with signs
in parentheses indicating the effect on the beta): Dividend payout (-), Asset growth (+),
Leverage (+), Liquidity (-), Asset size (-), Earning variability (+), Earnings Beta (slope in
earnings regressed on economy wide earnings) (+). Such relations can be used to make
an educated guess about the beta of an asset without historical data on the returns—but
with data on (at least some) of these fundamental variables.

5.4 Multi-Index Models

5.4.1 Overview

The multi-index model is just a multivariate extension of the single-index model (5.3)
PK
Ri D ai C  
kD1 bi k Ik C ei , where (5.11)
E.ei / D 0, Cov ei ; Ik D 0, and Cov.ei ; ej / D 0:


As an example, there could be two indices: the stock market return and an interest rate.
An ad-hoc approach is to first try a single-index model and then test if the residuals are
approximately uncorrelated. If not, then adding a second index might improve the model.
It is often found that it takes several indices to get a reasonable approximation—but
that a single-index model is equally good (or better) at “forecasting” the covariance over
a future period. This is much like the classical trade-off between in-sample fit (requires a
large model) and forecasting (often better with a small model).
The types of indices vary, but one common set captures the “business cycle” and
includes things like the market return, interest rate (or some measure of the yield curve
slope), GDP growth, inflation, and so forth. Another common set of indices are industry
indices.
It turns out (see below) that the calculations of the covariance matrix are much simpler

97
if the indices are transformed to be uncorrelated so we get the model
PK
Ri D ai C kD1 bi k Ik C ei ; where (5.12)
E.ei / D 0, Cov .ei ; Ik / D 0, Cov.ei ; ej / D 0 (unless i D j /, and
Cov.Ik ; Ih / D 0 (unless k D h).

If this transformation of the indices is linear (and non-singular, so it is can be reversed if

we want to), then the fit of the regression is unchanged.

5.4.2 “Rotating” the Indices

There are several ways of transforming the indices to make them uncorrelated, but the fol-
lowing regression approach is perhaps the simplest and may also give the best possibility
of interpreting the results:

1. Let the first transformed index equal the original index, I1 D I1 (possibly de-
meaned). This would often be the market return.

2. Regress the second original index on the first transformed index, I2 D 0 C 1 I1 C
"2 . Then, let the second transformed index be the fitted residual, I2 D "O2 .

3. Regress the third original index on the first two transformed indices, I3 D 0 C
1 I1 C 2 I2 C "3 . Then, let I3 D "O3 . Follow the same idea for all subsequent
indices.

Recall that the fitted residual (from Least Squares) is always uncorrelated with the
regressor (by construction). In this case, this means that I2 is uncorrelated with I1 (step
2) and that I3 is uncorrelated with both I1 and I2 (step 3). The correlation matrix of the
first three rotated indices is therefore
02 31 2 3
I1 1 0 0
Corr @4I2 5A D 40 1 05 : (5.13)
B6 7C 6 7

I3 0 0 1

This recursive approach also helps in interpreting the transformed indices. Suppose
the first index is the market return and that the second original index is an interest rate.
The first transformed index (I1 ) is then clearly the market return. The second transformed

98
index (I2 ) can then be interpreted as the interest rate minus the interest rate expected at the
current stock market return—that is, the part of the interest rate that cannot be explained
by the stock market return.
More generally, let the kth index (k D 1; 2; : : : ; K) be

Ik D "Ok , where "Ok is the fitted residual from the regression (5.14)
Ik D ık1 C ksD11 ks Is C "k : (5.15)
P

Notice that for the first index (k D 1), the regression is only I1 D ı11 C "1 , so I1 equals
the demeaned I1 .

5.4.3 Multi-Index Model after “Rotating” the Indices

To see why the transformed indices are very convenient for calculating the covariance
matrix, consider a two-index model. Then, (5.12) implies that the variance of asset i is

i i D Var .ai C bi1 I1 C bi 2 I2 C ei /

2
D bi1 Var .I1 / C bi22 Var .I2 / C Var .ei / : (5.16)

Similarly, the covariance of assets i and j is

ij D Cov ai C bi1 I1 C bi 2 I2 C ei ; aj C bj1 I1 C bj 2 I2 C ej

D bi1 bj1 Var .I1 / C bi 2 bj 2 Var .I2 / : (5.17)

More generally, with n assets and K indices we can define b1 to be an n  1 vector

of the slope coefficients for the first index (bi1 ; bj1 ) and b2 the vector of slope coefficients
for the second index and so on. Also, let ˙ to be an n  n matrix with the variances of
the residuals along the diagonal. The covariance matrix of the returns is then

Cov.R/ D b1 b10 Var .I1 / C b2 b20 Var .I2 / C : : : C bK bK

0
Var .IK / C ˙ (5.18)
PK
D kD1 bk bk0 Var .Ik / C ˙: (5.19)

See Figure 5.5 for an example.

99
 Correlations, data Difference in correlations: data − model

1 0.5

0.5 0

0 −0.5
25 20 25 20
15 10 20 25 15 10 20 25
5 5 10 15 5 5 10 15
Portfolio Portfolio

25 FF US portfolios, 1957:1−2012:12
Indices (factors): US market, SMB, HML

Figure 5.5: Correlations of US portfolios

5.4.4 Multi-Index Model as a Method for Portfolio Choice

The factor loadings (betas) can be used for more than just constructing the covariance ma-
trix. In fact, the factor loadings are often used directly in portfolio choice. The reason is
simple: the betas summarize how different assets are exposed to the big risk factors/return
drivers. The betas therefore provide a way to understand the broad features of even com-
plicated portfolios. Combined this with the fact that many analysts and investors have
fairly little direct information about individual assets, but are often willing to form opin-
ions about the future relative performance of different asset classes (small vs large firms,
equity vs bonds, etc)—and the role for factor loadings becomes clear.
See Figures 5.6–5.7 for an illustration.

5.5 Principal Component Analysis

Principal component analysis (PCA) can help us determine how many factors that are
needed to explain a cross-section of asset returns.
Let z t D R t RN t be an n  1 vector of demeaned returns with covariance matrix ˙.
The first principal component (pc1t ) is the (normalized) linear combinations of z t that
account for as much of the variability as possible—and its variance is denoted 1 . The

100
 US portfolios, βm, 1957:1−2012:12 US portfolios, βSMBres
1.5
1.4
1
1.2
β

β
0.5

1 0

0 10 20 0 10 20
Portfolio Portfolio

US portfolios, βHMLres

0.5
β

−0.5
0 10 20
Portfolio

Figure 5.6: Loading (betas) of rotated factors

j th (j  2) principal component (pcjt ) is similar (and its variance is denoted j ), except

that is must be uncorrelated with all lower principal components. Remark 5.4 gives a a
formal definition.

Remark 5.4 (Principal component analysis) Consider the zero mean N 1 vector z t with
covariance matrix ˙ . The first (sample) principal component is pc1t D w10 z t , where w1
is the eigenvector associated with the largest eigenvalue (1 ) of ˙ . This value of w1
solves the problem maxw w 0 ˙w subject to the normalization w 0 w D 1. The eigenvalue
1 equals Var.pc1t / D w10 ˙w1 . The j th principal component solves the same problem,
but under the additional restriction that wi0 wj D 0 for all i < j . The solution is the
eigenvector associated with the j th largest eigenvalue j (which equals Var.pcjt / D
wj0 ˙ wj ).

101
 Factor exposure of small growth stocks Factor exposure of large value stocks
HML (res)

HML (res)
Market
Market

SMB (res)

SMB (res)
The factor exposure is measured as abs(β)
The factors are rotated to become uncorrelated

Figure 5.7: Absolute loading (betas) of rotated factors

Let the i th eigenvector be the i th column of the n  n matrix

W D Πw1


wn : (5.20)

We can then calculate the n  1 vector of principal components as

pc t D W 0 z t : (5.21)

Since the eigenvectors are orthogonal it can be shown that W 0 D W 1

, so the expression
can be inverted as
z t D Wpc t : (5.22)

This shows that the i th eigenvector (the i th column of W ) can be interpreted as the effect
of the ith principal component on each of the elements in z t . However, the sign of column
j of W can be changed without any effects (except that the pcjt also changes sign), so
we can always reinterpret a negative coefficient as a positive exposure (to pcjt ).

102
 25 FF US portfolios, eigenvectors, monthly data 1957:1−2012:12
0.5
1st (83.7%)
2nd (6.2%)
3rd (3.3%)

−0.5
0 5 10 15 20 25

Figure 5.8: Eigenvectors for US portfolio returns

Example 5.5 (PCA with 2 series) With two series we have

" #0 " # " #0 " #
w11 z1t w12 z1t
pc1t D and pc2t D or
w21 z2t w22 z2t
" # " #0 " #
pc1t w11 w12 z1t
D and
pc2t w21 w22 z2t
" # " #" #
z1t w11 w12 pc1t
D :
z2t w21 w22 pc2t

For instance, w12 shows how pc2t affects z1t , while w22 shows how pc2t affects z2t .

Remark 5.6 (Data in matrices ) Transpose (5.21) to get pc t0 D z t0 W , where the dimen-
sions are 1  n, 1  n and n  n respectively. If we form a T  n matrix of data Z by
putting z t in row t, then the T  N matrix of principal components can be calculated as
P C D ZW .

Notice that (5.22) shows that all n data series in z t can be written in terms of the n prin-
cipal components. Since the principal components are uncorrelated (Cov.pci t ; pcjt / D
0/), we can think of the sum of their variances (˙inD1 i ) as the “total variation” of the
series in z t . In practice, it is common to report the relative importance of principal com-

103
ponent j as
relative importance of pcj D j =˙inD1 i : (5.23)

For instance, if it is found that the first two principal components account for 75% for the
total variation among many asset returns, then a two-factor model is likely to be a good
approximation.

5.6 Estimating Expected Returns

The starting point for forming estimates of future mean excess returns is typically histor-
ical excess returns. Excess returns are preferred to returns, since this avoids blurring the
risk compensation (expected excess return) with long-run movements in inflation (and
therefore interest rates). The expected excess return for the future period is typically
formed as a judgmental adjustment of the historical excess return. Evidence suggest that
the adjustments are hard to make.
It is typically hard to predict movements (around the mean) of asset returns, but a few
variables seem to have some predictive power, for instance, the slope of the yield curve,
the earnings/price yield, and the book value–market value ratio. Still, the predictive power
is typically low.
Makridakis, Wheelwright, and Hyndman (1998) 10.1 show that there is little evidence
that the average stock analyst beats (on average) the market (a passive index portfolio).
In fact, less than half of the analysts beat the market. However, there are analysts which
seem to outperform the market for some time, but the autocorrelation in over-performance
is weak. The evidence from mutual funds is similar. For them it is typically also found
that their portfolio weights do not anticipate price movements.
It should be remembered that many analysts also are sales persons: either of a stock
(for instance, since the bank is underwriting an offering) or of trading services. It could
well be that their objective function is quite different from minimizing the squared forecast
errors—or whatever we typically use in order to evaluate their performance. (The number
of litigations in the US after the technology boom/bust should serve as a strong reminder
of this.)

104
Bibliography
Amemiya, T., 1985, Advanced econometrics, Harvard University Press, Cambridge, Mas-
sachusetts.

Blume, M. E., 1971, “On the Assessment of Risk,” Journal of Finance, 26, 1–10.

Elton, E. J., M. J. Gruber, S. J. Brown, and W. N. Goetzmann, 2010, Modern portfolio

theory and investment analysis, John Wiley and Sons, 8th edn.

Greene, W. H., 2003, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 5th edn.

Makridakis, S., S. C. Wheelwright, and R. J. Hyndman, 1998, Forecasting: methods and

applications, Wiley, New York, 3rd edn.

105
6 Testing CAPM and Multifactor Models
Reference: Elton, Gruber, Brown, and Goetzmann (2010) 15
More advanced material is denoted by a star ( ). It is not required reading.

6.1 Market Model

Let Riet D Ri t Rf t be the excess return on asset i in excess over the riskfree asset, and
let Rmt
e
be the excess return on the market portfolio. The basic implication of CAPM is
that the expected excess return of an asset (E Riet ) is linearly related to the expected excess
return on the market portfolio (E Rmt
e
) according to

Cov .Ri ; Rm /
E Riet D ˇi E Rmt
e
, where ˇi D : (6.1)
Var .Rm /
Consider the regression

Riet D ˛i C bi Rmt
e
C "i t , where (6.2)
E "i t D 0 and Cov.Rmt
e
; "i t / D 0:

The two last conditions are automatically imposed by LS. Take expectations of the regres-
sion to get
E Riet D ˛i C bi E Rmt
e
: (6.3)

Notice that the LS estimate of bi is the sample analogue to ˇi in (6.1). It is then clear that
CAPM implies that the intercept (˛i ) of the regression should be zero, which is also what
empirical tests of CAPM focus on.
This test of CAPM can be given two interpretations. If we assume that Rmt is the
correct benchmark (the tangency portfolio for which (6.1) is true by definition), then it
is a test of whether asset Ri t is correctly priced. This is typically the perspective in
performance analysis of mutual funds. Alternatively, if we assume that Ri t is correctly
priced, then it is a test of the mean-variance efficiency of Rmt . This is the perspective of
CAPM tests.

106
 The t-test of the null hypothesis that ˛i D 0 uses the fact that, under fairly mild
conditions, the t-statistic has an asymptotically normal distribution, that is

˛O i d
! N.0; 1/ under H0 W ˛i D 0: (6.4)
Std.˛O i /
Note that this is the distribution under the null hypothesis that the true value of the inter-
cept is zero, that is, that CAPM is correct (in this respect, at least).
The test assets are typically portfolios of firms with similar characteristics, for in-
stance, small size or having their main operations in the retail industry. There are two
main reasons for testing the model on such portfolios: individual stocks are extremely
volatile and firms can change substantially over time (so the beta changes). Moreover,
it is of interest to see how the deviations from CAPM are related to firm characteristics
(size, industry, etc), since that can possibly suggest how the model needs to be changed.
The results from such tests vary with the test assets used. For US portfolios, CAPM
seems to work reasonably well for some types of portfolios (for instance, portfolios based
on firm size or industry), but much worse for other types of portfolios (for instance, port-
folios based on firm dividend yield or book value/market value ratio). Figure 6.1 shows
some results for US industry portfolios.

6.1.1 Interpretation of the CAPM Test

Instead of a t-test, we can use the equivalent chi-square test

˛O i2 d
! 21 under H0 : ˛i D 0: (6.5)
Var.˛O i /
Tables (A.2)–(A.1) list critical values for t- and chi-square tests
It is quite straightforward to use the properties of minimum-variance frontiers (see
Gibbons, Ross, and Shanken (1989), and also MacKinlay (1995)) to show that the test
statistic in (6.5) can be written

˛O i2 .SRc /2 .SRm /2
D ; (6.6)
Var.˛O i / Œ1 C .SRm /2 =T
where SRm is the Sharpe ratio of the market portfolio (as before) and SRc is the Sharpe
ratio of the tangency portfolio when investment in both the market return and asset i is
possible. (Recall that the tangency portfolio is the portfolio with the highest possible

107
 US industry portfolios, 1970:1-2012:12 US industry portfolios, 1970:1-2012:12
15 15

Mean excess return

Mean excess return

10 10
D
A D
A
GC G
I FH JB E I FH CJ
5 5 BE

Excess market return: 5.6%

0 0
0 0.5 1 1.5 0 5 10 15
β (against the market) Predicted mean excess return (with α = 0)

alpha pval StdErr CAPM

all NaN 0.04 NaN Factor: US market
A (NoDur) 3.62 0.01 8.70 alpha and StdErr are in annualized %
B (Durbl) -1.21 0.55 13.66
C (Manuf ) 0.70 0.48 6.37
D (Enrgy) 4.06 0.07 14.75
E (HiTec) -1.82 0.32 11.93
F (Telcm) 1.82 0.29 11.10
G (Shops) 1.37 0.35 9.51
H (Hlth ) 2.13 0.21 11.39
I (Utils) 2.87 0.11 11.64
J (Other) -0.65 0.55 6.99

Figure 6.1: CAPM regressions on US industry indices

Sharpe ratio.) If the market portfolio has the same (squared) Sharpe ratio as the tangency
portfolio of the mean-variance frontier of Ri t and Rmt (so the market portfolio is mean-
variance efficient also when we take Ri t into account) then the test statistic, ˛O i2 = Var.˛O i /,
is zero—and CAPM is not rejected.
Proof. ( Proof of (6.6)) From the CAPM regression (6.2) we have
" # " # " # " #
Riet ˇi2 m2 C Var."i t / ˇi m2 ei ˛i C ˇi em
Cov e
D , and D :
Rmt ˇi m2 m2 em em

Suppose we use this information to construct a mean-variance frontier for both Ri t and
Rmt , and we find the tangency portfolio, with excess return Rct
e
. It is straightforward to
show that the square of the Sharpe ratio of the tangency portfolio is e0 ˙ 1 e , where
e is the vector of expected excess returns and ˙ is the covariance matrix. By using the
covariance matrix and mean vector above, we get that the squared Sharpe ratio for the

108
tangency portfolio, e0 ˙ 1
e , (using both Ri t and Rmt ) is
2  e 2
ec ˛i2

m
D C ;
c Var."i t / m
which we can write as
˛i2
.SRc /2 D C .SRm /2 :
Var."i t /
Combine this with (6.8) which shows that Var.˛O i / D Œ1 C .SRm /2  Var."i t /=T .
This is illustrated in Figure 6.2 which shows the effect of adding an asset to the invest-
ment opportunity set. In this case, the new asset has a zero beta (since it is uncorrelated
with all original assets), but the same type of result holds for any new asset. The basic
point is that the market model tests if the new assets moves the location of the tangency
portfolio. In general, we would expect that adding an asset to the investment opportunity
set would expand the mean-variance frontier (and it does) and that the tangency portfolio
changes accordingly. However, the tangency portfolio is not changed by adding an asset
with a zero intercept. The intuition is that such an asset has neutral performance com-
pared to the market portfolio (obeys the beta representation), so investors should stick to
the market portfolio.

6.1.2 Econometric Properties of the CAPM Test

A common finding from Monte Carlo simulations is that these tests tend to reject a true
null hypothesis too often when the critical values from the asymptotic distribution are
used: the actual small sample size of the test is thus larger than the asymptotic (or “nom-
inal”) size (see Campbell, Lo, and MacKinlay (1997) Table 5.1). The practical conse-
quence is that we should either used adjusted critical values (from Monte Carlo or boot-
strap simulations)—or more pragmatically, that we should only believe in strong rejec-
tions of the null hypothesis.
To study the power of the test (the frequency of rejections of a false null hypothesis)
we have to specify an alternative data generating process (for instance, how much extra
return in excess of that motivated by CAPM) and the size of the test (the critical value to
use). Once that is done, it is typically found that these tests require a substantial deviation
from CAPM and/or a long sample to get good power. The basic reason for this is that asset
returns are very volatile. For instance, suppose that the standard OLS assumptions (iid

109
 MV frontiers before and after (α = 0) MV frontiers before and after (α = 0.05)
Solid curves: 2 assets,
0.1 0.1
Dashed curves: 3 assets
Mean

Mean
0.05 0.05

0 0
0 0.05 0.1 0 0.05 0.1
Std Std
The new asset has the abnormal return α
compared to the market (of 2 assets)

MV frontiers before and after (α = −0.04) Means 0.0800 0.0500 α + β(Rm − Rf )

Cov 0.0256 0.0000 0.0000
0.1 matrix 0.0000 0.0144 0.0000
0.0000 0.0000 0.0144
Mean

0.05 Tang N =2 α=0 α = 0.05 α = −0.04

portf 0.47 0.47 0.31 0.82
0.53 0.53 0.34 0.91
0 NaN 0.00 0.34 -0.73
0 0.05 0.1
Std

Figure 6.2: Effect on MV frontier of adding assets

residuals that are independent of the market return) are correct. Then, it is straightforward
to show that the variance of Jensen’s alpha is
" #
.em /2
Var.˛O i / D 1 C
Var."i t /=T (6.7)
Var Rm e

D Œ1 C .SRm /2  Var."i t /=T; (6.8)

where SRm is the Sharpe ratio of the market portfolio. We see that the uncertainty about
the alpha is high when the residual is volatile and when the sample is short, but also when
the Sharpe ratio of the market is high. Note that a large market Sharpe ratio means that
the market asks for a high compensation for taking on risk. A bit uncertainty about how
risky asset i is then translates in a large uncertainty about what the risk-adjusted return
should be.

110
Example 6.1 Suppose we have monthly data with b ˛ i D 0:2% (that is, 0:2%  12 D 2:4%
p
per year), Std ."i t / D 3% (that is, 3%  12  10% per year) and a market Sharpe ratio
p
of 0:15 (that is, 0:15  12  0:5 per year). (This corresponds well to US CAPM
regressions for industry portfolios.) A significance level of 10% requires a t-statistic (6.4)
of at least 1.65, so
0:2
p p  1:65 or T  626:
1 C 0:152 3= T
We need a sample of at least 626 months (52 years)! With a sample of only 26 years (312
months), the alpha needs to be almost 0.3% per month (3.6% per year) or the standard
deviation of the residual just 2% (7% per year). Notice that cumulating a 0.3% return
over 25 years means almost 2.5 times the initial value.

Proof. ( Proof of (6.8)) Consider the regression equation y t D x t0 b C " t . With iid
errors that are independent of all regressors (also across observations), the LS estimator,
bOLs , is asymptotically distributed as
p d
T .bOLs b/ ! N.0;  2 ˙xx1 /, where  2 D Var." t / and ˙xx D plim˙ tD1
T
x t x t0 =T:

When the regressors are just a constant (equal to one) and one variable regressor, f t , so
x t D Œ1; f t 0 , then we have
" # " #
PT 1 T 1 f t 1 E f t
˙xx D E t D1 x t x t0 =T D E D , so
P
T t D1 f t f t2 E f t E f t2
" # " #
 2
E f t
2
E f t  2
Var.f t / C .E f t /2
E f t
 2 ˙xx1 D D :
E f t2 .E f t /2 E ft 1 Var.f t / E ft 1

(In the last line we use Var.f t / D E f t2 .E f t /2 :)

6.1.3 Several Assets

In most cases there are several (n) test assets, and we actually want to test if all the ˛i (for
i D 1; 2; :::; n) are zero. Ideally we then want to take into account the correlation of the
different alphas.
While it is straightforward to construct such a test, it is also a bit messy. As a quick
way out, the following will work fairly well. First, test each asset individually. Second,
form a few different portfolios of the test assets (equally weighted, value weighted) and

111
test these portfolios. Although this does not deliver one single test statistic, it provides
plenty of information to base a judgement on. For a more formal approach, see Section
6.1.4.
A quite different approach to study a cross-section of assets is to first perform a CAPM
regression (6.2) and then the following cross-sectional regression
T
Riet =T D C ˇOi C ui ;
X
(6.9)
t D1

where TtD1 Riet =T is the (sample) average excess return on asset i . Notice that the es-
P

timated betas are used as regressors and that there are as many data points as there are
assets (n).
There are severe econometric problems with this regression equation since the regres-
sor contains measurement errors (it is only an uncertain estimate), which typically tend
to bias the slope coefficient towards zero. To get the intuition for this bias, consider an
extremely noisy measurement of the regressor: it would be virtually uncorrelated with the
dependent variable (noise isn’t correlated with anything), so the estimated slope coeffi-
cient would be close to zero.
If we could overcome this bias (and we can by being careful), then the testable im-
plications of CAPM is that D 0 and that  equals the average market excess return.
We also want (6.9) to have a high R2 —since it should be unity in a very large sample (if
CAPM holds).

6.1.4 Several Assets: SURE Approach

This section outlines how we can set up a formal test of CAPM when there are several
test assets.
For simplicity, suppose we have two test assets. Stack (6.2) for the two equations are

e e
R1t D ˛1 C b1 Rmt C "1t ; (6.10)
e e
R2t D ˛2 C b2 Rmt C "2t (6.11)

where E "i t D 0 and Cov.Rmt e

; "i t / D 0. This is a system of seemingly unrelated regres-
sions (SURE)—with the same regressor (see, for instance, Wooldridge (2002) 7.7). In
this case, the efficient estimator (GLS) is LS on each equation separately. Moreover, the

112
covariance matrix of the coefficients is particularly simple.
To see what the covariances of the coefficients are, write the regression equation for
asset 1 (6.10) on a traditional form
" # " #
e 1 ˛1
R1t D x t0 ˇ1 C "1t , where x t D e
; ˇ1 D ; (6.12)
Rmt b1

and similarly for the second asset (and any further assets).
Define
XT XT
˙O xx D x t x t0 =T , and O ij D "Oi t "Ojt =T; (6.13)
t D1 tD1

where "Oi t is the fitted residual of asset i . The key result is then that the (estimated)
asymptotic covariance matrix of the vectors ˇOi and ˇOj (for assets i and j ) is

Cov.ˇOi ; ˇOj / D O ij ˙O xx1 =T: (6.14)

(In many text books, this is written O ij .X 0 X/ 1 .)

The null hypothesis in our two-asset case is

H0 W ˛1 D 0 and ˛2 D 0: (6.15)

In a large sample, the estimator is normally distributed (this follows from the fact that
the LS estimator is a form of sample average, so we can apply a central limit theorem).
Therefore, under the null hypothesis we have the following result. From (6.8) we know
that the upper left element of ˙xx1 =T equals Œ1 C .SRm /2 =T . Then
" # " # " # !
˛O 1 0 11 12
N ; Œ1 C .SRm /2 =T (asymptotically). (6.16)
˛O 2 0 12 22

In practice we use the sample moments for the covariance matrix. Notice that the zero
means in (6.16) come from the null hypothesis: the distribution is (as usual) constructed
by pretending that the null hypothesis is true. In practice we use the sample moments for
the covariance matrix. Notice that the zero means in (6.16) come from the null hypothesis:
the distribution is (as usual) constructed by pretending that the null hypothesis is true.
We can now construct a chi-square test by using the following fact.

Remark 6.2 If the n  1 vector v  N.0; ˝/, then v 0 ˝ 1

v  2n .

113
 To apply this, form the test static
" #0 " # 1 " #
˛O 1 11 12 ˛O 1
T Œ1 C .SRm /2  1
 22 : (6.17)
˛O 2 12 22 ˛O 2

This can also be transformed into an F test, which might have better small sample prop-
erties.

6.1.5 Representative Results of the CAPM Test

One of the more interesting studies is Fama and French (1993) (see also Fama and French
(1996)). They construct 25 stock portfolios according to two characteristics of the firm:
the size (by market capitalization) and the book-value-to-market-value ratio (BE/ME). In
June each year, they sort the stocks according to size and BE/ME. They then form a 5  5
matrix of portfolios, where portfolio ij belongs to the i th size quintile and the j th BE/ME
quintile:
small size, low B/M : : : : : : : : : small size, high B/M
2 3
6 :: ::: 7
6 : 7
::
6 7
: : :
6 7
6 : 7
: :::
6 7
6
4 :
:
7
5
large size, low B/M large size, high B/M
Tables 6.1–6.2 summarize some basic properties of these portfolios.

Book value/Market value

1 2 3 4 5
Size 1 3:3 9:2 9:6 11:7 13:2
2 5:4 8:4 10:5 10:8 12:0
3 5:7 8:9 8:8 10:3 12:0
4 6:8 6:7 8:6 9:7 9:6
5 5:2 5:8 6:1 5:9 7:3

Table 6.1: Mean excess returns (annualised %), US data 1957:1–2012:12. Size 1: smallest
20% of the stocks, Size 5: largest 20% of the stocks. B/M 1: the 20% of the stocks with
the smallest ratio of book to market value (growth stocks). B/M 5: the 20% of the stocks
with the highest ratio of book to market value (value stocks).

114
 Book value/Market value
1 2 3 4 5
Size 1 1:4 1:2 1:1 1:0 1:1
2 1:4 1:2 1:0 1:0 1:1
3 1:3 1:1 1:0 1:0 1:0
4 1:2 1:1 1:0 1:0 1:0
5 1:0 0:9 0:9 0:8 0:9

Table 6.2: Beta against the market portfolio, US data 1957:1–2012:12. Size 1: smallest
20% of the stocks, Size 5: largest 20% of the stocks. B/M 1: the 20% of the stocks with
the smallest ratio of book to market value (growth stocks). B/M 5: the 20% of the stocks
with the highest ratio of book to market value (value stocks).

Histogram of small growth stocks Histogram of large value stocks

0.1 0.1
mean, std: mean, std:
0.28 7.99 0.61 5.02

0.05 0.05

0 0
−20 −10 0 10 20 −20 −10 0 10 20
Monthly excess return, % Monthly excess return, %

Monthly data on two U.S. indices, 1957:1-2012:12

Sample size: 672
Solid line: estimated normal distribution

Figure 6.3: Comparison of small growth stock and large value stocks

They run a traditional CAPM regression on each of the 25 portfolios (monthly data
1963–1991)—and then study if the expected excess returns are related to the betas as they
should according to CAPM (recall that CAPM implies E Riet D ˇi  where  is the risk
premium (excess return) on the market portfolio).
However, it is found that there is almost no relation between E Riet and ˇi (there is
a cloud in the ˇi  E Riet space, see Cochrane (2001) 20.2, Figure 20.9). This is due
to the combination of two features of the data. First, within a BE/ME quintile, there is
a positive relation (across size quantiles) between E Riet and ˇi —as predicted by CAPM

115
 Fit of CAPM
18

16

14
Mean excess return, %

12

10

6 US data 1957:1-2012:12
25 FF portfolios (B/M and size)
4 p-value for test of model: 0.00

4 6 8 10 12 14 16 18
Predicted mean excess return (CAPM), %

Figure 6.4: CAPM, FF portfolios

(see Cochrane (2001) 20.2, Figure 20.10). Second, within a size quintile there is a negative
relation (across BE/ME quantiles) between E Riet and ˇi —in stark contrast to CAPM (see
Cochrane (2001) 20.2, Figure 20.11).
Figure 6.1 shows some results for US industry portfolios and Figures 6.4–6.6 for US
size/book-to-market portfolios.

6.1.6 Representative Results on Mutual Fund Performance

Mutual fund evaluations (estimated ˛i ) typically find (i) on average neutral performance
(or less: trading costs&fees); (ii) large funds might be worse; (iii) perhaps better perfor-
mance on less liquid (less efficient?) markets; and (iv) there is very little persistence in
performance: ˛i for one sample does not predict ˛i for subsequent samples (except for
bad funds).

116
 Fit of CAPM
18

16

14
Mean excess return, %

12

lines connect same size

10

6 1 (small)
2
3
4
4 5 (large)

4 6 8 10 12 14 16 18
Predicted mean excess return (CAPM), %

Figure 6.5: CAPM, FF portfolios

6.2 Calendar Time and Cross Sectional Regression

To investigate how the performance (alpha) or exposure (betas) of different investors/funds

are related to investor/fund characteristics, we often use the calendar time (CalTime) ap-
proach. First define M discrete investor groups (for instance, age 18–30, 31–40, etc) and
calculate their respective average excess returns (RNjt
e
for group j )

1 P
RNjt
e
D Re ; (6.18)
Nj i 2Groupj i t

where Nj is the number of individuals in group j .

Then, we run a factor model

RNjt
e
D x t0 ˇj C vjt ; for j D 1; 2; : : : ; M (6.19)

where x t typically includes a constant and various return factors (for instance, excess re-
turns on equity and bonds). By estimating these M equations as a SURE system with

117
 Fit of CAPM
18

16

14
Mean excess return, %

12

lines connect same B/M

10

6 1 (low)
2
3
4
4 5 (high)

4 6 8 10 12 14 16 18
Predicted mean excess return (CAPM), %

Figure 6.6: CAPM, FF portfolios

White’s (or Newey-West’s) covariance estimator, it is straightforward to test various hy-

potheses, for instance, that the intercept (the “alpha”) is higher for the M th group than
for the for first group.

Example 6.3 (CalTime with two investor groups) With two investor groups, estimate the
following SURE system

RN 1t
e
D x t0 ˇ1 C v1t ;
RN 2t
e
D x t0 ˇ2 C v2t :

The CalTime approach is straightforward and the cross-sectional correlations are fairly
easy to handle (in the SURE approach). However, it forces us to define discrete investor
groups—which makes it hard to handle several different types of investor characteristics
(for instance, age, trading activity and income) at the same time.
The cross sectional regression (CrossReg) approach is to first estimate the factor

118
model for each investor

Riet D x t0 ˇi C "i t ; for i D 1; 2; : : : ; N (6.20)

and to then regress the (estimated) betas for the pth factor (for instance, the intercept) on
the investor characteristics
ˇOpi D zi0 cp C wpi : (6.21)

In this second-stage regression, the investor characteristics zi could be a dummy variable

(for age roup, say) or a continuous variable (age, say). Notice that using a continuos
investor characteristics assumes that the relation between the characteristics and the beta
is linear—something that is not assumed in the CalTime approach. (This saves degrees of
freedom, but may sometimes be a very strong assumption.) However, a potential problem
with the CrossReg approach is that it is often important to account for the cross-sectional
correlation of the residuals.

6.3 Several Factors

In multifactor models, (6.2) is still valid—provided we reinterpret bi and Rmt

e
as vectors,
so bi Rmt stands for bi o Rot C bip Rpt C :::
e e e

Riet D ˛ C bi o Rot
e e
C bip Rpt C ::: C "i t : (6.22)

In this case, (6.2) is a multiple regression, but the test (6.4) still has the same form (the
standard deviation of the intercept will be different, though).
Fama and French (1993) also try a multi-factor model. They find that a three-factor
model fits the 25 stock portfolios fairly well (two more factors are needed to also fit the
seven bond portfolios that they use). The three factors are: the market return, the return
on a portfolio of small stocks minus the return on a portfolio of big stocks (SMB), and
the return on a portfolio with high BE/ME minus the return on portfolio with low BE/ME
(HML). This three-factor model is rejected at traditional significance levels, but it can still
capture a fair amount of the variation of expected returns.

Remark 6.4 (Returns on long-short portfolios ) Suppose you invest x USD into asset i,
but finance that by short-selling asset j . (You sell enough of asset j to raise x USD.)
The net investment is then zero, so there is no point in trying to calculate an overall

119
 alpha pval StdErr
US industry portfolios, 1970:1-2012:12 all NaN 0.00 NaN
15 A (NoDur) 2.70 0.04 8.47
Mean excess return

B (Durbl) -4.96 0.01 12.36

C (Manuf ) -0.46 0.62 6.08
10 D (Enrgy) 2.90 0.19 14.30
AD E (HiTec) 1.46 0.34 9.96
H GC
F F (Telcm) 1.43 0.41 10.86
5 E I J B
G (Shops) 0.76 0.60 9.40
H (Hlth ) 4.35 0.01 10.66
0 I (Utils) 0.33 0.85 10.45
0 5 10 15 J (Other) -3.00 0.00 5.86
Predicted mean excess return (with α = 0)

Fama-French model
Factors: US market, SMB (size), and HML (book-to-market)
alpha and StdErr are in annualized %

Figure 6.7: Fama-French regressions on US industry indices

return like “value today/investment yesterday - 1.” Instead, the convention is to calculate
an excess return of your portfolio as Ri Rj (or equivalently, Rie Rje ). This excess
return essentially says: if your exposure (how much you invested) is x, then you have
earned x.Ri Rj /. To make this excess return comparable with other returns, you add
the riskfree rate: Ri Rj C Rf , implicitly assuming that your portfolio consists includes
a riskfree investment of the same size as your long-short exposure (x).

Chen, Roll, and Ross (1986) use a number of macro variables as factors—along with
traditional market indices. They find that industrial production and inflation surprises are
priced factors, while the market index might not be.
Figure 6.7 shows some results for the Fama-French model on US industry portfolios
and Figures 6.8–6.10 on the 25 Fama-French portfolios.

6.4 Fama-MacBeth

Reference: Cochrane (2001) 12.3; Campbell, Lo, and MacKinlay (1997) 5.8; Fama and
MacBeth (1973)
The Fama and MacBeth (1973) approach is a bit different from the regression ap-
proaches discussed so far. The method has three steps, described below.

120
 Fit of FF model
18

16

14
Mean excess return, %

12

10

6 US data 1957:1-2012:12
25 FF portfolios (B/M and size)
4 p-value for test of model: 0.00

4 6 8 10 12 14 16 18
Predicted mean excess return (FF), %

Figure 6.8: FF, FF portfolios

 First, estimate the betas ˇi (i D 1; : : : ; n) from (6.2) (this is a time-series regres-

sion). This is often done on the whole sample—assuming the betas are constant.
Sometimes, the betas are estimated separately for different sub samples (so we
could let ˇOi carry a time subscript in the equations below).

 Second, run a cross sectional regression for every t. That is, for period t, estimate
 t from the cross section (across the assets i D 1; : : : ; n) regression

Riet D 0t ˇOi C "i t ; (6.23)

where ˇOi are the regressors. (Note the difference to the traditional cross-sectional
approach discussed in (6.9), where the second stage regression regressed E Riet on
ˇOi , while the Fama-French approach runs one regression for every time period.)

121
 Fit of FF model
18

16

14
Mean excess return, %

12

lines connect same size

10

6 1 (small)
2
3
4
4 5 (large)

4 6 8 10 12 14 16 18
Predicted mean excess return (FF), %

Figure 6.9: FF, FF portfolios

 Third, estimate the time averages

T
1X
"Oi D "Oi t for i D 1; : : : ; n, (for every asset) (6.24)
T t D1
T
O D 1 O t :
X
(6.25)
T t D1

The second step, using ˇOi as regressors, creates an errors-in-variables problem since
ˇOi are estimated, that is, measured with an error. The effect of this is typically to bias the
estimator of  t towards zero (and any intercept, or mean of the residual, is biased upward).
One way to minimize this problem, used by Fama and MacBeth (1973), is to let the assets
be portfolios of assets, for which we can expect some of the individual noise in the first-
step regressions to average out—and thereby make the measurement error in ˇOi smaller.
If CAPM is true, then the return of an asset is a linear function of the market return and an
error which should be uncorrelated with the errors of other assets—otherwise some factor

122
 Fit of FF model
18

16

14
Mean excess return, %

12

lines connect same B/M

10

6 1 (low)
2
3
4
4 5 (high)

4 6 8 10 12 14 16 18
Predicted mean excess return (FF), %

Figure 6.10: FF, FF portfolios

is missing. If the portfolio consists of 20 assets with equal error variance in a CAPM
regression, then we should expect the portfolio to have an error variance which is 1/20th
as large.
We clearly want portfolios which have different betas, or else the second step regres-
sion (6.23) does not work. Fama and MacBeth (1973) choose to construct portfolios
according to some initial estimate of asset specific betas. Another way to deal with the
errors-in-variables problem is to adjust the tests.
We can test the model by studying if "i D 0 (recall from (6.24) that "i is the time
average of the residual for asset i , "it ), by forming a t-test "Oi = Std.O"i /. Fama and MacBeth
(1973) suggest that the standard deviation should be found by studying the time-variation
in "Oi t . In particular, they suggest that the variance of "Oi t (not "Oi ) can be estimated by the
(average) squared variation around its mean
T
1X
Var.O"it / D .O"i t "Oi /2 : (6.26)
T t D1

123
Since "Oi is the sample average of "Oi t , the variance of the former is the variance of the latter
divided by T (the sample size)—provided "Oi t is iid. That is,
T
1 1 X
Var.O"i / D Var.O"i t / D 2 .O"i t "Oi /2 : (6.27)
T T t D1

A similar argument leads to the variance of O

T
O 1 X O O 2:
Var./ D 2 . t / (6.28)
T t D1

Fama and MacBeth (1973) found, among other things, that the squared beta is not
significant in the second step regression, nor is a measure of non-systematic risk.

A Statistical Tables

n Critical values
10% 5% 1%
10 1:81 2:23 3:17
20 1:72 2:09 2:85
30 1:70 2:04 2:75
40 1:68 2:02 2:70
50 1:68 2:01 2:68
60 1:67 2:00 2:66
70 1:67 1:99 2:65
80 1:66 1:99 2:64
90 1:66 1:99 2:63
100 1:66 1:98 2:63
Normal 1:64 1:96 2:58

Table A.1: Critical values (two-sided test) of t distribution (different degrees of freedom)
and normal distribution.

124
 n Critical values
10% 5% 1%
1 2:71 3:84 6:63
2 4:61 5:99 9:21
3 6:25 7:81 11:34
4 7:78 9:49 13:28
5 9:24 11:07 15:09
6 10:64 12:59 16:81
7 12:02 14:07 18:48
8 13:36 15:51 20:09
9 14:68 16:92 21:67
10 15:99 18:31 23:21

Table A.2: Critical values of chisquare distribution (different degrees of freedom, n).

Bibliography
Campbell, J. Y., A. W. Lo, and A. C. MacKinlay, 1997, The econometrics of financial
markets, Princeton University Press, Princeton, New Jersey.

Chen, N.-F., R. Roll, and S. A. Ross, 1986, “Economic forces and the stock market,”
Journal of Business, 59, 383–403.

Cochrane, J. H., 2001, Asset pricing, Princeton University Press, Princeton, New Jersey.

Elton, E. J., M. J. Gruber, S. J. Brown, and W. N. Goetzmann, 2010, Modern portfolio

theory and investment analysis, John Wiley and Sons, 8th edn.

Fama, E., and J. MacBeth, 1973, “Risk, return, and equilibrium: empirical tests,” Journal
of Political Economy, 71, 607–636.

Fama, E. F., and K. R. French, 1993, “Common risk factors in the returns on stocks and
bonds,” Journal of Financial Economics, 33, 3–56.

Fama, E. F., and K. R. French, 1996, “Multifactor explanations of asset pricing anoma-
lies,” Journal of Finance, 51, 55–84.

Gibbons, M., S. Ross, and J. Shanken, 1989, “A test of the efficiency of a given portfolio,”
Econometrica, 57, 1121–1152.

125
MacKinlay, C., 1995, “Multifactor models do not explain deviations from the CAPM,”
Journal of Financial Economics, 38, 3–28.

Wooldridge, J. M., 2002, Econometric analysis of cross section and panel data, MIT
Press.

126
7 Time Series Analysis
Reference: Newbold (1995) 17 or Pindyck and Rubinfeld (1998) 13.5, 16.1–2, and 17.2
More advanced material is denoted by a star ( ). It is not required reading.

7.1 Descriptive Statistics

The sth autocovariance of y t is estimated by

b
Cov .y t ; y t s / D
PT
t D1 .y t N .y t
y/ s N =T , where yN D
y/
PT
tD1 y t =T: (7.1)

The conventions in time series analysis are that we use the same estimated (using all data)
mean in both places and that we divide by T .
The sth autocorrelation is estimated as

Os D
b
Cov .y t ; y t s /
: (7.2)
c .y t /2
Std
Compared with a traditional estimate of a correlation we here impose that the standard
deviation of y t and y t p are the same (which typically does not make much of a differ-
ence).
The sampling properties of Os are complicated, but there are several useful large sam-
ple results for Gaussian processes (these results typically carry over to processes which
are similar to the Gaussian—a homoskedastic process with finite 6th moment is typically
enough, see Priestley (1981) 5.3 or Brockwell and Davis (1991) 7.2-7.3). When the true
autocorrelations are all zero (not 0 , of course), then for any i and j different from zero

p
" # " # " #!
Oi 0 1 0
T !d N ; : (7.3)
Oj 0 0 1

This result can be used to construct tests for both single autocorrelations (t-test or 2 test)
and several autocorrelations at once (2 test). In particular,
p d
T Os ! N.0; 1/; (7.4)

127
 p
so T Os can be used as a t-stat.

Example 7.1 (t-test) We want to test the hypothesis that 1 D 0. Since the N.0; 1/ dis-
tribution has 5% of the probability mass below -1.65 and another 5% above 1.65, we
p
can reject the null hypothesis at the 10% level if T jO1 j > 1:65. With T D 100, we
p
therefore need jO1 j > 1:65= 100 D 0:165 for rejection, and with T D 1000 we need
p
jO1 j > 1:65= 1000  0:052.
p
The Box-Pierce test follows directly from the result in (7.3), since it shows that T Oi
p
and T Oj are iid N(0,1) variables. Therefore, the sum of the square of them is distributed
as a 2 variable. The test statistics typically used is
L
X
QL D T Os2 !d L
2
: (7.5)
sD1

Example 7.2 (Box-Pierce) Let O1 D 0:165, and T D 100, so Q1 D 100  0:1652 D
2:72. The 10% critical value of the 21 distribution is 2.71, so the null hypothesis of no
autocorrelation is rejected.

The choice of lag order in (7.5), L, should be guided by theoretical considerations, but
it may also be wise to try different values. There is clearly a trade off: too few lags may
miss a significant high-order autocorrelation, but too many lags can destroy the power of
the test (as the test statistics is not affected much by increasing L, but the critical values
increase).
The pth partial autocorrelation is discussed in Section 7.4.6.

7.2 Stationarity

The process y t is (weakly) stationary if the mean, variance, and covariances are finite and
constant across time

E yt D  < 1 (7.6)
Var.y t / D 0 < 1 (7.7)
Cov.y t ; y t s / D s < 1 (7.8)

128
 The autocorrelation function is just the autocorrelation coefficient s as a function of
s. Notice that
lim s D 0 for any stationary series. (7.9)
jsj!1

7.3 White Noise

The white noise process is the basic building block used in most other time series models.
It is characterized by a zero mean, a constant variance, and no autocorrelation

E "t D 0
Var ." t / D  2 , and
Cov ." t s ; " t / D 0 if s ¤ 0. (7.10)

If, in addition, " t is normally distributed, then it is said to be Gaussian white noise. This
process can clearly not be forecasted.
To construct a variable that has a non-zero mean, we can form

yt D  C "t ; (7.11)

where  is a constant. This process is most easily estimated by estimating the sample
mean and variance in the usual way (as in (7.1) with p D 0) or my OLS with a constant
as the only regressor.

7.4 Autoregression (AR)

7.4.1 AR(1)

In this section we study the first-order autoregressive process, AR(1), in some detail in
order to understand the basic concepts of autoregressive processes. The process is as-
sumed to have a zero mean (or is demeaned, that an original variable minus its mean, for
instance y t D x t xN t )—but it is straightforward to put in any mean or trend.
An AR(1) is
y t D ay t 1 C " t ; with Var." t / D  2 ; (7.12)

where " t is the white noise process in (7.10) which is uncorrelated with y t 1. If 1<

129
a < 1, then the effect of a shock eventually dies out: y t is stationary.
The AR(1) model can be estimated with OLS (since " t and y t 1 are uncorrelated) and
the usual tools for testing significance of coefficients and estimating the variance of the
residual all apply.
The basic properties of an AR(1) process are (provided jaj < 1)

Var .y t / D  2 =.1 a2 / (7.13)

Corr .y t ; y t s / D as ; (7.14)

so the variance and autocorrelation are increasing in a (assuming a > 0).

See Figure 7.1 for an illustration.

Remark 7.3 (Autocorrelation and autoregression). Notice that the OLS estimate of a
in (7.12) is essentially the same as the sample autocorrelation coefficient in (7.2). This
b
follows from the fact that the slope coefficient is Cov .y t ; y t 1 / =Var.y
c t 1 /. The denomi-
nator can be a bit different since a few data points are left out in the OLS estimation, but
the difference is likely to be small.

Example 7.4 With a D 0:85 and  2 D 0:52 , we have Var .y t / D 0:25=.1 0:852 /  0:9,
which is much larger than the variance of the residual. (Why?)

If a D 1 in (7.12), then we get a random walk. It is clear from the previous analysis
that a random walk is non-stationary—that is, the effect of a shock never dies out. This
implies that the variance is infinite and that the standard tools for testing coefficients etc.
are invalid. The solution is to study changes in y instead: y t y t 1 . In general, processes
with the property that the effect of a shock never dies out are called non-stationary or unit
root or integrated processes. Try to avoid them.

130
 Autocorrelations Partial autocorrelations
1 1
AR(1), 0.85
0.5 0.5 MA(1), 0.85

0 0

−0.5 AR(1), 0.85 −0.5

MA(1), 0.85
−1 −1
0 5 10 0 5 10
Periods Periods

Autocorrelations Partial autocorrelations

1 1
AR(1), -0.85
0.5 0.5 MA(1), -0.85

0 0

−0.5 AR(1), -0.85 −0.5

MA(1), -0.85
−1 −1
0 5 10 0 5 10
Periods Periods

Figure 7.1: Autocorrelations and partial autocorrelations

7.4.2 More on the Properties of an AR(1) Process

Solve (7.12) backwards by repeated substitution

y t D a.ay t 2 C " t 1 / C " t (7.15)

„ ƒ‚ …
yt 1

D a2 y t 2 C a" t 1 C "t (7.16)

::
: (7.17)
K
X
KC1
Da yt K 1 C as " t s : (7.18)
sD0

The factor aKC1 y t K 1 declines monotonically to zero if 0 < a < 1 as K increases, and
declines in an oscillating fashion if 1 < a < 0. In either case, the AR(1) process is

131
 Slope coefficient (b) R2
Slope with 90% conf band
0.5
0.1
0
0.05
−0.5
0
0 20 40 60 0 20 40 60
Return horizon (months) Return horizon (months)

Monthly US stock returns 1957:1-2012:12

Regression: rt = a + brt−1 + ǫt

Figure 7.2: Predicting US stock returns (various investment horizons) with lagged returns.

covariance stationary and we can then take the limit as K ! 1 to get

y t D " t C a" t 1 C a2 " t 2 C :::

1
X
D as " t s : (7.19)
sD0

Since " t is uncorrelated over time, y t 1 and " t are uncorrelated. We can therefore
calculate the variance of y t in (7.12) as the sum of the variances of the two components
on the right hand side

Var .y t / D Var .ay t 1/ C Var ." t /

D a2 Var .y t 1/ C Var ." t /
D Var ." t / =.1 a2 /; since Var .y t 1/ D Var .y t / . (7.20)

In this calculation, we use the fact that Var .y t 1 / and Var .y t / are equal. Formally, this
follows from that they are both linear functions of current and past "s terms (see (7.19)),
which have the same variance over time (" t is assumed to be white noise).
Note from (7.20) that the variance of y t is increasing in the absolute value of a, which
is illustrated in Figure 7.3. The intuition is that a large jaj implies that a shock have effect
over many time periods and thereby create movements (volatility) in y.

132
 Forecast with 90% conf band Forecast with 90% conf band
5 5

0 0

Intial value: 3 Intial value: 0

−5 −5
0 5 10 0 5 10
Forecasting horizon Forecasting horizon

AR(1) model:
yt+1 = 0.85yt + ǫt+1 , with σ = 0.5

Figure 7.3: Properties of AR(1) process

Similarly, the covariance of y t and y t 1 is

Cov .y t ; y t 1/ D Cov .ay t 1 C "t ; yt 1/

D a Cov .y t 1; yt 1/

D a Var .y t / : (7.21)

We can then calculate the first-order autocorrelation as

Cov .y t ; y t 1 /
Corr .y t ; y t 1/ D
Std.y t / Std.y t 1 /
D a: (7.22)

It is straightforward to show that

Corr .y t ; y t s / D Corr .y t Cs ; y t / D as : (7.23)

7.4.3 Forecasting with an AR(1)

Suppose we have estimated an AR(1). To simplify the exposition, we assume that we

actually know a and Var." t /, which might be a reasonable approximation if they were
estimated on long sample.

133
 We want to forecast y t C1 using information available in t . From (7.12) we get

y t C1 D ay t C " t C1 : (7.24)

Since the best guess of " t C1 is that it is zero, the best forecast and the associated forecast
error are

E t y t C1 D ay t , and (7.25)
y t C1 E t y t C1 D " t C1 with variance  2 . (7.26)

We may also want to forecast y t C2 using the information in t. To do that note that
(7.12) gives

y t C2 D ay t C1 C " t C2
D a.ay t C " t C1 / C " t C2
„ ƒ‚ …
y t C1

D a2 y t C a" t C1 C " t C2 : (7.27)

Since the E t " t C1 and E t " t C2 are both zero, we get that

E t y tC2 D a2 y t ; and (7.28)

y t C2 E t y tC2 D a" tC1 C " t C2 with variance a2  2 C  2 : (7.29)

More generally, we have

E t y t Cs D as y t , (7.30)
 
Var .y t Cs E t y t Cs / D 1 C a2 C a4 C ::: C a2.s 1/
2 (7.31)
a2s 1 2
D 2  : (7.32)
a 1
Example 7.5 If y t D 3; a D 0:85 and  D 0:5, then the forecasts and the forecast error
variances become

Horizon s E t y t Cs Var .y tCs E t y t Cs /

1 0:85  3 D 2:55 0:25
0:852  3 D 2:17 0:852 C 1  0:52 D 0:43


2
0:8550 1
25 0:8525  3 D 0:05 0:852 1
 0:52 D 0:90

134
Notice that the point forecast converge towards zero and the variance of the forecast error
variance to the unconditional variance (see Example 7.4).

If the shocks " t , are normally distributed, then we can calculate 90% confidence in-
tervals around the point forecasts in (7.25) and (7.28) as

90% confidence band of E t y tC1 W ay t ˙ 1:65   (7.33)

p
90% confidence band of E t y t C2 W a2 y t ˙ 1:65  a2  2 C  2 : (7.34)

(Recall that 90% of the probability mass is within the interval 1:65 to 1:65 in the N(0,1)
distribution). To get 95% confidence bands, replace 1.65 by 1.96. Figure 7.3 gives an
example.

Example 7.6 Continuing Example 7.5, we get the following 90% confidence bands

Horizon s
p
1 2:55 ˙ 1:65  0:25  Œ1:7; 3:4
p :
2 2:17 ˙ 1:65  0:43  Œ1:1; 3:2
p
25 0:05 ˙ 1:65  0:90  Œ 1:5; 1:6

Remark 7.7 (White noise as special case of AR(1).) When a D 0 in (7.12), then the
AR(1) collapses to a white noise process. The forecast is then a constant (zero) for all
forecasting horizons, see (7.30), and the forecast error variance is also the same for all
horizons, see (7.32).

7.4.4 Adding a Constant to the AR(1)

The discussion of the AR(1) worked with a zero mean variable, but that was just for
convenience (to make the equations shorter). One way to work with a variable x t with
a non-zero mean, is to first estimate its sample mean xN t and then let the y t in the AR(1)
model (7.12) be a demeaned variable y t D x t xN t .
To include a constant  in the theoretical expressions, we just need to substitute x t 
for y t everywhere. For instance, in (7.12) we would get

xt  D a .x t 1 / C " t or
x t D .1 a/  C ax t 1 C "t : (7.35)

135
Estimation by LS will therefore give an intercept that equals .1 a/  and a slope coef-
ficient that equals a.

7.4.5 AR(p)

The pth-order autoregressive process, AR(p), is a straightforward extension of the AR(1)

y t D a1 y t 1 C a2 y t 2 C :::ap y t p C "t : (7.36)

All the previous calculations can be made on this process as well—it is just a bit messier.
This process can also be estimated with OLS since " t is uncorrelated with lags of y t .
Adding a constant is straightforward by substituting x t  for y t everywhere

7.4.6 Partial Autocorrelations

The pth partial autocorrelation tries to measure the direct relation between y t and y t p ,
where the indirect effects of y t 1 ; :::; y t pC1 are eliminated. For instance, if y t is gen-
erated by an AR(1) model, then the 2nd autocorrelation is a2 , whereas the 2nd partial
autocorrelation is zero. The partial autocorrelation is therefore a way to gauge how many
lags that are needed in an AR(p) model.
In practice, the first partial autocorrelation is estimated by a in an AR(1)

y t D ay t 1 C "t : (7.37)

The second partial autocorrelation is estimated by the second slope coefficient (a2 ) in an
AR(2)
y t D a1 y t 1 C a2 y t 2 C " t ; (7.38)

and so forth. The general pattern is that the pth partial autocorrelation is estimated by the
slope coefficient of the pth lag in an AR(p), where we let p go from 1,2,3...
See Figure 7.1 for an illustration.

7.4.7 Forecasting with an AR(2)

As an example, consider making a forecast of y t C1 based on the information in t by using

an AR(2)
y t C1 D a1 y t C a2 y t 1 C " t C1 : (7.39)

136
 Forecast made in t − 1 and Actual Forecast made in t − 2 and Actual
R2=0.88 R2=0.66
4 4

2 2
%

%
0 Forecast 0
Actual

−2 −2
1995 2000 2005 2010 1995 2000 2005 2010

Comparison of forecast errors from

Model: AR(2) of US 4-quarter GDP growth
autoregression (AR) and random walk (RW):
Estimated on data for 1947-1994
1-quarter 2-quarter
R2 is corr(forecast,actual)2 for 1995– MSE(AR)/MSE(RW) 0.70 0.72
yt and Et−s yt are plotted in t MAE(AR)/MAE(RW) 0.95 0.98
R2 (AR)/R2 (RW) 1.07 1.27

Figure 7.4: Forecasting with an AR(2)

This immediately gives the one-period point forecast

E t y t C1 D a1 y t C a2 y t 1: (7.40)

We can use (7.39) to write y t C2 as

y t C2 D a1 y t C1 C a2 y t C " t C2
D a1 .a1 y t C a2 y t 1 C " t C1 / C a2 y t C " t C2
„ ƒ‚ …
y tC1

D .a12 C a2 /y t C a1 a2 y t 1 C a1 " t C1 C " tC2 : (7.41)

Figure 7.4 gives an empirical example.

The expressions for the forecasts and forecast error variances quickly get somewhat
messy—and even more so with an AR of higher order than two. There is a simple, and
approximately correct, shortcut that can be taken. Note that both the one-period and two-
period forecasts are linear functions of y t and y t 1 . We could therefore estimate the

137
following two equations with OLS

y t C1 D a1 y t C a2 y t 1 C " t C1 (7.42)
y t C2 D b1 y t C b2 y t 1 C v t C2 : (7.43)

Clearly, (7.42) is the same as (7.39) and the estimated coefficients can therefore be used
to make one-period forecasts, and the variance of " t C1 is a good estimator of the variance
of the one-period forecast error. The coefficients in (7.43) will be very similar to what we
get by combining the a1 and a2 coefficients as in (7.41): b1 will be similar to a12 C a2 and
b2 to a1 a2 (in an infinite sample they should be identical). Equation (7.43) can therefore
be used to make two-period forecasts, and the variance of v t C2 can be taken to be the
forecast error variance for this forecast.

7.5 Moving Average (MA)

A qt h -order moving average process is

y t D " t C 1 " t 1 C ::: C q " t q; (7.44)

where the innovation " t is white noise (usually Gaussian). It is straightforward to add a
constant to capture a non-zero mean.
Estimation of MA processes is typically done by setting up the likelihood function
and then using some numerical method to maximize it; LS does not work at all since the
right hand side variables are unobservable. This is one reason why MA models play a
limited role in applied work. Moreover, most MA models can be well approximated by
an AR model of low order.
The autocorrelations and partial autocorrelations (for different lags) can help us gauge
if the time series looks more like an AR or an MA. In an AR(p) model, the autocorrela-
tions decay to zero for long lags, while the p C1 partial autocorrelation (and beyond) goes
abruptly to zero. The reverse is true for an MA model. See Figure 7.1 for an illustration.

138
7.6 ARMA(p,q)

When the autocorrelations and partial autocorrelations look a bit like both an AR and an
MA model, then a combination (ARMA) might be appropriate.
Autoregressive-moving average models add a moving average structure to an AR
model. For instance, an ARMA(2,1) could be

y t D a1 y t 1 C a2 y t 2 C " t C 1 " t 1; (7.45)

where " t is white noise. This type of model is much harder to estimate than the autore-
gressive model (use MLE). The appropriate specification of the model (number of lags
of y t and " t ) is often unknown. The Box-Jenkins methodology is a set of guidelines for
arriving at the correct specification by starting with some model, study the autocorrelation
structure of the fitted residuals and then changing the model.
It is straightforward to add a constant to capture a non-zero mean.
Most ARMA models can be well approximated by an AR model—provided we add
some extra lags. Since AR models are so simple to estimate, this approximation approach
is often used.

Remark 7.8 In an ARMA model, both the autocorrelations and partial autocorrelations
decay to zero for long lags.

To choose a model, study the ACF and PACF—and check that residual are close to
white noise (or at least not autocorrelated). To avoid overfitting, “punish” models with
to many parameters. Akaike’s Information Criterion (AIC) and the Bayesian information
criterion (BIC) are
pCqC1
AIC D ln O 2 C 2 (7.46)
T
p C q C1
BIC D ln O 2 C ln T: (7.47)
T
trade-off between fit (low O 2 ) and number of parameters (p C q). Choose the model with
the lowest AIC or BIC. (AIC often exaggerates the length)

139
 Autocorr and partial autocorr t-stat, autocorrelation of residual from AR(p)
1
autocorr 6
pac
0.5 4
2
0
0

−0.5 −2
2 4 6 8 10 2 4 6 8 10
lag p, no. of lags in AR

AIC and BIC, different AR(p) models Data:

0.8 US 4-quarter GDP growth, 1947Q1-2011Q4

0.6

0.4

0.2 AIC
BIC
0
2 4 6 8 10
p, no. of lags in AR

Figure 7.5: Example of choosing lag length in an AR model

7.7 VAR(p)

The vector autoregression is a multivariate version of an AR(1) process: we can think of

y t and " t in (7.36) as vectors and the ai as matrices.
For instance the VAR(1) of two variables (x t and z t ) is (in matrix form)
" # " #" # " #
x t C1 a11 a12 xt "xtC1
D C ; (7.48)
z t C1 a21 a22 zt "ztC1

or equivalently

x t C1 D a11 x t C a12 z t C "xtC1 ; and (7.49)

z t C1 D a21 x t C a22 z t C "ztC1 : (7.50)

Both (7.49) and (7.50) are regression equations, which can be estimated with OLS

140
(since "xt C1 and "ztC1 are uncorrelated with x t and z t ).
With the information available in t , that is, information about x t and z t , (7.49) and
(7.50) can be used to forecast one step ahead as

E t x t C1 D a11 x t C a12 z t (7.51)

E t z t C1 D a21 x t C a22 z t : (7.52)

We also want to make a forecast of x t C2 based on the information in t . Clearly, it

must be the case that

E t x t C2 D a11 E t x t C1 C a12 E t z t C1 (7.53)

E t z t C2 D a21 E t x t C1 C a22 E t z t C1 : (7.54)

We already have values for E t x t C1 and E t z t C1 from (7.51) and (7.52) which we can use.
For instance, for E t x t C2 we get

E t x tC2 D a11 .a11 x t C a12 z t / C a12 .a21 x t C a22 z t /

„ ƒ‚ … „ ƒ‚ …
E t x t C1 E t z tC1
2
D a11 C a12 a21 x t C .a12 a22 C a11 a12 / z t : (7.55)

This has the same form as the one-period forecast in (7.51), but with other coefficients.
Note that all we need to make the forecasts (for both t C 1 and t C 2) are the values in
period t (x t and z t ). This follows from that (7.48) is a first-order system where the values
of x t and z t summarize all relevant information about the future that is available in t .
The forecast uncertainty about the one-period forecast is simple: the forecast error
x tC1 E t x t C1 D "xtC1 . The two-period forecast error, x t C2 E t x t C2 , is a linear
combination of "xtC1 , "zt C1 , and "xt C2 . The calculations of the forecasting error variance
(as well as for the forecasts themselves) quickly get messy. This is even more true when
the VAR system is of a higher order.
As for the AR(p) model, a practical way to get around the problem with messy calcu-
lations is to estimate a separate model for each forecasting horizon. In a large sample, the
difference between the two ways is trivial. For instance, suppose the correct model is the
VAR(1) in (7.48) and that we want to forecast x one and two periods ahead. From (7.51)

141
and (7.55) we see that the regression equations should be of the form

x t C1 D ı1 x t C ı2 z t C u t C1 , and (7.56)
x t C2 D 1 x t C 2 z t C w t Cs : (7.57)

With estimated coefficients (OLS can be used), it is straightforward to calculate forecasts

and forecast error variances.
In a more general VAR(p) model we need to include p lags of both x and z in the
regression equations (p D 1 in (7.56) and (7.57)).

7.7.1 Granger Causality

If z t can help predict future x, over and above what lags of x itself can, then z is said to
Granger Cause x. This is a statistical notion of causality, and may not necessarily have
much to do with economic causality (Christmas cards may Granger cause Christmas).
In (7.56) z does Granger cause x if ı2 ¤ 0, which can be tested with an F-test. More
generally, there may be more lags of both x and z in the equation, so we need to test if all
coefficients on different lags of z are zero.

7.8 Impulse Response Function

Any stationary process can be rewritten on (“inverted to”) MA form

Example: AR(1)!MA(1)

y t D y t 1 C "t
D .y t 2 C " t 1 / C " t D  2 y t 2 C " t 1 C "t
„ ƒ‚ …
yt 1

::
:
D " t C " t 1 C  2"t 2 C ::: (7.58)

The MA form can be interpreted as giving the impulse response (the dynamic response

142
to a shock in period t). Set all "s D 0, except " t D 1. From (7.44) we have

yt D 1
y t C1 D 1
y t C2 D 2 ; etc (7.59)

See Figure 7.6 for an illustration.

143
 Impulse response (shock in t = 1)
1.5
AR(1), (0.85)
MA(1), (0.5)
ARMA(1,1), (0.85,0.5)

0.5

0
0 2 4 6 8 10
Periods

Figure 7.6: Impulse responses

7.9 Non-stationary Processes

7.9.1 Introduction

A trend-stationary process can be made stationary by subtracting a linear trend. The

simplest example is
y t D  C ˇt C " t (7.60)

where " t is white noise.

A unit root process can be made stationary only by taking a difference. The simplest
example is the random walk with drift

yt D  C yt 1 C "t ; (7.61)

where " t is white noise. The name “unit root process” comes from the fact that the largest
eigenvalues of the canonical form (the VAR(1) form of the AR(p)) is one. Such a process
is said to be integrated of order one (often denoted I(1)) and can be made stationary by
taking first differences. (So the first difference is an I(0) series.)

Example 7.9 (Non-stationary AR(2)) The process y t D 1:5y t 1 0:5y t 2 C " t can be

144
written " # " #" # " #
yt 1:5 0:5 yt 1 "t
D C ;
yt 1 1 0 yt 2 0
where the matrix has the eigenvalues 1 and 0.5 and is therefore non-stationary. Note that
subtracting y t 1 from both sides gives y t y t 1 D 0:5 .y t 1 y t 2 /C" t , so the variable
x t D y t y t 1 is stationary.

The distinguishing feature of unit root processes is that the effect of a shock never
vanishes. This is most easily seen for the random walk. Substitute repeatedly in (7.61) to
get

y t D  C . C y t 2 C "t 1/ C "t
::
:
t
X
D t C y0 C "s : (7.62)
sD1

The effect of " t never dies out: a non-zero value of " t gives a permanent shift of the level
of y t . This process is clearly non-stationary. See Figure 7.7 for an illustration.
A consequence of the permanent effect of a shock is that the variance of the con-
ditional distribution grows without bound as the forecasting horizon is extended. For
instance, for the random walk with drift, (7.62), the distribution conditional on the in-
formation in t D 0 is N.y0 C t; s 2 / if the innovations are normally distributed. This
means that the expected change is t and that the conditional variance grows linearly with
the forecasting horizon. The unconditional variance is therefore infinite and the standard
results on inference are not applicable.
In contrast, the conditional distribution from the trend stationary model, (7.60), is
N.st;  2 /.
A process could have two unit roots (integrated of order 2: I(2)). In this case, we need
to difference twice to make it stationary. Alternatively, a process can also be explosive,
that is, have eigenvalues outside the unit circle. In this case, the impulse response function
diverges.

Example 7.10 (Two unit roots.) Suppose y t in Example (7.9) is actually the first differ-

145
 Impulse response (shock in t = 1)
1.5
AR(1), (1.0)
MA(1), (0.5)
ARMA(1,1), (1.0,0.5)

0.5

0
0 2 4 6 8 10
Periods

Figure 7.7: Impulse responses

ence of some other series, y t D z t zt 1. We then have

zt zt 1 D 1:5 .z t 1 zt 2/ 0:5 .z t 2 zt 3/ C "t

z t D 2:5z t 1 2z t 2 C 0:5z t 3 C "t ;

which is an AR(3) with the following canonical form

2 3 2 32 3 2
3
zt 2:5 2 0:5 zt 1 "t
4 zt 1 5 D 4 1 0 0 5 4 zt 5 C 4 0 5:
6 7 6 76 7 6 7
2

zt 2 0 1 0 zt 3 0

The eigenvalues are 1, 1, and 0.5, so z t has two unit roots (integrated of order 2: I(2) and
needs to be differenced twice to become stationary).

Example 7.11 (Explosive AR(1).) Consider the process y t D 1:5y t 1 C " t . The eigen-
value is then outside the unit circle, so the process is explosive. This means that the
impulse response to a shock to " t diverges (it is 1:5s for s periods ahead).

Remark 7.12 (Lag operator ) A common and convenient way of dealing with leads and

146
lags is the lag operator, L. It is such that

Ls y t D y t s

For instance, the AR(1) model

y t D  y t 1 C " t , or
„ƒ‚…
Ly t

.1 L/ y t D " t , or
.L/y t D " t ;

where .L/ D .1 L/ is a lag polynomial. Similarly, an ARMA(2,1) can be written

yt 1 y t 1 2 y t 2 D " t C ˛1 " t 1
2 L2 y t D .1 C ˛1 L/ " t :


1 1 L

7.9.2 Spurious Regressions

Strong trends often causes problems in econometric models where y t is regressed on x t .

In essence, if no trend is included in the regression, then x t will appear to be significant,
just because it is a proxy for a trend. The same holds for unit root processes, even if
they have no deterministic trends. However, the innovations accumulate and the series
therefore tend to be trending in small samples. A warning sign of a spurious regression is
when R2 > DW statistics.
See Figure 7.8 for an empirical example and Figures 7.9–7.11 for a Monte Carlo
simulation.
For trend-stationary data, this problem is easily solved by detrending with a linear
trend (before estimating or just adding a trend to the regression).
However, this is usually a poor method for a unit root processes. What is needed is a
first difference. For instance, a first difference of the random walk with drift is

y t D y t yt 1

D  C "t ; (7.63)

which is white noise (any finite difference, like y t y t s , will give a stationary series),

147
 Real GDP and deflator, US 1947Q1-2011Q4 Residual from deflator = a + b × gdp
2 R2 : 0.96
t-stat for slope: 80.65 0.2
1.5
log deflator

0.1

%
1 0
−0.1
0.5
−0.2 Auto (t) and DW stats: 15.99 0.01
0
0 0.5 1 1.5 2 1960 1980 2000
log GDP

Figure 7.8: Example of a spurious regression

Distribution of bLS , ρ = 0.2 Distribution of bLS , ρ = 1

6 6

4 4

2 2

0 0
−0.5 0 0.5 −0.5 0 0.5
b b

Illustration of spurious regressions

bLS is the LS estimate of b in
y and x are uncorrelated AR(1) processes: yt = a + bxt + ut , T = 200
yt = ρyt−1 + ǫt
xt = ρxt−1 + ηt Number of simulations: 25000
where ǫt and ηt are uncorrelated

Figure 7.9: Distribution of LS estimator when y t and x t are independent AR(1) processes

so we could proceed by applying standard econometric tools to
y t .

One may then be tempted to try first-differencing all non-stationary series, since it
may be hard to tell if they are unit root process or just trend-stationary. For instance, a
first difference of the trend stationary process, (7.60), gives

yt yt 1 D ˇ C "t "t 1: (7.64)

148
 Distribution of R2 , ρ = 0.2 Distribution of R2 , ρ = 1
50
15
40
30 10
20
5
10
0 0
0 0.5 1 0 0.5 1
R2 R2

Distribution of autocorr, ρ = 0.2 Distribution of autocorr, ρ = 1

15 15

10 10

5 5

0 0
0 0.5 1 0 0.5 1
Autocorrelations of residuals Autocorrelation of residuals

Figure 7.10: Distribution of R2 and autorrelation of residuals. See Figure 7.9

Distribution of t-stat, ρ = 0.2 Distribution of t-stat, ρ = 1

0.4 0.1

0.3

0.2 0.05

0.1

0 0
−5 0 5 −10 0 10 20
t-stat (slope = 0) t-stat (slope = 0)

Figure 7.11: Distribution of t-statistics. See Figure 7.9

Its unclear if this is an improvement: the trend is gone, but the errors are now of MA(1)
type (in fact, non-invertible, and therefore tricky, in particular for estimation).

149
7.9.3 Testing for a Unit Root

Suppose we run an OLS regression of

y t D ay t 1 C "t ; (7.65)

where the true value of jaj < 1. The asymptotic distribution of the LS estimator is
p
T .aO a/  N 0; 1 a2 : (7.66)

(The variance follows from the standard OLS formula where the variance of the estimator
is  2 .X 0 X=T / 1 . Here plim X 0 X=T D Var .y t / which we know is  2 = 1 a2 ).

It is well known (but not easy to show) that when a D 1, then aO is biased towards
zero in small samples. In addition, the asymptotic distribution is no longer (7.66). In
fact, there is a discontinuity in the limiting distribution as we move from a stationary to
a non-stationary variable. This, together with the small sample bias means that we have
to use simulated critical values for testing the null hypothesis of a D 1 based on the OLS
estimate from (7.65).
In practice, the approach is to run the regression (7.65) with a constant (and perhaps
even a time trend), calculate the test statistic
aO 1
DF D ; (7.67)
Std.a/
O
and reject the null of non-stationarity if DF is less than the critical values published by
Dickey and Fuller ( 2:86 at the 5% level if the regression has a constant, and 3:41 if
the regression includes a trend).
With more dynamics (to capture any serial correlation in " t in (7.65)), do an aug-
mented DickeyFuller test (ADF)

y t D ı C 1 y t 1 C 2 y t 2 C "2t , or

y t D ı C .1 C 2 1/ y t 1 2
y t 1 C "2t ; (7.68)

and test if 1 C 2 1 D 0 (against the alternative, < 0) The critical values are as for the
DF test. If "2t is autocorrelated, then add further lags.

150
 The KPSS test has stationarity as the null hypothesis. It has three steps. First, regress

yt D a C "t : (7.69)

Second, define
Pt
St D Os
sD1 " for t D 1; :::; T and let (7.70)
O 2 D Var.O" t /: (7.71)

Third, the test statistic is

1 PT
KPSS D S 2 =O 2 (7.72)
T 2 t D1 t
Reject stationarity if KPSS > 0:463 (a 5% critical value). We could also include a linear
trend in (KPSSReg). The 5% critical value is then 0.146.
In principle, distinguishing between a stationary and a non-stationary series is very
difficult (and impossible unless we restrict the class of processes, for instance, to an
AR(2)), since any sample of a non-stationary process can be arbitrary well approximated
by some stationary process et vice versa. The lesson to be learned, from a practical point
of view, is that strong persistence in the data generating process (stationary or not) invali-
dates the usual results on inference. We are usually on safer ground to apply the unit root
results in this case, even if the process is actually stationary.

7.9.4 Cointegration

An exception to the “spurious regression” result: Y t and X t are I(1) but share a common
stochastic trend such that
y t ˛ ˇx t is I(0). (7.73)

In this case, OLS works fine: it is actually very good (super consistent), ˇO converges to
true value ˇ faster than in standard theory. The intuition is that if ˇO ¤ ˇ, then " t are I(1)
and therefore have high sample variance: OLS will pick ˇO close to ˇ.
In (7.73), we call .1; ˇ/ the cointegrating vector, since
" #
h i y
t
1 ˇ is I(0) (7.74)
xt

Example 7.13 Y t is GDP, x t is private consumption. Suppose both are driven by the

151
non-stationary productivity of the economy, A t , plus other stationary stuff (z t ; w t )

y t D A t C z t
x t D ıA t C w t

From the second equation A t D .x t w t /=ı, use in first equation

yt D xt C zt wt
„ƒ‚… ı „ƒ‚… ı …
I.1/ I.1/
„ ƒ‚
I.0/

To test if y t and x t are cointegrated, we need to study three things. First, does it make
sense? Look at data, and consider the (economic) theory. Second, are both x t and y t I(I)?
Do Dickey-Fuller tests, etc. Third, are (a; O in from the regression
O b)

y t D a C bx t C " t (7.75)

such that "O t is I(0)? To determine the latter, do an ADF test on "O t , but use special critical
values—H0 : no cointegration (so " t is I.1/). 5% critical values: 3:34 (if x t is a scalar).
One way to incorporate the cointegration in a model of the short-run dynamics is to
use a Error-Correction Model, for instance,

y t D ı C 1
x t 1 .y t 1 ˇx t 1/ C " t or perhaps (7.76)

D ı C 1
x t 1 C 1
y t 1 .y t 1 ˇx t 1/ C "t

Recall: .y t ; x t / are I(1), but y t 1 ˇx t 1 is I(0), so all terms in (7.76) are I(0). We
typically do not put the intercept into the cointegrating relation (as there is already another
intercept in the equation).
If > 0, then the system is driven ack to a stationary path for y ˇx: the “error
correction mechanism.” Can have more lags of both
y and
x.
Estimation is fairly straightforward (Engle-Granger’s 2-step method). First, estimate
the cointegrating vector. Second, use it in (7.76) and estimate the rest of the parameters.
(Standard inference applies to them.)

152
 t-stat for autocorr of residual
in ADF test of gdp or c
Real GDP and consumption (US)
gdp c
2 log GDP 1 lag in ADF test 5.82 1.28
log consumption 2 lags in ADF test -0.51 -0.36
1.5 3 lags in ADF test 0.21 -0.08
%

1 ADF test of series

gdp c
0.5 1 lag in ADF test -1.79 -0.23
2 lags in ADF test -1.97 -1.32
0 3 lags in ADF test -1.56 -1.37
1960 1980 2000

t-stat for autocorr of residual

in ADF test of gdp − a − b × c
gdp − a − b × c
0.05 1 lag in ADF test -1.19
2 lags in ADF test -0.04
3 lags in ADF test -0.26
0
%

ADF test
−0.05 1 lag in ADF test -4.03
2 lags in ADF test -4.16
3 lags in ADF test -4.29
−0.1
1960 1980 2000

Figure 7.12: Unit root tests, US quarterly macro data

Bibliography
Brockwell, P. J., and R. A. Davis, 1991, Time series: theory and methods, Springer Verlag,
New York, second edn.

Newbold, P., 1995, Statistics for business and economics, Prentice-Hall, 4th edn.

Pindyck, R. S., and D. L. Rubinfeld, 1998, Econometric models and economic forecasts,
Irwin McGraw-Hill, Boston, Massachusetts, 4ed edn.

Priestley, M. B., 1981, Spectral analysis and time series, Academic Press.

153

gdp
Coint res t 1 0:10
. 2:56/

gdp t 1 0:15
.1:61/

c t 1 0:33
.3:00/

gdp t 2 0:02
.0:24/

c t 2 0:22
.2:34/
const 0:00
.2:01/
R2 0:25
obs 257:00

Table 7.1: Error-correction model for log real US GDP growth, 1947Q1-2011Q4. Num-
bers in parentheses are t-stats. The ’Coint res’ is the residual from regressing the log GDP
level on the log consumption level.

154
8 Predicting Asset Returns
Reference (medium): Elton, Gruber, Brown, and Goetzmann (2010) 17 (efficient markets)
and 26 (earnings estimation)
Additional references: Campbell, Lo, and MacKinlay (1997) 2 and 7; Cochrane (2001)
20.1
More advanced material is denoted by a star ( ). It is not required reading.

8.1 Autocorrelations

8.1.1 Autocorrelation Coefficients and the Box-Pierce Test

The autocovariances of the y t process can be estimated as

T
1 X
Os D .y t N .y t
y/ s N ; with
y/ (8.1)
T t D1Cs
T
1X
yN D yt : (8.2)
T t D1

(We typically divide by T in (8.1) even if we have only T s full observations to estimate
s from.) Autocorrelations are then estimated as

Os D Os = O0 : (8.3)

The sampling properties of Os are complicated, but there are several useful large sam-
ple results for Gaussian processes (these results typically carry over to processes which
are similar to the Gaussian—a homoskedastic process with finite 6th moment is typically
enough, see Priestley (1981) 5.3 or Brockwell and Davis (1991) 7.2-7.3). When the true
autocorrelations are all zero (not 0 , of course), then for any i and j different from zero

p
" # " # " #!
Oi 0 1 0
T !d N ; : (8.4)
Oj 0 0 1

155
This result can be used to construct tests for both single autocorrelations (t-test or 2 test)
and several autocorrelations at once (2 test).

Example 8.1 (t-test) We want to test the hypothesis that 1 D 0. Since the N.0; 1/ dis-
tribution has 5% of the probability mass below -1.65 and another 5% above 1.65, we
p
can reject the null hypothesis at the 10% level if T jO1 j > 1:65. With T D 100, we
p
therefore need jO1 j > 1:65= 100 D 0:165 for rejection, and with T D 1000 we need
p
jO1 j > 1:65= 1000  0:052.
p
The Box-Pierce test follows directly from the result in (8.4), since it shows that T Oi
p
and T Oj are iid N(0,1) variables. Therefore, the sum of the square of them is distributed
as a 2 variable. The test statistics typically used is
L
X
QL D T Os2 !d L
2
: (8.5)
sD1

Example 8.2 (Box-Pierce) Let O1 D 0:165, and T D 100, so Q1 D 100  0:1652 D
2:72. The 10% critical value of the 21 distribution is 2.71, so the null hypothesis of no
autocorrelation is rejected.

The choice of lag order in (8.5), L, should be guided by theoretical considerations, but
it may also be wise to try different values. There is clearly a trade off: too few lags may
miss a significant high-order autocorrelation, but too many lags can destroy the power of
the test (as the test statistics is not affected much by increasing L, but the critical values
increase).

8.1.2 Autoregressions

An alternative way of testing autocorrelations is to estimate an AR model

y t D c C a1 y t 1 C a2 y t 2 C ::: C ap y t p C "t ; (8.6)

and then test if all slope coefficients (a1 ; a2 ; :::; ap ) are zero with a 2 or F test. This
approach is somewhat less general than the Box-Pierce test, but most stationary time
series processes can be well approximated by an AR of relatively low order.
See Figure 8.4 for an illustration.

156
 SMI SMI daily excess returns, %
10
SMI
bill portfolio 5
4
0
2
−5

0 −10
2000 2010 2000 2010

Daily SMI data, 1993:5-2013:5

1st order autocorrelation of returns (daily, weekly, monthly): 0.03 -0.11 0.04
1st order autocorrelation of absolute returns (daily, weekly, monthly): 0.28 0.31 0.19

Figure 8.1: Time series properties of SMI

Autocorr, daily excess returns Autocorr, weekly excess returns

Autocorr with 90% conf band around 0
0.2 S&P 500, 1979:1-2013:4 0.2

0.1 0.1

0 0

−0.1 −0.1
1 2 3 4 5 1 2 3 4 5
lag (days) lag (weeks)

Autocorr, daily abs(excess returns) Autocorr, weekly abs(excess returns)

0.2 0.2

0.1 0.1

0 0

−0.1 −0.1
1 2 3 4 5 1 2 3 4 5
lag (days) lag (weeks)

Figure 8.2: Predictability of US stock returns

157
 Autocorr, excess returns, smallest decile Autocorr, excess returns, 5th decile

0.1 0.1

0 0

−0.1 −0.1
1 2 3 4 5 1 2 3 4 5
lag (days) lag (days)

Autocorr, excess returns, largest decile Autocorr with 90% conf band around 0
US daily data 1979:1-2012:12

0.1

−0.1
1 2 3 4 5
lag (days)

Figure 8.3: Predictability of US stock returns, size deciles

The autoregression can also allow for the coefficients to depend on the market situ-
ation. For instance, consider an AR(1), but where the autoregression coefficient may be
different depending on the sign of last period’s return

y t D c C aı.y t 1  0/y t 1 C bı.y t 1 > 0/y t 1 C " t , where (8.7)

(
1 if q is true
ı.q/ D
0 else.

See Figure 8.5 for an illustration.

Inference of the slope coefficient in autoregressions on returns for longer data horizons
than the data frequency (for instance, analysis of weekly returns in a data set consisting
of daily observations) must be done with care. If only non-overlapping returns are used
(use the weekly return for a particular weekday only, say Wednesdays), the standard LS

158
 Slope coefficient (b) R2

0.5 Slope with 90% conf band

0.1
0
0.05
−0.5
0
0 20 40 60 0 20 40 60
Return horizon (months) Return horizon (months)

Scatter plot, 36 month returns Monthly US stock returns 1926:1-2012:12

2
Regression: rt = a + brt−1 + ǫt
1
Return

−1

−2
−2 −1 0 1 2
lagged return

Figure 8.4: Predictability of US stock returns

expression for the standard deviation of the autoregressive parameter is likely to be rea-
sonable. This is not the case, if overlapping returns (all daily data on weekly returns) are
used.

Remark 8.3 (Overlapping returns ) Consider an AR(1) for the two-period return, y t 1C

yt
y t C1 C y t C2 D a C b2 .y t 1 C y t / C " t C2 :

Two successive observations with non-overlapping returns are then

y t C1 C y t C2 D a C b2 .y t 1 C y t / C " t C2
y t C3 C y t C4 D a C b2 .y t C1 C y t C2 / C " t C4 :

Suppose that y t is not autocorrelated, so the slope coefficient b2 D 0. We can then write

159
 Autoregression coeff, after negative returns Autoregression coeff, after positive returns
0.1 0.1
with 90% conf band around 0
0.05 S&P 500 (daily), 1979:1-2013:4 0.05

0 0

−0.05 −0.05

−0.1 −0.1
1 2 3 4 5 1 2 3 4 5
lag (days) lag (days)

Based on the following regression:

rt = α + β(1 − Qt−1 )rt−1 + γQt−1 rt−1 + ǫt
Qt−1 = 1 if rt−1 > 0, and zero otherwise

Figure 8.5: Predictability of US stock returns, results from a regression with interactive
dummies

the residuals as

" t C2 D a C y tC1 C y tC2

" t C4 D a C y tC3 C y tC4 ;

which are uncorrelated. Compare this to the case where we use overlapping data. Two
successive observations are then

y t C1 C y t C2 D a C b2 .y t 1 C y t / C " t C2
y t C2 C y t C3 D a C b2 .y t C y t C1 / C " t C3 :

As before, b2 D 0 if y t has no autocorrelation, so the residuals become

" t C2 D a C y tC1 C y t C2
" t C3 D a C y t C2 C y t C3 ;

which are correlated since y t C2 shows up in both. This demonstrates that overlapping
return data introduces autocorrelation of the residuals—which has to be handled in order
to make correct inference.

160
8.1.3 Autoregressions versus Autocorrelations

It is straightforward to see the relation between autocorrelations and the AR model when
the AR model is the true process. This relation is given by the Yule-Walker equations.
For an AR(1), the autoregression coefficient is simply the first autocorrelation coeffi-
cient. For an AR(2), y t D a1 y t 1 C a2 y t 2 C " t , we have

Cov.y t ; y t / Cov.y t ; a1 y t 1 C a2 y t 2 C " t /

2 3 2 3

4 Cov.y t 1 ; y t / 5 D 4 Cov.y t 1 ; a1 y t 1 C a2 y t 2 C " t / 5

6 7 6 7

Cov.y t 2 ; y t / Cov.y t 2 ; a1 y t 1 C a2 y t 2 C " t /

a1 Cov.y t ; y t 1 / C a2 Cov.y t ; y t 2 / C Cov.y t ; " t /
2 3

D 4 a1 Cov.y t 1 ; y t 1 / C a2 Cov.y t 1 ; y t 2 / 5 , or
6 7

a1 Cov.y t C a2 Cov.y t
2; yt 1/ 2; yt 2/

a1 1 C a2 2 C Var." t /
2 3 2 3
0
4 1 5 D 4 a1 0 C a2 1 5: (8.8)
6 7 6 7

2 a1 1 C a2 0

To transform to autocorrelation, divide by 0 . The last two equations are then

" # " # " # " #
1 a1 C a2  1 1 a1 = .1 a2 /
D or D : (8.9)
2 a1 1 C a2 2 a12 = .1 a2 / C a2

If we know the parameters of the AR(2) model (a1 , a2 , and Var." t /), then we can
solve for the autocorrelations. Alternatively, if we know the autocorrelations, then we can
solve for the autoregression coefficients. This demonstrates that testing if all the autocor-
relations are zero is essentially the same as testing if all the autoregressive coefficients are
zero. Note, however, that the transformation is non-linear, which may make a difference
in small samples.

8.1.4 Variance Ratios

A variance ratio is another way to measure predictability. It is defined as the variance of

a q-period return divided by q times the variance of a 1-period return
P 
q 1
Var y
sD0 t s
VRq D : (8.10)
q Var.y t /

161
 Variance Ratio, 1926- Variance Ratio, 1957-
1.5 VR with 90% conf band 1.5

1 1

0.5 0.5
0 20 40 60 0 20 40 60
Return horizon (months) Return horizon (months)

Monthly US stock returns 1926:1-2012:12

The confidence bands use the asymptotic

sampling distribution of the variance ratios

Figure 8.6: Variance ratios, US excess stock returns

To see that this is related to predictability, consider the 2-period variance ratio.

Var.y t C y t 1 /
VR2 D (8.11)
2 Var.y t /
Var .y t / C Var .y t 1 / C 2 Cov .y t ; y t 1/
D
2 Var .y t /
Cov .y t ; y t 1 /
D1C
Var .y t /
D 1 C 1 : (8.12)

It is clear from (8.12) that if y t is not serially correlated, then the variance ratio is unity;
a value above one indicates positive serial correlation and a value below one indicates
negative serial correlation. The same applies to longer horizons.
The estimation of VRq is typically not done by replacing the population variances in
(8.10) with the sample variances, since this would require using non-overlapping long
returns—which wastes a lot of data points. For instance, if we have 24 years of data and
we want to study the variance ratio for the 5-year horizon, then 4 years of data are wasted.

162
Instead, we typically rely on a transformation of (8.10)
P 
q 1
Var y
sD0 t s
VRq D
q Var.y t /
q 1
X  jsj

D 1 s or
q
sD .q 1/
q 1 
X s
D1C2 1 s : (8.13)
sD1
q

To estimate VRq , we first estimate the autocorrelation coefficients (using all available data
points for each estimation) and then calculate (8.13).

b
Remark 8.4 ( Sampling distribution of V Rq ) Under the null hypothesis that there is no
autocorrelation, (8.4) and (8.13) give

b
" q 1 
p 
 #

d
X s 2
T V Rq 1 ! N 0; 4 1 :
sD1
q

b b
Example 8.5 (Sampling distributions of V R2 and V R3 )
p 
b 
b
T V R2 1 !d N .0; 1/ or V R2 !d N .1; 1=T /
p 
b 
b
and T V R3 1 !d N .0; 20=9/ or V R3 !d N Œ1; .20=9/=T  :

The results in CLM Table 2.5 and 2.6 (weekly CRSP stock index returns, early 1960s
to mid 1990s) show variance ratios above one and increasing with the number of lags, q.
The results for individual stocks in CLM Table 2.7 show variance ratios close to, or even
below, unity. Cochrane Tables 20.5–6 report weak evidence for more mean reversion in
multi-year returns (annual NYSE stock index,1926 to mid 1990s).
See Figure 8.6 for an illustration.

8.2 Other Predictors and Methods

There are many other possible predictors of future stock returns. For instance, both the
dividend-price ratio and nominal interest rates have been used to predict long-run returns,
and lagged short-run returns on other assets have been used to predict short-run returns.

163
8.2.1 Lead-Lags

Stock indices have more positive autocorrelation than (most) individual stocks: there
should therefore be fairly strong cross-autocorrelations across individual stocks. (See
Campbell, Lo, and MacKinlay (1997) Tables 2.7 and 2.8.) Indeed, this is also what is
found in US data where weekly returns of large size stocks forecast weekly returns of
small size stocks.
See Figure 8.7 for an illustration.

Regression of smallest decile on lag of Regression of 5th decile on lag of

self self
largest largest
0.1 0.1

0 0

−0.1 −0.1
1 2 3 4 5 1 2 3 4 5
lag (days) lag (days)

US size deciles
Regression of largest decile on lag of US daily data 1979:1-2012:12

self Multiple regression with lagged return on

0.1 self and largest deciles as regressors.
The figures show regression coefficients.

−0.1
1 2 3 4 5
lag (days)

Figure 8.7: Coefficients from multiple prediction regressions

164
 (Auto-)correlation matrix, monthly FF returns 1957:1-2012:12
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

1 0.18 0.19 0.18 0.15 0.15 0.22 0.20 0.20 0.17 0.15 0.23 0.21 0.18 0.20 0.16 0.22 0.20 0.18 0.18 0.16 0.18 0.18 0.16 0.14 0.15

2 0.17 0.17 0.18 0.16 0.16 0.21 0.19 0.20 0.18 0.16 0.22 0.22 0.19 0.22 0.18 0.21 0.21 0.19 0.19 0.18 0.17 0.18 0.17 0.15 0.17

3 0.17 0.18 0.19 0.17 0.17 0.20 0.20 0.20 0.19 0.17 0.21 0.21 0.20 0.22 0.18 0.21 0.21 0.20 0.20 0.18 0.18 0.18 0.18 0.17 0.17

4 0.18 0.20 0.20 0.19 0.19 0.21 0.21 0.22 0.21 0.19 0.22 0.23 0.22 0.23 0.20 0.23 0.23 0.22 0.22 0.20 0.19 0.20 0.19 0.18 0.18

5 0.22 0.23 0.25 0.23 0.24 0.25 0.25 0.26 0.26 0.25 0.26 0.27 0.26 0.28 0.25 0.26 0.27 0.27 0.27 0.25 0.22 0.24 0.23 0.24 0.25

6 0.12 0.12 0.12 0.10 0.10 0.15 0.14 0.14 0.12 0.10 0.18 0.16 0.14 0.16 0.11 0.18 0.16 0.14 0.14 0.13 0.15 0.14 0.12 0.11 0.12

7 0.13 0.14 0.15 0.13 0.13 0.16 0.15 0.16 0.15 0.13 0.18 0.18 0.16 0.18 0.15 0.18 0.18 0.17 0.16 0.15 0.15 0.15 0.13 0.14 0.15

8 0.12 0.13 0.14 0.12 0.12 0.15 0.15 0.16 0.15 0.13 0.17 0.18 0.17 0.19 0.16 0.18 0.19 0.18 0.18 0.15 0.16 0.16 0.15 0.16 0.16

9 0.12 0.13 0.14 0.13 0.13 0.14 0.14 0.15 0.15 0.14 0.15 0.18 0.17 0.19 0.16 0.17 0.18 0.18 0.18 0.16 0.14 0.15 0.15 0.16 0.17

10 0.12 0.14 0.15 0.14 0.15 0.15 0.16 0.17 0.17 0.16 0.16 0.18 0.19 0.20 0.19 0.17 0.19 0.20 0.19 0.19 0.16 0.17 0.17 0.18 0.20

11 0.07 0.07 0.08 0.06 0.06 0.11 0.09 0.10 0.08 0.06 0.13 0.13 0.10 0.13 0.08 0.13 0.13 0.11 0.11 0.10 0.12 0.11 0.10 0.10 0.10

12 0.10 0.11 0.13 0.10 0.10 0.13 0.13 0.14 0.12 0.11 0.15 0.16 0.14 0.16 0.13 0.16 0.17 0.16 0.15 0.12 0.15 0.13 0.12 0.13 0.13

13 0.11 0.12 0.14 0.12 0.12 0.14 0.13 0.15 0.14 0.12 0.15 0.16 0.15 0.17 0.14 0.17 0.17 0.16 0.16 0.13 0.14 0.12 0.12 0.13 0.14

14 0.08 0.09 0.12 0.10 0.11 0.12 0.12 0.13 0.13 0.12 0.13 0.15 0.14 0.16 0.14 0.15 0.16 0.17 0.15 0.14 0.14 0.13 0.13 0.15 0.15

15 0.10 0.11 0.13 0.11 0.11 0.12 0.13 0.13 0.13 0.12 0.13 0.15 0.15 0.16 0.14 0.15 0.15 0.16 0.15 0.13 0.13 0.12 0.11 0.13 0.15

16 0.08 0.08 0.08 0.05 0.05 0.10 0.09 0.09 0.07 0.05 0.12 0.11 0.08 0.10 0.05 0.12 0.10 0.09 0.08 0.07 0.10 0.09 0.07 0.06 0.07

17 0.09 0.11 0.12 0.10 0.10 0.13 0.13 0.13 0.13 0.10 0.14 0.15 0.13 0.15 0.12 0.15 0.15 0.15 0.13 0.11 0.12 0.11 0.11 0.11 0.11

18 0.09 0.10 0.12 0.10 0.10 0.11 0.12 0.12 0.12 0.10 0.12 0.14 0.12 0.14 0.11 0.14 0.14 0.15 0.13 0.11 0.11 0.10 0.10 0.11 0.11

19 0.07 0.08 0.11 0.09 0.09 0.10 0.10 0.11 0.11 0.10 0.11 0.13 0.11 0.14 0.11 0.13 0.13 0.14 0.12 0.11 0.12 0.10 0.11 0.11 0.13

20 0.08 0.09 0.12 0.10 0.12 0.10 0.12 0.13 0.13 0.12 0.11 0.14 0.14 0.15 0.13 0.13 0.14 0.15 0.14 0.13 0.12 0.11 0.10 0.12 0.14

21 0.05 0.06 0.07 0.05 0.04 0.08 0.08 0.08 0.06 0.04 0.11 0.09 0.06 0.08 0.03 0.11 0.07 0.07 0.05 0.04 0.09 0.06 0.04 0.03 0.03

22 0.05 0.06 0.08 0.06 0.05 0.08 0.08 0.08 0.08 0.06 0.09 0.10 0.08 0.10 0.06 0.10 0.09 0.09 0.06 0.06 0.07 0.07 0.05 0.06 0.06

23 0.03 0.04 0.06 0.04 0.04 0.05 0.06 0.06 0.06 0.05 0.06 0.08 0.07 0.08 0.04 0.08 0.07 0.07 0.05 0.05 0.05 0.05 0.05 0.05 0.05

24 0.04 0.04 0.06 0.05 0.05 0.06 0.06 0.06 0.07 0.06 0.07 0.08 0.07 0.08 0.06 0.08 0.08 0.09 0.06 0.06 0.07 0.05 0.05 0.05 0.07

25 0.07 0.07 0.09 0.07 0.08 0.10 0.10 0.10 0.10 0.09 0.11 0.12 0.11 0.11 0.11 0.13 0.12 0.11 0.10 0.09 0.13 0.09 0.09 0.09 0.09

Figure 8.8: Illustration of the cross-autocorrelations, Corr.R t ; R t k /, monthly FF data.

Dark colors indicate high correlations, light colors indicate low correlations.

8.2.2 Dividend-Price Ratio as a Predictor

One of the most successful attempts to forecast long-run returns is a regression of future
returns on the current dividend-price ratio (here in logs)
q
X
r t Cs D ˛ C ˇq .d t p t / C " t Cq : (8.14)
sD1

For instance, CLM Table 7.1, report R2 values from this regression which are close to
zero for monthly returns, but they increase to 0.4 for 4-year returns (US, value weighted
index, mid 1920s to mid 1990s).
See Figure 8.9 for an illustration.

8.2.3 Predictability but No Autocorrelation

The evidence for US stock returns is that long-run returns may perhaps be predicted by the
dividend-price ratio or interest rates, but that the long-run autocorrelations are weak (long-
run US stock returns appear to be “weak-form efficient” but not “semi-strong efficient”).
This should remind us of the fact that predictability and autocorrelation need not be the
same thing: although autocorrelation implies predictability, we can have predictability

165
 Slope coefficient (b) R2

Slope with 90% conf band

0.4 0.1

0.2 0.05

0 0
0 20 40 60 0 20 40 60
Return horizon (months) Return horizon (months)

Scatter plot, 36 month returns Monthly US stock returns 1926:1-2012:12

2
Regression: rt = a + b log(E/P)t−1 + ǫt
1
Return

−1

−2
−4 −3 −2 −1
lagged log(E/P)

Figure 8.9: Predictability of US stock returns

without autocorrelation.

8.3 Out-of-Sample Forecasting Performance

8.3.1 In-Sample versus Out-of-Sample Forecasting

References: Goyal and Welch (2008), and Campbell and Thompson (2008)
Goyal and Welch (2008) find that the evidence of predictability of equity returns dis-
appears when out-of-sample forecasts are considered. Campbell and Thompson (2008)
claim that there is still some out-of-sample predictability, provided we put restrictions on
the estimated models.
Campbell and Thompson (2008) first report that only few variables (earnings price
ratio, T-bill rate and the inflation rate) have significant predictive power for one-month

166
stock returns in the full sample (1871–2003 or early 1920s–2003, depending on predictor).
To gauge the out-of-sample predictability, they estimate the prediction equation using
data up to and including t 1, and then make a forecast for period t. The forecasting
performance of the equation is then compared with using the historical average as the
predictor. Notice that this historical average is also estimated on data up to an including
t 1, so it changes over time. Effectively, they are comparing the forecast performance
of two models estimated in a recursive way (long and longer sample): one model has just
an intercept, the other has also a predictor. The comparison is done in terms of the RMSE
and an “out-of-sample R2 ”
1 XT 1 XT
2
ROS D1 .r t rOt /2 = .r t rQt /2 ; (8.15)
T t Ds T t Ds

where s is the first period with an out-of-sample forecast, rOt is the forecast based on the
prediction model (estimated on data up to and including t 1) and rQt is the prediction from
some benchmark model (also estimated on data up to and including t 1). In practice,
the paper uses the historical average (also estimated on data up to and including t 1) as
the benchmark prediction. That is, the benchmark prediction is that the return in t will
equal the historical average.
The evidence shows that the out-of-sample forecasting performance is very weak—as
claimed by Goyal and Welch (2008).
It is argued that forecasting equations can easily give strange results when they are
estimated on a small data set (as they are early in the sample). They therefore try different
restrictions: setting the slope coefficient to zero whenever the sign is “wrong,” setting
the prediction (or the historical average) to zero whenever the value is negative. This
improves the results a bit—although the predictive performance is still weak.
See Figure 8.10 for an illustration.

8.3.2 Trading Strategies

Another way to measure predictability and to illustrate its economic importance is to

calculate the return of a dynamic trading strategy, and then measure the “performance”
of this strategy in relation to some benchmark portfolios. The trading strategy should, of
course, be based on the variable that is supposed to forecast returns.
A common way (since Jensen, updated in Huberman and Kandel (1987)) is to study

167
 Out-of-sample R2 , E/P regression Out-of-sample R2 , max(E/P regression,0)

0 0

−0.2 −0.2

−0.4 −0.4

100 200 300 100 200 300

Length of data window, months Length of data window, months

US stock returns (1-year, in excess of riskfree) 1926:1-2012:12

Estimation is done on moving data window,

forecasts are made out of sample for: 1957:1-2012:12

Figure 8.10: Predictability of US stock returns, in-sample and out-of-sample

the performance of a portfolio by running the following regression

R1t Rf t D ˛ C ˇ.Rmt Rf t / C " t , with (8.16)

E " t D 0 and Cov.Rmt Rf t ; " t / D 0;

where R1t Rf t is the excess return on the portfolio being studied and Rmt Rf t the
excess returns of a vector of benchmark portfolios (for instance, only the market portfolio
if we want to rely on CAPM; returns times conditional information if we want to allow
for time-variation in expected benchmark returns). Neutral performance (mean-variance
intersection, that is, that the tangency portfolio is unchanged and the two MV frontiers
intersect there) requires ˛ D 0, which can be tested with a t test.
See Figure 8.11 for an illustration.

8.3.3 More Evidence on Out-of-Sample Forecasting Performance

Figures 8.12–8.16 illustrate the out-of-sample performance on daily returns. Figure 8.12
shows that extreme S&P 500 returns are followed by mean-reverting movements the fol-
lowing day—which suggests that a trading strategy should sell after a high return and buy
after a low return. However, extreme returns are rare, so Figure 8.13 tries a simpler strate-

168
 Buy winners and sell losers

excess return
alpha
8

Monthly US data 1957:1-2012:12, 25 FF portfolios (B/M and size)

2
Buy (sell) the 5 assets with highest (lowest) return
over the last month
0
0 2 4 6 8 10 12
Evaluation horizon, days

Figure 8.11: Predictability of US stock returns, momentum strategy

Average return after ”events”

5

Daily S&P 500 excess returns 1979:1-2013:4

−5
−10 −5 0 5 10
Lagged return (bins)

Figure 8.12: Short-run predictability of US stock returns, out-of-sample

gies: buy after a negative return (or hold T-bills), or instead buy after a positive return (or
hold T-bills). It turns out that the latter has a higher average return, which suggests that the
extreme mean-reverting movements in Figure 8.12 are actually dominated by smaller mo-

169
 Average return after ”events” Average return on strategies
10 10

freq
5 riskfree 5

0.47 0.53
3.64 3.53
0 0
Rt−1 < 0 Rt−1 ≥ 0 average Rt−1 < 0 Rt−1 ≥ 0 always

Strategies (rebalanced daily):

Daily S&P 500 returns 1979:1-2013:4
hold stocks if condition is met;
The returns are annualized
otherwise, hold T-bills

Figure 8.13: Short-run predictability of US stock returns, out-of-sample

Average return after ”event”, smallest decile Average return after ”event”, largest decile

20 20

10 10

0 0

−10 −10
Rt−1 < 0 Rt−1 ≥ 0 always Rt−1 < 0 Rt−1 ≥ 0 always

US size deciles (daily) 1979:1-2012:12

Strategies (rebalanced daily):

hold stocks if condition is met;
otherwise, hold T-bills

Figure 8.14: Short-run predictability of US stock returns, out-of-sample

mentum type changes (positive autocorrelation). However, always holding the S&P 500
index seems¨ to dominate both strategies—basically because stocks always outperform
T-bills (in this setting). Notice that these strategies assume that you are always invested,
in either stocks or the T-bill. In contrast, Figure 8.14 shows that the momentum strategy
works reasonably well on small stocks.

170
 Out-of-sample R2 , excess returns Average excess return on strategy
0.05
6

4
0
2 historical mean (2y)
historical mean (2y) AR(lag)
AR(lag) 0 always invested
−0.05
1 2 3 4 5 1 2 3 4 5
lag (days) lag (days)

S&P 500 daily excess returns, 1979:1-2013:4

The strategies are based on forecasts
The out-of-sample R2 measures of excess returns:
the fit relative to forecasting 0 (a) forecast> 0: long in stock, short in riskfree
(b) forecast≤ 0: no investment

Figure 8.15: Short-run predictability of US stock returns, out-of-sample

Figure 8.15 shows out-of-sample R2 and average returns of different strategies. The
evidence suggests that an autoregressive model for the daily S&P 500 excess returns per-
forms worse than forecasting zero (and so does using the historical average). In addition,
the strategies based on the predicted excess return (from either the AR model or the histor-
ical returns) are worse than always being invested into the index. Notice that the strategies
here allow for borrowing at the riskfree rate and also for leaving the market, so they are
potentially more powerful than in the earlier figures. Figures 8.16 compares the results for
small and large stocks—and illustrates that there is more predictability for small stocks.
Figures 8.17–8.19 illustrate the out-of-sample performance on long-run returns. Fig-
ure 8.17 shows average one-year return on S&P 500 for different bins of the p/e ratio (at
the beginning of the year). The figure illustrates that buying when the market is underval-
ued (low p/e) might be a winning strategy. To implement simple strategies based on this
observation, 8.18 splits up the observation in (approximately) half: after low and after
high p/e values. The results indicate that buying after low p/e ratios is better than after
high p/e ratios, but that staying invested in the S&P 500 index all the time is better than
sometimes switching over to T-bills. The reason is that even the low stock returns are
higher than the interest rate.
Figure 8.19 studies the out-of-sample R2 for simple forecasting models, and also al-

171
 Out-of-sample R2 , smallest decile Out-of-sample R2 , largest decile
0.05 0.05
historical mean (2y)
AR(lag)

0 0

US size deciles (daily) 1979:1-2012:12

−0.05 −0.05
1 2 3 4 5 1 2 3 4 5
lag (days) lag (days)

Avg excess return on strategy, smallest decile Avg excess return on strategy, largest decile
15 15
historical mean (2y)
AR(lag)
10 10 always invested

5 5

0 0
1 2 3 4 5 1 2 3 4 5
lag (days) lag (days)

Figure 8.16: Short-run predictability of US stock returns, out-of-sample. See Figure 8.15
for details on the strategies.

lows for somewhat more flexible strategies (where we borrow at the riskfree rate and are
allowed to leave the market). The evidence again suggests that it is hard to predict 1-year
S&P 500 returns.

8.3.4 Technical Analysis

Main reference: Bodie, Kane, and Marcus (2002) 12.2; Neely (1997) (overview, foreign
exchange market)
Further reading: Murphy (1999) (practical, a believer’s view); The Economist (1993)
(overview, the perspective of the early 1990s); Brock, Lakonishok, and LeBaron (1992)
(empirical, stock market); Lo, Mamaysky, and Wang (2000) (academic article on return
distributions for “technical portfolios”)

172
 Average excess return after ”event”
8 S&P 500 1-year returns 1957:1-2012:12
The p/e is measured at the beginning of the year
The frequency for each bin is also reported
6

0.12 0.22 0.41 0.12 0.13

0

−2

p/e < 10 10 < p/e < 15 15 < p/e < 20 20 < p/e < 25 25 < p/e average

Figure 8.17: Long-run predictability of US stock returns, out-of-sample

Average return after ”event” Average return on strategy

10 10
freq
riskfree
5 5
0.48 0.52
6.13 3.55
0 0
p/e < 17 p/e > 17 average p/e < 17 p/e > 17 always

Strategies:
S&P 500 1-year returns 1957:1-2012:12
hold stocks if condition is met;
The p/e is measured at the beginning of the year
otherwise, hold T-bills

Figure 8.18: Long-run predictability of US stock returns, out-of-sample

General Idea of Technical Analysis

Technical analysis is typically a data mining exercise which looks for local trends or
systematic non-linear patterns. The basic idea is that markets are not instantaneously

173
 Out-of-sample R2 , excess returns Average excess return on strategy
0.1
0 0.04
−0.1
−0.2 0.02
historical mean (10y)
−0.3 historical mean (10y) E/P
E/P 0 always invested
−0.4
20 40 60 20 40 60
Return horizon (months) Return horizon (months)

Monthly US stock returns in excess of riskfree rate

Estimation is done on moving data window,
forecasts are made out of sample for 1957:1-2012:12

The strategies are based on forecasts

The out-of-sample R2 measures of excess returns:
the fit relative to forecasting 0 (a) forecast > 0: long in stock, short in riskfree
(b) forecast ≤ 0: no investment

Figure 8.19: Long-run predictability of US stock returns, out-of-sample

efficient: prices react somewhat slowly and predictably to news. The logic is essentially
that an observed price move must be due to some news (exactly which one is not very
important) and that old patterns can tell us where the price will move in the near future.
This is an attempt to gather more detailed information than that used by the market as a
whole. In practice, the technical analysis amounts to plotting different transformations
(for instance, a moving average) of prices—and to spot known patterns. This section
summarizes some simple trading rules that are used.

Technical Analysis and Local Trends

Many trading rules rely on some kind of local trend which can be thought of as positive
autocorrelation in price movements (also called momentum1 ).
A moving average rule is to buy if a short moving average (equally weighted or ex-
ponentially weighted) goes above a long moving average. The idea is that event signals
a new upward trend. Let S (L) be the lag order of a short (long) moving average, with
1
In physics, momentum equals the mass times speed.

174
S < L and let b be a bandwidth (perhaps 0.01). Then, a MA rule for period t could be

buy in t if MA t 1 .S/ > MA t 1 .L/.1 C b/

2 3

4 sell in t if MA t 1 .S/ < MA t 1 .L/.1 b/ 5 , where (8.17)

6 7

no change otherwise
MA t 1 .S/ D .p t 1 C : : : C pt S /=S:

The difference between the two moving averages is called an oscillator

oscillator t D MA t .S/ MA t .L/; (8.18)

(or sometimes, moving average convergence divergence, MACD) and the sign is taken
as a trading signal (this is the same as a moving average crossing, MAC).2 A version of
the moving average oscillator is the relative strength index3 , which is the ratio of average
price level (or returns) on “up” days to the average price (or returns) on “down” days—
during the last z (14 perhaps) days. Yet another version is to compare the oscillator t to an
moving average of the oscillator (also called a signal line).
The trading range break-out rule typically amounts to buying when the price rises
above a previous peak (local maximum). The idea is that a previous peak is a resistance
level in the sense that some investors are willing to sell when the price reaches that value
(perhaps because they believe that prices cannot pass this level; clear risk of circular
reasoning or self-fulfilling prophecies; round numbers often play the role as resistance
levels). Once this artificial resistance level has been broken, the price can possibly rise
substantially. On the downside, a support level plays the same role: some investors are
willing to buy when the price reaches that value. To implement this, it is common to let
the resistance/support levels be proxied by minimum and maximum values over a data
2
Yes, the rumour is true: the tribe of chartists is on the verge of developing their very own language.
3
Not to be confused with relative strength, which typically refers to the ratio of two different asset prices
(for instance, an equity compared to the market).

175
window of length L. With a bandwidth b (perhaps 0.01), the rule for period t could be

buy in t if P t > M t 1 .1 C b/
2 3

4 sell in t if P t < m t 1 .1 b/ 5 , where (8.19)

6 7

no change otherwise
Mt 1 D max.p t 1; : : : ; pt S /

mt 1 D min.p t 1 ; : : : ; p t S /:

When the price is already trending up, then the trading range break-out rule may be
replaced by a channel rule, which works as follows. First, draw a trend line through
previous lows and a channel line through previous peaks. Extend these lines. If the price
moves above the channel (band) defined by these lines, then buy. A version of this is to
define the channel by a Bollinger band, which is ˙2 standard deviations from a moving
data window around a moving average.
A head and shoulder pattern is a sequence of three peaks (left shoulder, head, right
shoulder), where the middle one (the head) is the highest, with two local lows in between
on approximately the same level (neck line). (Easier to draw than to explain in a thousand
words.) If the price subsequently goes below the neckline, then it is thought that a negative
trend has been initiated. (An inverse head and shoulder has the inverse pattern.)
Clearly, we can replace “buy” in the previous rules with something more aggressive,
for instance, replace a short position with a long.
The trading volume is also often taken into account. If the trading volume of assets
with declining prices is high relative to the trading volume of assets with increasing prices,
then this is interpreted as a market with selling pressure. (The basic problem with this
interpretation is that there is a buyer for every seller, so we could equally well interpret
the situations as if there is a buying pressure.)

Technical Analysis and Mean Reversion

If we instead believe in mean reversion of the prices, then we can essentially reverse the
previous trading rules: we would typically sell when the price is high. See Figure 8.20
and Table 8.1.
Some investors argue that markets show periods of mean reversion and then periods
with trends—and that both can be exploited. Clearly, the concept of support and resistance

176
levels (or more generally, a channel) is based on mean reversion between these points. A
new trend is then supposed to be initiated when the price breaks out of this band.

Inverted MA rule, S&P 500

1350
MA(3) and MA(25), bandwidth 0.01
Circles at the bottom (top) margin indicates buys (sells)
1300

1250

1200
Long MA (-)
Long MA (+)
Short MA
1150
Jan Feb Mar Apr
1999

Figure 8.20: Examples of trading rules

Mean Std
All days 0:032 1:165
After buy signal 0:054 1:716
After neutral signal 0:047 0:943
After sell signal 0:007 0:903

Table 8.1: Returns (daily, in %) from technical trading rule (Inverted MA rule). Daily
S&P 500 data 1990:1-2013:4

8.4 Security Analysts

Reference: Makridakis, Wheelwright, and Hyndman (1998) 10.1 and Elton, Gruber,
Brown, and Goetzmann (2010) 26

177
 Hold index if MA(3) > MA(25) Hold index if Pt > max(Pt−1 , ..., Pt−5 )

4 4

2 2
SMI
Rule
0 0
2000 2010 2000 2010

Daily SMI data

Weekly rebalancing: hold index or riskfree

Figure 8.21: Examples of trading rules

8.4.1 Evidence on Analysts’ Performance

Makridakis, Wheelwright, and Hyndman (1998) 10.1 shows that there is little evidence
that the average stock analyst beats (on average) the market (a passive index portfolio).
In fact, less than half of the analysts beat the market. However, there are analysts which
seem to outperform the market for some time, but the autocorrelation in over-performance
is weak. The evidence from mutual funds is similar. For them it is typically also found
that their portfolio weights do not anticipate price movements.
It should be remembered that many analysts also are sales persons: either of a stock
(for instance, since the bank is underwriting an offering) or of trading services. It could
well be that their objective function is quite different from minimizing the squared forecast
errors—or whatever we typically use in order to evaluate their performance. (The number
of litigations in the US after the technology boom/bust should serve as a strong reminder
of this.)

8.4.2 Do Security Analysts Overreact?

The paper by Bondt and Thaler (1990) compares the (semi-annual) forecasts (one- and
two-year time horizons) with actual changes in earnings per share (1976-1984) for several
hundred companies. The paper has regressions like

Actual change D ˛ C ˇ.forecasted change/ C residual,

178
and then studies the estimates of the ˛ and ˇ coefficients. With rational expectations (and
a long enough sample), we should have ˛ D 0 (no constant bias in forecasts) and ˇ D 1
(proportionality, for instance no exaggeration).
The main findings are as follows. The main result is that 0 < ˇ < 1, so that the
forecasted change tends to be too wild in a systematic way: a forecasted change of 1% is
(on average) followed by a less than 1% actual change in the same direction. This means
that analysts in this sample tended to be too extreme—to exaggerate both positive and
negative news.

8.4.3 High-Frequency Trading Based on Recommendations from Stock Analysts

Barber, Lehavy, McNichols, and Trueman (2001) give a somewhat different picture.
They focus on the profitability of a trading strategy based on analyst’s recommendations.
They use a huge data set (some 360,000 recommendations, US stocks) for the period
1985-1996. They sort stocks in to five portfolios depending on the consensus (average)
recommendation—and redo the sorting every day (if a new recommendation is published).
They find that such a daily trading strategy gives an annual 4% abnormal return on the
portfolio of the most highly recommended stocks, and an annual -5% abnormal return on
the least favourably recommended stocks.
This strategy requires a lot of trading (a turnover of 400% annually), so trading costs
would typically reduce the abnormal return on the best portfolio to almost zero. A less
frequent rebalancing (weekly, monthly) gives a very small abnormal return for the best
stocks, but still a negative abnormal return for the worst stocks. Chance and Hemler
(2001) obtain similar results when studying the investment advise by 30 professional
“market timers.”

8.4.4 Economic Experts

Several papers, for instance, Bondt (1991) and Söderlind (2010), have studied whether
economic experts can predict the broad stock markets. The results suggests that they
cannot. For instance, Söderlind (2010) show that the economic experts that participate in
the semi-annual Livingston survey (mostly bank economists) (ii) forecast the S&P worse
than the historical average (recursively estimated), and that their forecasts are strongly
correlated with recent market data (which in itself, cannot predict future returns).

179
8.4.5 The Characteristics of Individual Analysts’ Forecasts in Europe

Bolliger (2001) studies the forecast accuracy (earnings per share) of European (13 coun-
tries) analysts for the period 1988–1999. In all, some 100,000 forecasts are studied. It
is found that the forecast accuracy is positively related to how many times an analyst has
forecasted that firm and also (surprisingly) to how many firms he/she forecasts. The ac-
curacy is negatively related to the number of countries an analyst forecasts and also to the
size of the brokerage house he/she works for.

8.4.6 Bond Rating Agencies versus Stock Analysts

Ederington and Goh (1998) use data on all corporate bond rating changes by Moody’s
between 1984 and 1990 and the corresponding earnings forecasts (by various stock ana-
lysts).
The idea of the paper by Ederington and Goh (1998) is to see if bond ratings drive
earnings forecasts (or vice versa), and if they affect stock returns (prices).

1. To see if stock returns are affected by rating changes, they first construct a “normal”
return by a market model:

normal stock return t = ˛ C ˇ  return on stock index t ,

where ˛ and ˇ are estimated on a normal time period (not including the rating
change). The abnormal return is then calculated as the actual return minus the
normal return. They then study how such abnormal returns behave, on average,
around the dates of rating changes. Note that “time” is then measured, individually
for each stock, as the distance from the day of rating change. The result is that there
are significant negative abnormal returns following downgrades, but zero abnormal
returns following upgrades.

2. They next turn to the question of whether bond ratings drive earnings forecasts or
vice versa. To do that, they first note that there are some predictable patterns in
revisions of earnings forecasts. They therefore fit a simple autoregressive model
of earnings forecasts, and construct a measure of earnings forecast revisions (sur-
prises) from the model. They then relate this surprise variable to the bond ratings.
In short, the results are the following:

180
 (a) both earnings forecasts and ratings react to the same information, but there is
also a direct effect of rating changes, which differs between downgrades and
upgrades.
(b) downgrades: the ratings have a strong negative direct effect on the earnings
forecasts; the returns react ever quicker than analysts
(c) upgrades: the ratings have a small positive direct effect on the earnings fore-
casts; there is no effect on the returns

A possible reason for why bond ratings could drive earnings forecasts and prices is
that bond rating firms typically have access to more inside information about firms than
stock analysts and investors.
A possible reason for the observed asymmetric response of returns to ratings is that
firms are quite happy to release positive news, but perhaps more reluctant to release bad
news. If so, then the information advantage of bond rating firms may be particularly large
after bad news. A downgrading would then reveal more new information than an upgrade.
The different reactions of the earnings forecasts and the returns are hard to reconcile.

8.4.7 International Differences in Analyst Forecast Properties

Ang and Ciccone (2001) study earnings forecasts for many firms in 42 countries over the
period 1988 to 1997. Some differences are found across countries: forecasters disagree
more and the forecast errors are larger in countries with low GDP growth, less accounting
disclosure, and less transparent family ownership structure.
However, the most robust finding is that forecasts for firms with losses are special:
forecasters disagree more, are more uncertain, and are more overoptimistic about such
firms.

8.4.8 Analysts and Industries

Boni and Womack (2006) study data on some 170,000 recommendations for a very large
number of U.S. companies for the period 1996–2002. Focusing on revisions of recom-
mendations, the papers shows that analysts are better at ranking firms within an industry
than ranking industries.

181
8.4.9 Insiders

Corporate insiders used to earn superior returns, mostly driven by selling off stocks before
negative returns. (There is little/no systematic evidence of insiders gaining by buying
before high returns.) Actually, investors who followed the insider’s registered transactions
(in the U.S., these are made public six weeks after the reporting period), also used to earn
some superior returns. It seems as if these patterns have more or less disappeared.

Bibliography
Ang, J. S., and S. J. Ciccone, 2001, “International differences in analyst forecast proper-
ties,” mimeo, Florida State University.

Barber, B., R. Lehavy, M. McNichols, and B. Trueman, 2001, “Can investors profit from
the prophets? Security analyst recommendations and stock returns,” Journal of Fi-
nance, 56, 531–563.

Bodie, Z., A. Kane, and A. J. Marcus, 2002, Investments, McGraw-Hill/Irwin, Boston,

5th edn.

Bolliger, G., 2001, “The characteristics of individual analysts’ forecasts in Europe,”

mimeo, University of Neuchatel.

Bondt, W. F. M. D., 1991, “What do economists know about the stock market?,” Journal
of Portfolio Management, 17, 84–91.

Bondt, W. F. M. D., and R. H. Thaler, 1990, “Do security analysts overreact?,” American
Economic Review, 80, 52–57.

Boni, L., and K. L. Womack, 2006, “Analysts, industries, and price momentum,” Journal
of Financial and Quantitative Analysis, 41, 85–109.

Brock, W., J. Lakonishok, and B. LeBaron, 1992, “Simple technical trading rules and the
stochastic properties of stock returns,” Journal of Finance, 47, 1731–1764.

Brockwell, P. J., and R. A. Davis, 1991, Time series: theory and methods, Springer Verlag,
New York, second edn.

182
Campbell, J. Y., A. W. Lo, and A. C. MacKinlay, 1997, The econometrics of financial
markets, Princeton University Press, Princeton, New Jersey.

Campbell, J. Y., and S. B. Thompson, 2008, “Predicting the equity premium out of sam-
ple: can anything beat the historical average,” Review of Financial Studies, 21, 1509–
1531.

Chance, D. M., and M. L. Hemler, 2001, “The performance of professional market timers:
daily evidence from executed strategies,” Journal of Financial Economics, 62, 377–
411.

Cochrane, J. H., 2001, Asset pricing, Princeton University Press, Princeton, New Jersey.

Cuthbertson, K., 1996, Quantitative financial economics, Wiley, Chichester, England.

Ederington, L. H., and J. C. Goh, 1998, “Bond rating agencies and stock analysts: who
knows what when?,” Journal of Financial and Quantitative Analysis, 33, 569–585.

Elton, E. J., M. J. Gruber, S. J. Brown, and W. N. Goetzmann, 2010, Modern portfolio

theory and investment analysis, John Wiley and Sons, 8th edn.

Ferson, W. E., S. Sarkissian, and T. T. Simin, 2003, “Spurious regressions in financial

economics,” Journal of Finance, 57, 1393–1413.

Goyal, A., and I. Welch, 2008, “A comprehensive look at the empirical performance of
equity premium prediction,” Review of Financial Studies 2008, 21, 1455–1508.

Granger, C. W. J., 1992, “Forecasting stock market prices: lessons for forecasters,” Inter-
national Journal of Forecasting, 8, 3–13.

Huberman, G., and S. Kandel, 1987, “Mean-variance spanning,” Journal of Finance, 42,
873–888.

Lo, A. W., H. Mamaysky, and J. Wang, 2000, “Foundations of technical analysis: com-
putational algorithms, statistical inference, and empirical implementation,” Journal of
Finance, 55, 1705–1765.

Makridakis, S., S. C. Wheelwright, and R. J. Hyndman, 1998, Forecasting: methods and

applications, Wiley, New York, 3rd edn.

183
Murphy, J. J., 1999, Technical analysis of the financial markets, New York Institute of
Finance.

Neely, C. J., 1997, “Technical analysis in the foreign exchange market: a layman’s guide,”
Federal Reserve Bank of St. Louis Review.

Priestley, M. B., 1981, Spectral analysis and time series, Academic Press.

Söderlind, P., 2010, “Predicting stock price movements: regressions versus economists,”
Applied Economics Letters, 17, 869–874.

The Economist, 1993, “Frontiers of finance,” pp. 5–20.

184
9 Maximum Likelihood Estimation
Reference: Verbeek (2008) 2 and 4
More advanced material is denoted by a star ( ). It is not required reading.

9.1 Maximum Likelihood

A different route to create a estimate is to maximize the likelihood function.

To understand the principle of maximum likelihood estimation, consider the following
examples.

9.1.1 Example: Estimating the Mean with ML

Suppose we know x t  N.;  2 /, but we don’t know the value of  (for now, assume
we know the variance). Since x t is a random variable, there is a probability of every
observation and the density function of x t is

1 .x t /2
 
1
L D pdf .x t / D p exp ; (9.1)
2 2 2 2

where L stands for “likelihood.” The basic idea of maximum likelihood estimation (MLE)
is to pick model parameters to make the observed data have the highest possible proba-
bility. Here this gives O D x t . This is the maximum likelihood estimator in this example.
What if there are T observations, x1 ; x2 ,...xT ? In the simplest case where xi and xj
are independent, then the joint pdf is just the product of the individual pdfs (for instance,
pdf.xi ; xj / D pdf.xi / pdf.xj /) so

L D pdf .x1 /  pdf .x2 /  :::  pdf .xT / (9.2)

1 .x1 /2 .x2 /2 .xT /2
  
2 T =2
D .2 / exp C C ::: C (9.3)
2 2 2 2
Take logs (log likelihood)
T 1 
ln L D ln.2 2 / /2 C .x2 /2 C ::: C .xT /2 : (9.4)

.x1
2 2 2
185
 The derivative with respect to  is

@ ln L 1
D 2 Œ.x1 / C .x2 / C ::: C .xT / : (9.5)
@ 

To maximize the likelihood function find the value of O that makes @ ln L=@ D 0, which
is the usual sample average

O D .x1 C x2 C ::: C xT / =T: (9.6)

Remark 9.1 (Coding the log likelihood function) Many software packages want just
the likelihood contribution of data point t (not the full sample). Here it is ln L t D
1
2
ln.2 2 / 21 2 .x t /2 .

9.1.2 Example: Estimating the Variance with ML

To estimate the variance, use (9.4) and find the value  2 that makes @ ln L=@ 2 D 0

@ ln L
0D
@ 2
T 1 1 
D C /2 C .x2 /2 C ::: C .xT /2 ; (9.7)

2 .x1
2 2 2 2. 2 /2
so
1
O 2 D .x1 /2 C .x2 /2 C ::: C .xT /2 : (9.8)

T
Notice that we divide by T , not by T 1, so O 2 must be biased, but the bias disappears
as T ! 1

9.1.3 MLE of a Regression

To apply this idea to a (multiple) regression model

yt D ˇ0xt C ut ; (9.9)

we could assume that u t is iid N.0;  2 /. The probability density function of u t is

 
1 1 2 2
pdf .u t / D p exp u = : (9.10)
2 2 2 t

186
Since the errors are independent, we get the joint pdf of the u1 ; u2 ; : : : ; uT by multiplying
the marginal pdfs of each of the errors

L D pdf .u1 /  pdf .u2 /  :::  pdf .uT /

1 u21 u22 u2T
  
2 T =2
D .2 / exp C 2 C ::: C 2 : (9.11)
2 2  

Substitute y t ˇ 0 x t for u t and take logs to get the log likelihood function of the sample
XT
ln L D ln L t , where (9.12)
t D1
1 1 1
2
ln L t D ln .2/ ln. 2 / yt ˇ0xt = 2 : (9.13)
2 2 2
Suppose (for simplicity) that we happen to know the value of  2 . It is then clear that
this likelihood function is maximized by minimizing the last term, which is proportional
to the sum of squared errors: LS is ML when the errors are iid normally distributed (but
only then). (This holds also when we do not know the value of  2 —just slightly messier
to show it.) See Figure 9.1.
Maximum likelihood estimators have very nice properties, provided the basic distri-
butional assumptions are correct, that is, if we maximize the right likelihood function.
In that case, MLE is typically the most efficient/precise estimators (at least in very large
samples). ML also provides a coherent framework for testing hypotheses (including the
Wald, LM, and LR tests).

Example 9.2 Consider the regression model yi D ˇ1 xi Cui , where we (happen to) know
that ui  N.0; 1/.Suppose we have the following data
h i h i h i h i
y1 y2 y3 D 1:5 0:6 2:1 and x1 x2 x3 D 1 0 1 :

Suppose .y1 ; x1 / D . 1:5; 1/. Try different values of ˇ2 on observation 1

ˇ2 u1 Density function value of u1

1:6 1:5 1:6  . 1/ D 0:1 0:40
1:8 1:5 1:8  . 1/ D 0:3 0:38
2:0 1:5 2:0  . 1/ D 0:5 0:35

187
Observation 1 favours ˇ D 1:6; see Figure 9.2. Do the same for observations 2 and 3:

ˇ2 u2 Density function value of u2

1:6 0:6 1:6  0 D 0:6 0:33
1:8 0:6 1:8  0 D 0:6 0:33
2:0 0:6 2:0  0 D 0:6 0:33

ˇ2 u3 Density function value of u3

1:6 2:1 1:6  1 D 0:5 0:35
1:8 2:1 1:8  1 D 0:3 0:38
2:0 2:1 2:0  1 D 0:1 0:40
To sum up, observation 1 favours ˇ D 1:6, observation 2 is neutral, and observation 3
favours ˇ D 2. The estimate is a “compromise” that maximises the joint density (the
product of the individual densities since the ui are independent)

ˇ2 pdf.u1 /  pdf.u2 /  pdf.u3 /

1:6 0:40  0:33  0:35  0:047
1:8 0:38  0:33  0:38  0:048
2:0 0:35  0:33  0:40  0:047

so 1:8 has the highest likelihood value of these three alternatives (it is actually the opti-
mum). See Figure 9.2.

Example 9.3 Consider the simple regression where we happen to know that the intercept
is zero, y t D ˇ1 x t C u t . Suppose we have the following data
h i h i h i h i
y1 y2 y3 D 1:5 0:6 2:1 and x1 x2 x3 D 1 0 1 :

Suppose ˇ2 D 2, then we get the following values for u t D y t 2x t and its square
2 3 2 3 2 3
1:5 2  . 1/ 0:5 0:25
4 0:6 2  0 5 D 4 0:65 with the square 40:365 with sum 0:62:
6 7 6 7 6 7

2:1 21 0:1 0:01

188
 OLS, y = b0 + b1 x + u, b0 = 0 Sum of squared errors
10
2 y: -1.5 -0.6 2.1
x: -1.0 0.0 1.0
1 b1 : 1.8 (OLS)
0 5
y

−1 Data
fitted
−2
0
−1 −0.5 0 0.5 1 0 1 2 3 4
x b1

Log likelihood
0

−5

−10

−15
0 1 2 3 4
b1

Figure 9.1: Example of OLS and ML estimation

Now, suppose instead that ˇ2 D 1:8, then we get

2 3 2 3 2 3
1:5 1:8  . 1/ 0:3 0:09
4 0:6 1:8  0 5 D 4 0:65 with the square 40:365 with sum 0:54:
6 7 6 7 6 7

2:1 1:8  1 0:3 0:09

The latter choice of ˇ2 will certainly give a larger value of the likelihood function (it is
actually the optimum). See Figure 9.1.

189
 Pdf value of u1 = -1.5 −β×(-1) Pdf value of u2 = -0.6 −β×0
0.4 0.4

0.35 0.35

0.3 0.3

0.25 0.25

0.2 0.2
1 1.5 2 2.5 3 1 1.5 2 2.5 3
β β

Pdf value of u3 = 2.1 −β×1 Pdf value of (u1 , u2 , u3 )

0.4

0.35 0.04

0.3
0.02
0.25

0.2 0
1 1.5 2 2.5 3 1 1.5 2 2.5 3
β β

Figure 9.2: Example of OLS and ML estimation

9.1.4 MLE of a Regression with GARCH(1,1) Errors

Consider a regression model where the residuals are uncorrelated across time, but have
time-varying volatility

y t D b 0 x t C u t ; where u t is N.0;  t2 /: (9.14)

The variance follows the GARCH(1,1) process

 t2 D ! C ˛u2t 1 C ˇ t2 1 : (9.15)

(It is assumed that ! > 0; ˛; ˇ  0; and ˛ C ˇ < 1.)

To estimate this model (that is, the parameters in (b; !; ˛; ˇ), we could use a numerical

190
optimization routine to maximize the log likelihood function
T
X 1 1 1 u2t
ln L D L t , where L t D ln .2/ ln  t2 : (9.16)
t D1
2 2 2  t2

This means, in effect, that the optimization routine searches for the values of (b; !; ˛; ˇ)
that makes the value of the log likelihood function as large as possible.

Remark 9.4 To perform the estimation, we also need to supply the optimization routine
with a starting value for 12 and make sure that the restrictions on the GARCH parameters
are fulfilled.

9.2 Key Properties of MLE

No general results on small-sample properties of MLE: can be biased or not...

MLE have very nice asymptotic (large-sample) properties, provided we maximize the
right likelihood function. If so, then

1. MLE is consistent (Pr.jˇO ˇj > any number/ gets very small as T gets large)

2. MLE is the most efficient/precise estimator, at least asymptotically (efficient =

smallest variance)

O are normally distributed,

3. MLE estimates ()
p
T .O / !d N.0; V /; (9.17)
1 @2 ln L
V D I./ with I./ D E =T: (9.18)
@@
(I./ is called the “information matrix”). The information matrix can also be
2
written I./ D E @ @log@Lt , where ln L t is the log likelihood contribution of obser-
vation t.

4. ML also provides a coherent framework for testing hypotheses (including the Wald,
LM, and LR tests).

191
9.2.1 Example of the Information Matrix

Differentiate (9.5) (and assume we know  2 ) to get

@2 ln L T
D : (9.19)
@@ 2
The information matrix is
@2 ln L 1
I./ D E =T D 2 ; (9.20)
@@ 
which we combine with (9.17)–(9.18) to get
p
T .O / !d N.0;  2 / or O !d N.;  2 =T /: (9.21)

This is the standard expression for the distribution of a sample average.

9.3 Three Test Principles

Wald test. Estimate  with MLE, check if O H0 is too large. Example: t-test and F-test
Likelihood ratio test. Estimate  with MLE as usual, estimate again by imposing the
O
H0 restrictions, test if ln L./ ln L.“O with H0 restrictions”/ D 0. Example: compare
the R2 from a model without and with a restriction that some coefficient equals 1/3
Lagrange multiplier test. Estimate  under the H0 restrictions, check if @ ln L=@ D 0
for unconstrained model is true when evaluated at “O with H0 restrictions”

9.4 QMLE

A MLE based on the wrong likelihood function (distribution) may still be useful.
Suppose we use the likelihood function L and get estimates O by

@ ln L
D0 (9.22)
@
If L is wrong, then we are maximizing the wrong thing. With some luck, we still get
reasonable (consistent) estimates.

Example 9.5 (LS and QMLE) In a linear regression, y t D x t0 ˇ C " t , the first order
O t D 0.
condition for MLE based on the assumption that " t  N.0;  2 / is ˙ tTD1 .y t x t0 ˇ/x

192
This has an expected value of zero (at the true parameters), even if the shocks have a, say,
t22 distribution (which would define the correct likelihood function).

The example suggests that if

@ ln L
E D 0; (9.23)
@
then the estimates are still consistent. We are doing quasi-MLE (or pseudo-MLE).
p
With QMLE, T .O / !d N.0; V /, but

@ ln L t @ ln L t 0
  
V D I./ E 1
I./ 1 (9.24)
@ @
Practical implication: this is perhaps a “safer” way of constructing tests—since it is
less restrictive than assuming that we have the exactly correct likelihood function.

Bibliography
Verbeek, M., 2008, A guide to modern econometrics, Wiley, Chichester, 3rd edn.

193
10 ARCH and GARCH
Reference: Bodie, Kane, and Marcus (2005) 13.4
Reference (advanced): Taylor (2005) 8–9; Verbeek (2004) 8; Campbell, Lo, and MacKin-
lay (1997) 12; Franses and van Dijk (2000)

10.1 Heteroskedasticity

10.1.1 Descriptive Statistics of Heteroskedasticity

Time-variation in volatility (heteroskedasticity) is a common feature of macroeconomic

and financial data.
The perhaps most straightforward way to gauge heteroskedasticity is to estimate a
time-series of variances on “rolling samples.” For a zero-mean variable, u t , this could
mean
 t2 D .u2t 1 C u2t 2 C : : : C u2t q /=q; (10.1)

where the latest q observations are used. Notice that  t2 depends on lagged information,
and could therefore be thought of as the prediction (made in t 1) of the volatility in
t . This method can be used for detecting both (general) time variation in volatility—
and the estimates (for instance, over a month) are sometimes called realised volatility.
Alternatively, this method can also be used to gauge seasonality in volatility by estimating
the variance for each “season,” for instance, Mondays.
See Figures 10.1 and 10.2 for examples.
Unfortunately, this method can produce quite abrupt changes in the estimate. An
alternative is to apply an exponentially weighted moving average (EWMA) estimator of
volatility, which uses all data points since the beginning of the sample—but where recent
observations carry larger weights. The weight for lag s is .1 /s where 0 <  < 1, so

 t2 D .1 /.u2t 1 C u2t 2 C 2 u2t 3 C : : :/; (10.2)

194
 Realized std (44 days), annualized EWMA std, annualized, λ = 0.99
50 50
40 40
30 30
20 20
10 10
0 0
1980 1990 2000 2010 1980 1990 2000 2010

S&P 500 (daily) 1954:1-2013:4

AR(1) of excess returns EWMA std, annualized, λ = 0.9

50
40
30
20
10
0
1980 1990 2000 2010

Figure 10.1: Standard deviation

Standard deviation, different weekdays Standard deviation, different hours

0.04 0.05

0.04
0.035 0.03

0.02
0.03
Mon Tue Wed Thu Fri 0 6 12 18
Hour

5-minute data on EUR/USD changes, 1998:1-2011:11

Sample size: 1045414

Figure 10.2: Standard deviation for EUR/USD exchange rate changes

which can also be calculated in a recursive fashion as

 t2 D .1 /u2t 1 C  t2 1 : (10.3)

195
 Weight on lagged data (u2t−s ) in EWMA estimate of volatility

λ = 0.99
0.05 λ = 0.94

0.04
σt2 = (1 − λ)(u2t−1 + λu2t−2 + λ2 u2t−3 + ...)
0.03

0.02

0.01

0
0 20 40 60 80 100
lag, s

Figure 10.3: Weights on old data in the EMA approach to estimate volatility

The initial value (before the sample) could be assumed to be zero or (perhaps better) the
unconditional variance in a historical sample.
This methods is commonly used by practitioners. For instance, the RISK Metrics
uses this method with  D 0:94 for use on daily data. Alternatively,  can be chosen to
minimize some criterion function like ˙ tTD1 .u2t  t2 /2 .
See Figure 10.3 for an illustration of the weights.

10.1.2 Predicting Realised Volatility

Volatility is often predictable, at least for horizons up to a couple of months. See Tables
10.1–10.2 for examples of very simple prediction equations.

10.1.3 Heteroskedastic Residuals in a Regression

Suppose we have a regression model

y t D b0 C x1t b1 C x2t b2 C


C xk t bk C " t ; where (10.4)

E " t D 0 and Cov.xi t ; " t / D 0:

196
 VIX (solid) and realized volatility (dashed)

80 The realized volatility is

measured over the last 30 days
70

60

50

40

30

20

10

1990 1995 2000 2005 2010

Figure 10.4: VIX and realized volatility (variance)

(1) (2) (3)

lagged RV 0:75 0:26
.11:02/ .2:20/
lagged VIX 0:91 0:64
.12:60/ .7:54/
constant 3:97 2:62 1:20
.4:29/ . 2:06/ . 1:55/
R2 0:56 0:61 0:62
obs 5825:00 5845:00 5825:00

Table 10.1: Regression of 22-day realized S&P return volatility 1990:1-2013:5. All daily
observations are used, so the residuals are likely to be autocorrelated. Numbers in paren-
theses are t-stats, based on Newey-West with 30 lags.

In the standard case we assume that " t is iid (independently and identically distributed),
which rules out heteroskedasticity.
In case the residuals actually are heteroskedastic, least squares (LS) is nevertheless a
useful estimator: it is still consistent (we get the correct values as the sample becomes

197
 RV(EUR) RV(GBP) RV(CHF) RV(JPY)
lagged RV(EUR) 0:63
.6:94/
lagged RV(GBP) 0:72
.10:07/
lagged RV(CHF) 0:33
.2:50/
lagged RV(JPY) 0:56
.4:92/
constant 0:06 0:04 0:25 0:13
.1:71/ .1:40/ .3:36/ .1:93/
D(Tue) 0:12 0:08 0:13 0:11
.10:60/ .6:01/ .4:00/ .3:37/
D(Wed) 0:11 0:09 0:08 0:13
.8:60/ .6:56/ .3:64/ .4:09/
D(Thu) 0:12 0:09 0:13 0:15
.9:25/ .5:38/ .6:26/ .3:40/
D(Fri) 0:13 0:07 0:14 0:12
.6:13/ .3:83/ .8:47/ .3:81/
R2 0:40 0:52 0:11 0:31
obs 3629:00 3629:00 3629:00 3629:00

Table 10.2: Regression of daily realized variance 1998:1-2011:11. All exchange rates
are against the USD. The daily variances are calculated from 5 minute data. Numbers in
parentheses are t-stats, based on Newey-West with 1 lag.

really large)—and it is reasonably efficient (in terms of the variance of the estimates),
although not the most efficient (MLE is). However, the standard expression for the stan-
dard errors (of the coefficients) is (except in a special case, see below) not correct. This is
illustrated in Figure 10.5.
There are two ways to handle this problem. First, we could use some other estimation
method than LS that incorporates the structure of the heteroskedasticity. For instance,
combining the regression model (10.4) with an ARCH structure of the residuals—and
estimate the whole thing with maximum likelihood (MLE) is one way. As a by-product
we get the correct standard errors provided, of course, the assumed distribution is cor-
rect. Second, we could stick to OLS, but use another expression for the variance of the
coefficients: a “heteroskedasticity consistent covariance matrix,” among which “White’s

198
 Std of LS slope coefficient under heteroskedasticity
0.1
OLS formula
White’s
Simulated
0.08

0.06

0.04 Model: yt = 0.9xt + ǫt ,

where ǫt ∼ N (0, ht ), with ht = 0.5exp(αx2t )

bLS is the LS estimate of b in

0.02 yt = a + bxt + ut

Number of simulations: 25000

0
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2
α (effect of regressor on variance)

Figure 10.5: Variance of OLS estimator, heteroskedastic errors

covariance matrix” is the most common.

To test for heteroskedasticity, we can use White’s test of heteroskedasticity. The null
hypothesis is homoskedasticity, and the alternative hypothesis is the kind of heteroskedas-
ticity which can be explained by the levels, squares, and cross products of the regressors—
clearly a special form of heteroskedasticity. The reason for this specification is that if the
squared residual is uncorrelated with w t , then the usual LS covariance matrix applies—
even if the residuals have some other sort of heteroskedasticity (this is the special case
mentioned before).
To implement White’s test, let wi be the squares and cross products of the regressors.
For instance, if the regressors include .1; x1t ; x2t / then w t is the vector (1; x1t ; x2t ; x1t
2 2
; x1t x2t ; x2t )—
since .1; x1t ; x2t /  1 is .1; x1t ; x2t / and 1  1 D 1. The test is then to run a regression of
squared fitted residuals on w t
"O2t D w t0 C vi ; (10.5)

and to test if all the slope coefficients (not the intercept) in are zero. (This can be done
be using the fact that TR2  p2 , p D dim.wi / 1:)

199
10.1.4 Autoregressive Conditional Heteroskedasticity (ARCH)

Autoregressive heteroskedasticity is a special form of heteroskedasticity—and it is often

found in financial data which shows volatility clustering (calm spells, followed by volatile
spells, followed by...).
To test for ARCH features, Engle’s test of ARCH is perhaps the most straightforward.
It amounts to running an AR(q) regression of the squared zero-mean variable (here de-
noted u t )
u2t D ! C a1 u2t 1 C : : : C aq u2t q C v t ; (10.6)

Under the null hypothesis of no ARCH effects, all slope coefficients are zero and the R2
of the regression is zero. (This can be tested by noting that, under the null hypothesis,
TR2  2q .) This test can also be applied to the fitted residuals from a regression like
(10.4). However, in this case, it is not obvious that ARCH effects make the standard
expression for the LS covariance matrix invalid—this is tested by White’s test as in (10.5).

10.2 ARCH Models

This section discusses the Autoregressive Conditional Heteroskedasticity (ARCH) model.

It is a model of how volatility depends on recent volatility.
There are two basic reasons for being interested in an ARCH model. First, if residuals
of the regression model (10.4) have ARCH features, then an ARCH model (that is, a
specification of exactly how the ARCH features are generated) can help us estimate the
regression model by maximum likelihood. Second, we may be interested in understanding
the ARCH features more carefully, for instance, as an input in a portfolio choice process
or option pricing.

10.2.1 Properties of ARCH(1)

In the ARCH(1) model the residual in the regression equation (10.4), or some other zero-
mean variable, can be written

u t  N.0;  t2 /, with (10.7)

 t2 D ! C ˛u2t 1 , with ! > 0 and 0  ˛ < 1: (10.8)

200
The non-negativity restrictions on ! and ˛ are needed in order to guarantee  t2 > 0. The
upper bound ˛ < 1 is needed in order to make the conditional variance stationary (more
later).
It is clear that the unconditional distribution of u t is non-normal. While the condi-
tional distribution of u t is N.0;  t2 /, the unconditional distribution of u t is a mixture of
normal distributions with different (and random) variances. It can be shown that the result
is a distribution which has fatter tails than a normal distribution with the same variance
(excess kurtosis)—which is a common feature of financial data.
It is straightforward to show that the ARCH(1) model implies that we in period t can
forecast the future conditional variance in t C s as (since  t2C1 is known in t. )
!
E t  t2Cs D N 2 C ˛ s 1
 t2C1 N 2 , with N 2 D (10.9)


;
1 ˛
where N 2 is the unconditional variance. The conditional volatility behaves like an AR(1),
and 0  ˛ < 1 is necessary to keep it positive and stationary.
See Figure 10.6 for an illustration of the fitted volatilities.
Proof. (of (10.9)) Notice that E t  t2C2 D !C˛ E t v t2C1 E t  t2C1 since v t is independent
of  t . Morover, E t v t2C1 D 1 and E t  t2C1 D  t2C1 (known in t). Combine to get E t  tC2 2
D
! C ˛ t C1 . Similarly, E t  t C3 D ! C ˛ E t  t C2 . Substitute for E t  t C2 to get E t  tC3 D
2 2 2 2 2

! C ˛.! C ˛ t2C1 /, which can be written as (10.9). Further periods follow the same
pattern.

10.2.2 Estimation of the ARCH(1) Model

The most common way to estimate the model is to assume that v t iid N.0; 1/ and to
set up the likelihood function. The log likelihood is easily found, since the model is
conditionally Gaussian. It is
T
X 1 1 1 u2t
ln L D L t , where L t D ln .2/ ln  t2 : (10.10)
t D1
2 2 2  t2

The estimates are found by maximizing the likelihood function (by choosing the parame-
ters). This is done by a numerical optimization routine, which should preferably impose
the constraints in (10.8).
If u t is just a zero-mean variable (no regression equation), then this just amounts to

201
 ARCH std, annualized GARCH std, annualized
50 50
40 40
30 30
20 20
10 10
0 0
1980 1990 2000 2010 1980 1990 2000 2010

S&P 500 (daily) 1954:1-2013:4

AR(1) of excess returns

with ARCH(1) or GARCH(1,1) errors

AR(1) coef: 0.09

ARCH coef: 0.31
GARCH coefs: 0.08 0.91

Figure 10.6: ARCH and GARCH estimates

choosing the parameters (! and ˛/ in (10.8). Instead, if u t is a residual from a regression

equation (10.4), then we instead need to choose both the regression coefficients (b0 ; :::; bk )
in (10.4) and the parameters (! and ˛) in (10.8). In either case, we need a starting value of
12 D ! C ˛u20 . This most common approach is to use the first observation as a “starting
point,” that is, we actually have a sample from (t D) 0 to T , but observation 0 is only used
to construct a starting value of 12 , and only observations 1 to T are used in the calculation
of the likelihood function value.
Notice that if we estimate a regression function and an ARCH model simultaneous
with MLE, then we automatically get the right standard errors of the regression coeffi-
cients from the information matrix. There is no need for using any adjusted (“White”)
values.

Remark 10.1 (Regression with ARCH(1) residuals) To estimate the full model (10.4) and
(10.8) by ML, we can do as follows.
First, guess values of the parameters b0 ; :::; bk , and !, and ˛. The guess of b0 ; :::; bk can
be taken from an LS estimation of (10.4), and the guess of ! and ˛ from an LS estimation
of "O2t D ! C ˛ "O2t 1 C " t where "O t are the fitted residuals from the LS estimation of (10.4).

202
Second, loop over the sample (first t D 1, then t D 2, etc.) and calculate u t D "O t from
(10.4) and  t2 from (10.8). Plug in these numbers in (10.10) to find the likelihood value.
Third, make better guesses of the parameters and do the second step again. Repeat until
the likelihood value converges (at a maximum).

Remark 10.2 (Imposing parameter constraints on ARCH(1).) To impose the restrictions

Q ˛/
in (10.8) when the previous remark is implemented, iterate over values of .b; !; Q and
let ! D !Q 2 and ˛ D exp.˛/=Œ1
Q C exp.˛/.
Q

It is sometimes found that the standardized values of u t , u t = t , still have too fat tails
compared with N.0; 1/. This would violate the assumption about a normal distribution in
(10.10). Estimation using other likelihood functions, for instance, for a t-distribution can
then be used. Or the estimation can be interpreted as a quasi-ML (is typically consistent,
but requires different calculation of the covariance matrix of the parameters).
It is straightforward to add more lags to (10.8). For instance, an ARCH(p) would be

 t2 D ! C ˛1 u2t 1 C : : : C ˛p u2t p: (10.11)

The form of the likelihood function is the same except that we now need p starting values
and that the upper boundary constraint should now be ˙jpD1 ˛j  1.

10.3 GARCH Models

Instead of specifying an ARCH model with many lags, it is typically more convenient to
specify a low-order GARCH (Generalized ARCH) model. The GARCH(1,1) is a simple
and surprisingly general model, where the volatility follows

 t2 D ! C ˛u2t 1 C ˇ t2 1 ,with (10.12)

! > 0; ˛; ˇ  0; and ˛ C ˇ < 1:

The non-negativity restrictions are needed in order to guarantee that  t2 > 0 in all
periods. The upper bound ˛ C ˇ < 1 is needed in order to make the  t2 stationary and
therefore the unconditional variance finite.

Remark 10.3 The GARCH(1,1) has many similarities with the exponential moving aver-
age estimator of volatility (10.3). The main differences are that the exponential moving

203
 GARCH std, annualized EWMA std, annualized, λ = 0.99
50 50
40 40
30 30
20 20
10 10
0 0
1980 1990 2000 2010 1980 1990 2000 2010

EWMA std, annualized, λ = 0.9

S&P 500 (daily) 1954:1-2013:4 50
40
AR(1) of excess returns
with GARCH(1,1) errors 30
AR(1) coef: 0.09 20
GARCH coefs: 0.08 0.91 10
0
1980 1990 2000 2010

Figure 10.7: Conditional standard deviation, estimated by GARCH(1,1) model

Daily DAX returns 1991:1-2013:4

Std of DAX, GARCH(1,1)
60 model: ut is N (0, σt2 ) with
σt2 = α0 + α1 u2t−1 + β1 σt−1
2

(ut is the de-meaned return)

40
Std

Coef Std err

α0 0.031 0.010
20
α1 0.082 0.011
β1 0.901 0.011
0
1995 2000 2005 2010

Figure 10.8: Results for a univariate GARCH model

average does not have a constant and volatility is non-stationary (the coefficients sum to
unity).

See Figure 10.7 for an example.

204
 The GARCH(1,1) corresponds to an ARCH.1/ with geometrically declining weights,
which is seen by solving (10.12) recursively by substituting for  t2 1 (and then  t2 2 ,  t2 3 ,
...)
1
2 ! X
t D C˛ ˇ j u2t 1 j : (10.13)
1 ˇ j D0

This suggests that a GARCH(1,1) might be a reasonable approximation of a high-order

ARCH.
Proof. (of (10.13)) Substitute for  t2 1 in (10.12), and then for  t2 2 , etc

t 1 2
‚ …„ ƒ

2 2 2 2
 t D ! C ˛u t 1 C ˇ ! C ˛u t 2 C ˇ t 2
D ! .1 C ˇ/ C ˛u2t 1 C ˇ˛u2t 2 C ˇ 2  t2 2
:
D ::

and we get (10.13).

Also, the GARCH(1,1) model implies that we in period t can forecast the future con-
ditional variance ( t2Cs ) as
!
E t  t2Cs D N 2 C .˛ C ˇ/s 1
 t2C1 N 2 , with N 2 D (10.14)


;
1 ˛ ˇ

which is of the same form as for the ARCH model (10.9), but where the sum of ˛ and ˇ
is like an AR(1) parameter.
Proof. (of (10.14)) Notice that E t  t2C2 D ! C ˛ E t v t2C1 E t  tC1
2
C ˇ t2C1 since v t is
independent of  t . Morover, E t v t2C1 D 1 and E t  t2C1 D  t2C1 (known in t). Combine to
get E t  t2C2 D ! C .˛ C ˇ/ tC1 2
. Similarly, E t  t2C3 D ! C .˛ C ˇ/ E t  t2C2 . Substitute
for E t  t2C2 to get E t  t2C3 D ! C .˛ C ˇ/Œ! C .˛ C ˇ/ tC1 2
, which can be written as
(10.14). Further periods follow the same pattern.
To estimate the model consisting of (10.4) and (10.12) we can still use the likelihood
function (10.10) and do a MLE (but we now have to choose a value of ˇ as well). We
typically create the starting value of u20 as in the ARCH(1) model, but this time we also
need a starting value of 02 . It is often recommended to use 02 D Var.u t /.

Remark 10.4 (Imposing parameter constraints on GARCH(1,1).) To impose the restric-

Q ˛;
tions in (10.12), iterate over values of .b; !; Q and let ! D ! 2 , ˛ D exp.˛/=Œ1
Q ˇ/ Q C

205
 QQ plot of AR residuals QQ plot of AR&GARCH residuals
6 6
S&P 500 returns (daily)
4 1954:1-2013:4 4
Empirical quantiles

Empirical quantiles
2 0.1th to 99.9th 2
0 percentile 0
−2 AR(1) −2 AR(1)&GARCH(1,1)
−4 Stand. residuals: −4 Stand. residuals:
ut /σ ut /σt
−6 −6
−5 0 5 −5 0 5
Quantiles from N(0,1), % Quantiles from N(0,1), %

Figure 10.9: QQ-plot of residuals

exp.˛/ Q and ˇ D exp.ˇ/=Œ1

Q C exp.ˇ/; Q C exp.˛/ Q
Q C exp.ˇ/.

See Figure 10.9 for evidence of how the residuals become more normally distributed
once the heteroskedasticity is handled.

Remark 10.5 (Value at Risk) The value at risk (as fraction of the investment) at the ˛
level (say, ˛ D 0:95) is VaR˛ D cdf 1 .1 ˛/, where cdf 1 ./ is the inverse of the cdf—so
cdf 1 .1 ˛/ is the 1 ˛ quantile of the return distribution. For instance, VaR0:95 D 0:08
says that there is only an 5% chance that the loss will be greater than 8% of the investment.
See Figure 10.10 for an illustration. When the return has an N.;  2 / distribution, then
VaR95% D . 1:64/. See Figure 10.11 for an example of time-varying VaR, based
on a GARCH model.

10.4 Non-Linear Extensions

A very large number of extensions have been suggested. I summarize a few of them,
which can be estimated by using the likelihood function (10.10) to do a MLE.
An asymmetric GARCH (Glosten, Jagannathan, and Runkle (1993)) can be con-

206
 Value at risk and density of returns

VaR95% = − (the 5% quantile)

-VaR95% Return

Figure 10.10: Value at risk

GARCH std, % Value at Risk95% (one day), %

5 10
4
3
5
2
1
0 0
1980 1990 2000 2010 1980 1990 2000 2010
S&P 500, daily data 1954:1-2013:4 The VaR is based on N()

The horizontal lines are from the

unconditional distribution

Figure 10.11: Conditional volatility and VaR

structed as

 t2 D ! C ˛u2t 1 C ˇ t2 1 C ı.u t 1 > 0/u2t 1 , where (10.15)

(
1 if q is true
ı.q/ D
0 else.

207
This means that the effect of the shock u2t 1 is ˛ if the shock was negative and ˛ C if
the shock was positive. With < 0, volatility increases more in response to a negative
u t 1 (“bad news”) than to a positive u t 1 .
The EGARCH (exponential GARCH, Nelson (1991)) sets

ju t 1j ut 1
ln  t2 D ! C ˛ C ˇ ln  t2 1 C (10.16)
t 1 t 1

Apart from being written in terms of the log (which is a smart trick to make  t2 > 0 hold
without any restrictions on the parameters), this is an asymmetric model. The ju t 1 j term
is symmetric: both negative and positive values of u t 1 affect the volatility in the same
way. The linear term in u t 1 modifies this to make the effect asymmetric. In particular,
if < 0, then the volatility increases more in response to a negative u t 1 (“bad news”)
than to a positive u t 1 .
Hentschel (1995) estimates several models of this type, as well as a very general
formulation on daily stock index data for 1926 to 1990 (some 17,000 observations). Most
standard models are rejected in favour of a model where  t depends on  t 1 and ju t 1
bj3=2 .

10.5 (G)ARCH-M

It can make sense to let the conditional volatility enter the mean equation—for instance,
as a proxy for risk which may influence the expected return.
We modify the “mean equation” (10.4) to include the conditional variance  t2 (taken
from any of the models for heteroskedasticity) as a regressor

y t D x t0 b C ' t2 C u t : (10.17)

Note that  t2 is predetermined, since it is a function of information in t 1. This model

can be estimated by using the likelihood function (10.10) to do MLE.

Remark 10.6 (Coding of (G)ARCH-M) We can use the same approach as in Remark
10.1, except that we use (10.17) instead of (10.4) to calculate the residuals (and that we
obviously also need a guess of ').

208
Example 10.7 (Theoretical motivation of GARCH-M) A mean variance investor solves

max˛ E Rp p2 k=2; subject to

Rp D ˛Rm C .1 ˛/Rf ;

where Rm is the return on the risky asset (the market index) and Rf is the riskfree return.
The solution is
1 E.Rm Rf /
˛D :
k m2
In equilibrium, this weight is one (since the net supply of bonds is zero), so we get

E.Rm Rf / D km2 ;

which says that the expected excess return is increasing in both the market volatility and
risk aversion (k).

10.6 Multivariate (G)ARCH

10.6.1 Different Multivariate Models

This section gives a brief summary of some multivariate models of heteroskedasticity.

Suppose u t is an n  1 vector. For instance, u t could be the residuals from n different
regressions or just n different demeaned return series.
We define the conditional (on the information set in t 1) covariance matrix of u t as

˙t D Et 1 u t u0t : (10.18)

Remark 10.8 (The vech operator) vech(A) of a matrix A gives a vector with the elements
on and below the principal diagonal
2 A3stacked on top of each other (column wise). For
" # a11
a11 a12
instance, vech D 4 a21 5.
6 7
a21 a22
a22

It may seem as if a multivariate (matrix) version of the GARCH(1,1) model would

be simple, but it is not. The reason is that it would contain far too many parameters.
Although we only need to care about the unique elements of ˙ t , that is, vech.˙ t /, this

209
still gives very many parameters

vech.˙ t / D C C Avech.u t 0
1ut 1/ C Bvech.˙ t 1 /: (10.19)

For instance, with n D 2 we have

2 3 2 2 3 2 3
11;t u1;t 1 11;t 1

4 21;t 5 D C C A 4 u1;t 1 u2;t 1 5 C B 4 21;t 5; (10.20)

6 7 6 7 6 7
1

22;t u22;t 1 22;t 1

where C is 3  1, A is 3  3, and B is 3  3. This gives 21 parameters, which is already

hard to manage. We have to limit the number of parameters. We also have to find a
way to impose restrictions so ˙ t is positive definite (compare the restrictions of positive
coefficients in (10.12)).

The Diagonal Model

The diagonal model assumes that A and B are diagonal. This means that every element
of ˙ t follows a univariate process. With n D 2 we have
2 3 2 3 2 32 2 3 2 32 3
11;t c1 a1 0 0 u1;t 1 b1 0 0 11;t 1
4 21;t 5 D 4c2 5 C 4 0 a2 0 5 4 u1;t 1 u2;t 1 5 C 4 0 b2 0 5 4 21;t 1 5 ;
6 7 6 7 6 76 7 6 76 7

22;t c3 0 0 a3 u22;t 1 0 0 b3 22;t 1

(10.21)
which gives 3 C 3 C 3 D 9 parameters (in C , A, and B, respectively). To make sure that
˙ t is positive definite we have to impose further restrictions. The obvious drawback of
this model is that there is no spillover of volatility from one variable to another.

The Constant Correlation Model

The constant correlation model assumes that every variance follows a univariate GARCH
process and that the conditional correlations are constant. With n D 2 the covariance
matrix is
" # "p #" # "p #
11;t 12;t 11;t 0 1 12 11;t 0
D p p (10.22)
12;t 22;t 0 22;t 12 1 0 22;t

210
and each of 11t and 22t follows a GARCH process. Assuming a GARCH(1,1) as in
(10.12) gives 7 parameters (2  3 GARCH parameters and one correlation), which is
convenient. The price is, of course, the assumption of no movements in the correlations.
To get a positive definite ˙ t , each individual GARCH model must generate a positive
variance (same restrictions as before), and that all the estimated (constant) correlations
are between 1 and 1.

Remark 10.9 (Estimating the constant correlation model) A quick (and dirty) method
for estimating is to first estimate the individual GARCH processes and then estimate the
p p
correlation of the standardized residuals u1t = 11;t and u2t = 22;t .

By also specifying how the correlation can change over time, we get a dynamic cor-
relation model. It is slightly harder to estimate.
See Figure 10.12 for an illustration and Figure 10.13 for a comparison with the EWMA
approach.

Bibliography
Bodie, Z., A. Kane, and A. J. Marcus, 2005, Investments, McGraw-Hill, Boston, 6th edn.

Campbell, J. Y., A. W. Lo, and A. C. MacKinlay, 1997, The econometrics of financial

markets, Princeton University Press, Princeton, New Jersey.

Franses, P. H., and D. van Dijk, 2000, Non-linear time series models in empirical finance,
Cambridge University Press.

Glosten, L. R., R. Jagannathan, and D. Runkle, 1993, “On the relation between the ex-
pected value and the volatility of the nominal excess return on stocks,” Journal of Fi-
nance, 48, 1779–1801.

Hentschel, L., 1995, “All in the family: nesting symmetric and asymmetric GARCH
models,” Journal of Financial Economics, 39, 71–104.

Nelson, D. B., 1991, “Conditional heteroskedasticity in asset returns,” Econometrica, 59,

347–370.

211
 Std of DAX, GARCH(1,1) Std of FTSE, GARCH(1,1)
60 60

40 40
Std

Std
20 20

0 0
1995 2000 2005 2010 1995 2000 2005 2010

Daily returns 1991:1-2013:4

Correlation of FTSE 100 and DAX 30
1 The standard deviations are annualized

DCC parameters: 0.046 0.949

Corr

0.5

DCC
CC
0
1995 2000 2005 2010

Figure 10.12: Results for multivariate GARCH models

Taylor, S. J., 2005, Asset price dynamics, volatility, and prediction, Princeton University
Press.

Verbeek, M., 2004, A guide to modern econometrics, Wiley, Chichester, 2nd edn.

212
 Correlation of FTSE 100 and DAX 30 Correlation of FTSE 100 and DAX 30
1 1
Corr

Corr
0.5 0.5

EWMA (λ = 0.95) EWMA (λ = 0.975)

0 0
1995 2000 2005 2010 1995 2000 2005 2010

Correlation of FTSE 100 and DAX 30

1
Corr

0.5

EWMA (λ = 0.99)
0
1995 2000 2005 2010

Figure 10.13: Time-varying correlations (different EWMA estimates)

213
11 Risk Measures
Reference: Hull (2006) 18; McDonald (2006) 25; Fabozzi, Focardi, and Kolm (2006)
4–5; McNeil, Frey, and Embrechts (2005); Alexander (2008)

11.1 Value at Risk

Value at risk and density of returns

VaR95% = − (the 5% quantile)

-VaR95% Return

Figure 11.1: Value at risk

The mean-variance framework is often criticized for failing to distinguish between

downside (considered to be risk) and upside (considered to be potential).
The 95% Value at Risk (VaR95% ) is a number such that there is only a 5% chance that
the loss ( R) is larger that VaR95%

Pr. R  VaR95% / D 5%: (11.1)

Here, 95% is the confidence level of the VaR. More generally, a there is only a 1 ˛

214
chance that the loss ( R) is larger that VaR˛ (the confidence level is ˛)

Pr. R  VaR˛ / D 1 ˛: (11.2)

Clearly, R VaR˛ is true when (and only when) R  VaR˛ , so (11.2) can also be
expressed as
Pr.R  VaR˛ / D cdfR . VaR˛ / D 1 ˛; (11.3)

where cdfR ./ is the cumulative distribution function of the returns. This says that VaR˛
is a number such that there is only a 1 ˛ (5%, say) chance that the return is below it.
See Figures 11.1–11.2 for illustrations. Using (11.3) allows us to work directly with the
return distribution (not the loss distribution), which is often convenient.

Example 11.1 (Quantile of a distribution) The 0.05 quantile is the value such that there
is only a 5% probability of a lower number, Pr.R quantile0:05 / D 0:05.

We can solve (11.3) for the value at risk, VaR˛ , as

VaR˛ D cdfR1 .1 ˛/, (11.4)

where cdfR1 ./ is the inverse of the cumulative distribution function of the returns, so
cdfR1 .1 ˛/ is the 1 ˛ quantile (or “critical value”) of the return distribution. For
instance, VaR95% is the (negative of the) 0:05 quantile of the return distribution.
To convert the value at risk into value terms (CHF, say), just multiply the VaR for
returns with the value of the investment (portfolio).
If the return is normally distributed, R  N.;  2 / and c1 ˛ is the 1 ˛ quantile of
a N(0,1) distribution (for instance, 1:64 for 1 ˛ D 0:05), then

VaR˛ D . C c1 ˛ /: (11.5)

This is illustrated in Figure 11.4.

Remark 11.2 (Critical values of N.;  2 /) If R  N.;  2 /, then there is a 5% proba-

bility that R   1:64, a 2.5% probability that R   1:96 , and a 1% probability
that R   2:33 .

Example 11.3 (VaR with R  N.;  2 /) If daily returns have  D 8% and  D 16%,
then the 1-day VaR95% D .0:08 1:64  0:16/  0:18; we are 95% sure that we will not

215
 Vaue at Risk, diffferent confidence levels
0.4
R ∼ N(0.08, 0.162 )

0.3
VaR/investment

0.2

0.1

0
0.9 0.92 0.94 0.96 0.98 1
Confidence level of VaR

Figure 11.2: Value at risk, different probability levels

loose more than 18% of the investment over one day, that is, VaR95% D 0:18. Similarly,
VaR97:5% D .0:08 1:96  0:16/  0:24.

Figure 11.3 shows the distribution and VaRs (for different probability levels) for the
daily S&P 500 returns. Two different VaRs are shown: based on a normal distribution
and as the empirical VaR (from the empirical quantiles of the distribution). While these
results are interesting, they are just time-averages in the sense of being calculated from
the unconditional distribution: time-variation in the distribution is not accounted for.
Figure 11.5 illustrates the VaR calculated from a time series model (to be precise,
an AR(1)+GARCH(1,1) model) for daily S&P returns. In this case, the VaR changes
from day to day as both the mean return (the forecast) as well as the standard error (of
the forecast error) do. Since the volatility clearly changes over time, this is crucial for a
reliable VaR model.
Notice that the value at risk in (11.5), that is, when the return is normally distributed,
is a strictly increasing function of the standard deviation (and the variance). This follows
from the fact that c1 ˛ < 0 (provided 1 ˛ < 50%, which is the relevant case). Minimiz-
ing the VaR at a given mean return therefore gives the same solution (portfolio weights)
as minimizing the variance at the same given mean return. In other cases, the portfolio
choice will be different (and perhaps complicated to perform).

216
 Distribution of daily S&P 500,1957:1-2013:5

normal pdf
0.6

0.5 dashed: -VaR from N()

solid: -VaR from empirical quantile
99.5%, 99% and 95% levels
0.4

0.3

0.2

0.1

0
−5 −4 −3 −2 −1 0 1 2 3 4 5
Daily excess return, %

Figure 11.3: Return distribution and VaR for S&P 500

Example 11.4 (VaR and regulation of bank capital) Bank regulations have used 3 times
the 99% VaR for 10-day returns as the required bank capital.

Notice that the return distribution depends on the investment horizon, so a value
at risk measure is typically calculated for a stated investment period (for instance, one
day). Multi-period VaRs are calculated by either explicitly constructing the distribution
of multi-period returns, or by making simplifying assumptions about the relation between
returns in different periods (for instance, that they are iid).

Remark 11.5 (Multi-period VaR) If the returns are iid, then a q-period return has the
mean q and variance q 2 , where  and  2 are the mean and variance of the one-period
p
returns respectively. If the mean is zero, then the q-day VaR is q times the one-day VaR.

Backtesting a VaR model amounts to checking if (historical) data fits with the VaR
numbers. For instance, we first find the VaR95% and then calculate what fraction of re-
turns that is actually below (the negative of ) this number. If the model is correct it should
be 5%. We then repeat this for VaR96% —only 4% of the returns should be below (the

217
 Density of N(0,1) Density of N(8, 162 )
5% quantile is c = −1.64 3 5% quantile is µ + cσ = −18
0.4
0.3 2
pdf

pdf
0.2
1
0.1
0 0
−3 c 0 3 −40 0 40
x R

cdf of N(8, 162 ) Inverse of cdf of N(8, 162 )

1
40
cdf

0.5 0

−40
0
−40 0 40 0 0.5 1
R cdf

Figure 11.4: Finding critical value of N(, 2 ) distribution

negative of ) this number. Figures 11.6–11.7 show results from backtesting a VaR model
where the volatility follows a GARCH process. It suggests that a GARCH model (to cap-
ture the time varying volatility), combined with the assumption that the return is normally
distributed (but with time-varying parameters), works relatively well.
The VaR concept has been criticized for having poor aggregation properties. In par-
ticular, the VaR for a portfolio is not necessarily (weakly) lower than the portfolio of the
VaRs, which contradicts the notion of diversification benefits. (To get this unfortunate
property, the return distributions must be heavily skewed.)
See Table 11.1 for an empirical comparison of the VaR with some alternative downside
risk measures (discussed below).

218
 GARCH std, % Value at Risk95% (one day), %
5 10
4
3
5
2
1
0 0
1980 1990 2000 2010 1980 1990 2000 2010
S&P 500, daily data 1954:1-2013:4 The VaR is based on N()

The horizontal lines are from the

unconditional distribution

Figure 11.5: Conditional volatility and VaR

Value at Risk95% (one day) and loss, %
10
VaR S&P 500, daily data 1954:1-2013:4
max(loss,0)
The VaR is based on GARCH(1,1) & N()
8
Loss > VaR95% in 0.051 of the cases
Negative losses are shown as zero
6

0
1980 1985 1990 1995 2000 2005 2010

Figure 11.6: Backtesting VaR from a GARCH model, assuming normally distributed
shocks

11.1.1 Value at Risk of a Portfolio

If the return distribution is normal with a zero mean, then the value at risk for asset i is

VaRi D 1:64i : (11.6)

219
 Backtesting VaR from GARCH(1,1) + N(), daily S&P 500 returns
0.1

0.09

0.08
Empirical Prob(loss > VaR)

0.07

0.06

0.05

0.04

0.03

0.02

0.01 Daily S&P 500 returns, 1954:1-2013:4

0
0.9 0.92 0.94 0.96 0.98 1
VaR confidence level (q in VaRq )

Figure 11.7: Backtesting VaR from a GARCH model, assuming normally distributed
shocks
Small growth Large value
Std 8:0 5:0
VaR (95%) 12:3 8:3
ES (95%) 17:2 10:8
SemiStd 5:5 3:4
Drawdown 79:7 52:3

Table 11.1: Risk measures of monthly returns of two stock indices (%), US data 1957:1-
2012:12.

It is then straightfoward to show that the VaR for a portfortfolio

Rp D w1 R1 C w2 R2 ; (11.7)

220
 Value at Risk and Expected Shortfall
3
-ES95% -VaR95%
2.5

1.5

0.5

0
−40 0 40
Return, %

Figure 11.8: Value at risk and expected shortfall

where w1 C w2 D 1 can be written

" #" #!1=2
h i 1 12 w1 Var1
VaRp D w1 Var1 w2 Var2 ; (11.8)
12 1 w2 Var2

where 12 is the correlation of R1 and R2 . The extension to n (instead of 2) assets is

straightforward.
This expression highlights the importance of both the individual VaRi values and the
correlation. Clearly, a worst case scenario is when the portfolio is long in all assets (wi >
0) and the correlation turns out to be perfect (12 D 1).
Proof. (of (11.8)) Recall that VaRp D 1:64p , and that

p2 D w12 11 C w22 22 C 2w1 w2 12 1 2 :

Use (11.6) to substitute as i DVaRi =1:64

p2 D w12 VaR21 =1:642 C w22 VaR22 =1:642 C 2w1 w2 12  VaR1  VaR2 =1:642 :

Multiply both sides by 1:642 and take the square root to get (11.8).

221
11.1.2 Index Models for Calculating the Value at Risk

Consider a multi-index model

R D a C b1 I1 C b2 I2 C : : : C bk Ik C e; or (11.9)
D a C b 0 I C e;

where b is a k  1 vector of the bi coefficients and I is also a k  1 vector of the Ii indices.

As usual, we assume E.e/ D 0 and Cov .e; Ii / D 0. This model can be used to generate
the inputs to a VaR model. For instance, the mean and standard deviation of the return are

 D a C b0 E I
p
 D b 0 Cov.I /b C Var.e/; (11.10)

which can be used in (11.5), that is, an assumption of a normal return distribution. If the
return is of a well diversified portfolio and the indices include the key stock indices, then
the idiosyncratic risk Var.e/ is close to zero. The RiskMetrics approach is to make this
assumption.
Stand-alone VaR is a way to assess the contribution of different factors (indices). For
instance, the indices in (11.9) could include: an equity indices, interest rates, exchange
rates and perhaps also a few commodity indices. Then, an equity VaR is calculated by
setting all elements in b, except those for the equity indices, to zero. Often, the intercept,
a, is also set to zero. Similarly, an interest rate VaR is calculated by setting all elements
in b, except referring to the interest rates, to zero. And so forth for an FX VaR and a
commodity VaR. Clearly, these different VaRs do not add up to the total VaR, but they still
give an indication of where the main risk comes from.
If an asset or a portfolio is a non-linear function of the indices, then (11.9) can be
thought of as a first-order Taylor approximation where bi represents the partial derivative
of the asset return with respect to index i . For instance, an option is a non-linear function
of the underlying asset value and its volatility (as well as the time to expiration and the
interest rate). This approach, when combined with the normal assumption in (11.5), is
called the delta-normal method.

222
11.2 Expected Shortfall

The expected shortfall (also called conditional VaR, average value at risk and expected
tail loss) is the expected loss when the return actually is below the VaR˛ , that is,

ES˛ D E.RjR  VaR˛ /: (11.11)

This might be more informative than the VaR˛ , which is the minimum loss that will happen
with a 1 ˛ probability.
For a normally distributed return R  N.;  2 / we have

.c1 ˛ /
ES˛ D C ; (11.12)
1 ˛
where ./ is the pdf or a N.0; 1/ variable and where c1 ˛ is the 1 ˛ quantile of a N(0,1)
distribution (for instance, 1:64 for 1 ˛ D 0:05).
Proof. (of (11.12)) If x  N.;  2 /, then E.xjx  b/ D  .b0 /=˚.b0 / where
b0 D .b /= and where ./ and ˚./ are the pdf and cdf of a N.0; 1/ variable
respectively. To apply this, use b D VaR˛ so b0 D c1 ˛ . Clearly, ˚.c1 ˛ / D 1 ˛ (by
definition of the 1 ˛ quantile). Multiply by 1.

Example 11.6 (ES) If  D 8% and  D 16%, the 95% expected shortfall is ES95% D
0:08 C 0:16. 1:64/=0:05  0:25 and the 97.5% expected shortfall is ES97:5% D
0:08 C 0:16. 1:96/=0:025  0:29.

Notice that the expected shortfall for a normally distributed return (11.12) is a strictly
increasing function of the standard deviation (and the variance). Minimizing the expected
shortfall at a given mean return therefore gives the same solution (portfolio weights) as
minimizing the variance at the same given mean return. In other cases, the portfolio
choice will be different (and perhaps complicated to perform).

11.3 Target Semivariance (Lower Partial 2nd Moment) and Max Draw-
down

Reference: Bawa and Lindenberg (1977) and Nantell and Price (1979)

223
 Probability density function (pdf) Contribution to variance
3 2
N (µ, σ ) pdf(x)(x − µ)2
µ = 0.08 0.04 Var(x) = area
2
σ = 0.16

0.02
1

0 0
−50 0 50 −50 0 50
x, % x, %

Contribution to target semivariance Target semivariance as function of σ2

0.03
pdf(x) min(x − 0.02, 0)2 Target level -0.01
0.04 target semivariance (x) = area Target level 0.02
0.02 Target level 0.05

0.02
0.01
N (0.08, σ2 )
0 0
−50 0 50 0 0.02 0.04 0.06
x, % σ2

Figure 11.9: Target semivariance as a function of mean and standard deviation for a
N(, 2 ) variable

Using the variance (or standard deviation) as a measure of portfolio risk (as a mean-
variance investor does) fails to distinguish between the downside and upside. As an alter-
native, one could consider using a target semivariance (lower partial 2nd moment) instead.
It is defined as
p .h/ D EŒmin.Rp h; 0/2 ; (11.13)

where h is a “target level” chosen by the investor. In the subsequent analysis it will be set
Rh
equal to the riskfree rate. (It can clearly also be written p .h/ D 1 .Rp h/2 f .Rp /dRp ,
where f ./ is the pdf of the portfolio return.)
In comparison with a variance

p2 D E.Rp p /2 ; (11.14)

224
the target semivariance differs on two accounts: (i) it uses the target level h as a reference
point instead of the mean p : and (ii) only negative deviations from the reference point
are given any weight. See Figure 11.9 for an illustration (based on a normally distributed
variable).

Std and mean

15 The markers for target semivariance (sv) indicate the std
of the portfolio that minimizes the target semivariance
at the given mean return

10
MV (risky)
Mean, %

MV (risky&riskfree)
target sv (risky)
target sv (risky&riskfree)
5

0
0 5 10 15
Std, %

Figure 11.10: Standard deviation and expected returns

max drawdown
price

time

Figure 11.11: Max drawdown

For a normally distributed variable, the target semivariance p .h/ is increasing in the
standard deviation (for a given mean)—see Remark 11.7. See also Figure 11.9 for an
illustration. This means that minimizing p .h/ at a given mean return gives the same
solution (portfolio weights) as minimizing p (or p2 ) at the same given mean return. As

225
 Level of return index
400
Small growth stocks
Large value stocks
300

200

100

0
1960 1970 1980 1990 2000 2010

Drawdown compared to earlier peak (in 5-year window), %

0

−20

−40

−60

−80
1960 1970 1980 1990 2000 2010

Figure 11.12: Drawdown

a result, with normally distributed returns, an investor who wants to minimize the lower
partial 2nd moment (at a given mean return) is behaving just like a mean-variance investor.
In other cases, the portfolio choice will be different (and perhaps complicated to perform).
See Figure 11.10 for an illustration.
An alternative measure is the (percentage) maximum drawdown over a given horizon,
for instance, 5 years, say. This is the largest loss from peak to bottom within the given
horizon–see Figure 11.11. This is a useful measure when the investor do not know exactly
when he/she has to exit the investment—since it indicates the worst (peak to bottom)
outcome over the sample.
See Figures 11.12–11.13 for an illustration of max drawdown.

226
 Level of return index
5
MSCI world
CT hedge funds
4 Global govt bonds

0
1995 2000 2005 2010

Drawdown compared to earlier peak (since start of sample), %

0

−20

−40

−60
1995 2000 2005 2010

Figure 11.13: Drawdown

Remark 11.7 (Target semivariance calculation for normally distributed variable ) For
an N.;  2 / variable, target semivariance around the target level h is

p .h/ D  2 a.a/ C  2 .a2 C 1/˚.a/, where a D .h /=;

where ./ and ˚./ are the pdf and cdf of a N.0; 1/ variable respectively. Notice that
p .h/ D  2 =2 for h D . See Figure 11.9 for a numerical illustration. It is straightfor-
ward (but a bit tedious) to show that
@p .h/
D 2˚.a/;
@
so the target semivariance is a strictly increasing function of the standard deviation.

227
 See Table 11.2 for an empirical comparison of the different risk measures.

Std VaR (95%) ES (95%) SemiStd Drawdown

Std 1:00 0:94 0:98 0:97 0:68
VaR (95%) 0:94 1:00 0:94 0:95 0:72
ES (95%) 0:98 0:94 1:00 0:98 0:67
SemiStd 0:97 0:95 0:98 1:00 0:68
Drawdown 0:68 0:72 0:67 0:68 1:00

Table 11.2: Correlation of rank of risk measures across the 25 FF portfolios (%), US data
1957:1-2012:12.

Bibliography
Alexander, C., 2008, Market Risk Analysis: Value at Risk Models, Wiley.

Bawa, V. S., and E. B. Lindenberg, 1977, “Capital market equilibrium in a mean-lower

partial moment framework,” Journal of Financial Economics, 5, 189–200.

Fabozzi, F. J., S. M. Focardi, and P. N. Kolm, 2006, Financial modeling of the equity
market, Wiley Finance.

Hull, J. C., 2006, Options, futures, and other derivatives, Prentice-Hall, Upper Saddle
River, NJ, 6th edn.

McDonald, R. L., 2006, Derivatives markets, Addison-Wesley, 2nd edn.

McNeil, A. J., R. Frey, and P. Embrechts, 2005, Quantitative risk management, Princeton
University Press.

Nantell, T. J., and B. Price, 1979, “An analytical comparison of variance and semivariance
capital market theories,” Journal of Financial and Quantitative Analysis, 14, 221–242.

228
12 Return Distributions (Univariate)
Sections denoted by a star ( ) is not required reading.

12.1 Estimating and Testing Distributions

Reference: Harvey (1989) 260, Davidson and MacKinnon (1993) 267, Silverman (1986);
Mittelhammer (1996), DeGroot (1986)

12.1.1 A Quick Recap of a Univariate Distribution

The cdf (cumulative distribution function) measures the probability that the random vari-
able Xi is below or at some numerical value xi ,

ui D Fi .xi / D Pr.Xi  xi /: (12.1)

For instance, with an N.0; 1/ distribution, F . 1:64/ D 0:05. Clearly, the cdf values
are between (and including) 0 and 1. The distribution of Xi is often called the marginal
distribution of Xi —to distinguish it from the joint distribution of Xi and Xj . (See below
for more information on joint distributions.)
The pdf (probability density function) fi .xi / is the “height” of the distribution in the
sense that the cdf F .xi / is the integral of the pdf from minus infinity to xi
Z xi
Fi .xi / D fi .s/ds: (12.2)
sD 1

(Conversely, the pdf is the derivative of the cdf, fi .xi / D @Fi .xi /=@xi .) The Gaussian
pdf (the normal distribution) is bell shaped.
Remark 12.1 (Quantile of a distribution) The ˛ quantile of a distribution (˛ ) is the value
of x such that there is a probability of ˛ of a lower value. We can solve for the quantile by
inverting the cdf, ˛ D F .˛ / as ˛ D F 1 .˛/. For instance, the 5% quantile of a N.0; 1/
distribution is 1:64 D ˚ 1 .0:05/, where ˚ 1 ./ denotes the inverse of an N.0; 1/ cdf.
See Figure 12.1 for an illustration.

229
 Density of N(0,1) Density of N(8, 162 )
5% quantile is c = −1.64 3 5% quantile is µ + cσ = −18
0.4
0.3 2
pdf

pdf
0.2
1
0.1
0 0
−3 c 0 3 −40 0 40
x R

cdf of N(8, 162 ) Inverse of cdf of N(8, 162 )

1
40
cdf

0.5 0

−40
0
−40 0 40 0 0.5 1
R cdf

Figure 12.1: Finding quantiles of a N(, 2 ) distribution

12.1.2 QQ Plots

Are returns normally distributed? Mostly not, but it depends on the asset type and on the
data frequency. Options returns typically have very non-normal distributions (in partic-
ular, since the return is 100% on many expiration days). Stock returns are typically
distinctly non-linear at short horizons, but can look somewhat normal at longer horizons.
To assess the normality of returns, the usual econometric techniques (Bera–Jarque
and Kolmogorov-Smirnov tests) are useful, but a visual inspection of the histogram and a
QQ-plot also give useful clues. See Figures 12.2–12.4 for illustrations.

Remark 12.2 (Reading a QQ plot) A QQ plot is a way to assess if the empirical distri-
bution conforms reasonably well to a prespecified theoretical distribution, for instance,
a normal distribution where the mean and variance have been estimated from the data.
Each point in the QQ plot shows a specific percentile (quantile) according to the empiri-

230
cal as well as according to the theoretical distribution. For instance, if the 2th percentile
(0.02 percentile) is at -10 in the empirical distribution, but at only -3 in the theoretical
distribution, then this indicates that the two distributions have fairly different left tails.

There is one caveat to this way of studying data: it only provides evidence on the
unconditional distribution. For instance, nothing rules out the possibility that we could
estimate a model for time-varying volatility (for instance, a GARCH model) of the returns
and thus generate a description for how the VaR changes over time. However, data with
time varying volatility will typically not have an unconditional normal distribution.

Daily returns Daily returns, zoomed vertically

25
8000
Number of days

Number of days
20
6000
15
4000 10
2000 5
0 0
−20 −10 0 10 −20 −10 0 10
Daily excess return, % Daily excess return, %

Daily returns, zoomed horizontally

8000 Daily S&P 500 returns, 1957:1-2013:5
Number of days

Solid line: estimated normal distribution

6000
4000
2000
0
−2 0 2
Daily excess return, %

Figure 12.2: Distribution of daily S&P returns

231
 QQ plot of daily S&P 500 returns
6
0.1st to 99.9th percentiles

2
Empirical quantiles

−2

−4

Daily S&P 500 returns, 1957:1-2013:5

−6

−6 −4 −2 0 2 4 6
Quantiles from estimated N(µ, σ2 ), %

Figure 12.3: Quantiles of daily S&P returns

12.1.3 Parametric Tests of Normal Distribution

The skewness, kurtosis and Bera-Jarque test for normality are useful diagnostic tools.
They are

Test statistic Distribution

3
skewness D T1 TtD1 xt 
P
N .0; 6=T /

4 (12.3)
kurtosis D T1 TtD1 xt 
P
N .3; 24=T /
Bera-Jarque D T6 skewness2 C 24
T
.kurtosis 3/2 22 :

This is implemented by using the estimated mean and standard deviation. The distribu-
tions stated on the right hand side of (12.3) are under the null hypothesis that x t is iid
N ;  2 . The “excess kurtosis” is defined as the kurtosis minus 3.

The intuition for the 22 distribution of the Bera-Jarque test is that both the skewness
and kurtosis are, if properly scaled, N.0; 1/ variables. It can also be shown that they,
under the null hypothesis, are uncorrelated. The Bera-Jarque test statistic is therefore a

232
 QQ plot of daily returns QQ plot of weekly returns
10
Empirical quantiles

Empirical quantiles
5
5
0 0

−5
−5
−10
−5 0 5 −10 −5 0 5 10
Quantiles from N(µ, σ2 ), % Quantiles from N(µ, σ2 ), %

QQ plot of monthly returns Circles denote 0.1th to 99.9th percentiles

Empirical quantiles

10 Daily S&P 500 returns, 1957:1-2013:5

−10

−20
−20 −10 0 10
Quantiles from N(µ, σ2 ), %

Figure 12.4: Distribution of S&P returns (different horizons)

sum of the square of two uncorrelated N.0; 1/ variables, which has a 22 distribution.

12.1.4 Nonparametric Tests of General Distributions

The Kolmogorov-Smirnov test is designed to test if an empirical distribution function,

EDF.x/, conforms with a theoretical cdf, F .x/. The empirical distribution function is
defined as the fraction of observations which are less or equal to x, that is,
T
1X
EDF .x/ D ı.x t  x/; where (12.4)
T t D1
(
1 if q is true
ı.q/ D
0 else.

233
 Empirical distribution function and theoretical cdf

1 When the plot of EDF takes a jump up,

the correct EDF value is the higher one

0.8

0.6

0.4
data
0.2 cdf of N(4,1)
EDF
0

3 3.5 4 4.5 5 5.5 6

x

Figure 12.5: Example of empirical distribution function

The EDF.x t / and F .x t / are often plotted against the sorted (in ascending order) sample
fx t gTtD1 .
See Figure 12.5 for an illustration.

Example 12.3 (EDF) Suppose we have a sample with three data points: Œx1 ; x2 ; x3  D
Œ5; 3:5; 4. The empirical distribution function is then as in Figure 12.5.

Define the absolute value of the maximum distance

DT D max jEDF .x t / F .x t /j : (12.5)

xt

Example 12.4 (Kolmogorov-Smirnov test statistic) Figure 12.5 also shows the cumula-
tive distribution function (cdf) of a normally distributed variable. The test statistic (12.5)
is then the largest difference (in absolute terms) of the EDF and the cdf—among the ob-
served values of x t .
p
We reject the null hypothesis that EDF.x/ D F .x/ if T D t > c, where c is a critical
value which can be calculated from
p  1
2i 2 c 2
X
lim Pr T DT  c D 1 2 . 1/i 1
e : (12.6)
T !1
i D1

234
 Kolmogorov-Smirnov test

1 The
√ K-S test statistic is
T × the length of the longest arrow

0.8

0.6

0.4
data
0.2 cdf of N(4,1)
EDF
0

3 3.5 4 4.5 5 5.5 6

x

Figure 12.6: K-S test

It can be approximated by replacing 1 with a large number (for instance, 100). For
instance, c D 1:35 provides a 5% critical value. See Figure 12.7. There is a corresponding
test for comparing two empirical cdfs.
Pearson’s 2 test does the same thing as the K-S test but for a discrete distribution.
Suppose you have K categories with Ni values in category i. The theoretical distribution
predicts that the fraction pi should be in category i , with Ki D1 pi D 1. Then
P

K
X .Ni Tpi /2 2
 K 1: (12.7)
i D1
Tpi

There is a corresponding test for comparing two empirical distributions.

12.1.5 Fitting a Mixture Normal Distribution to Data

Reference: Hastie, Tibshirani, and Friedman (2001) 8.5

A normal distribution often fits returns poorly. If we need a distribution, then a mixture

235
 Cdf of K-S test Quantiles of K-S distribution
1 2

cdf 1.5
0.5

c
1

0 0.5
0.5 1 1.5 2 0 0.5 1
c cdf

Quantiles of K-S distribution (zoomed)

p-value Critical value
1.8 0.150 1.138
0.100 1.224
1.6 0.050 1.358
c

1.4 0.025 1.480

0.010 1.628
1.2
0.85 0.9 0.95 1
cdf

p
Figure 12.7: Distribution of the Kolmogorov-Smirnov test statistics, T DT

of two normals is typically much better, and still fairly simple.

The pdf of this distribution is just a weighted average of two different (bell shaped)
pdfs of normal distributions (also called mixture components)

f .x t I 1 ; 2 ; 12 ; 22 ; / D .1 /.x t I 1 ; 12 / C .x t I 2 ; 22 /; (12.8)

where .xI i ; i2 / is the pdf of a normal distribution with mean i and variance i2 . It
thus contains five parameters: the means and the variances of the two components and
their relative weight ().
See Figures 12.8–12.10 for an illustration.

Remark 12.5 (Estimation of the mixture normal pdf) With 2 mixture components, the log
likelihood is just
XT
LL D ln f .x t I 1 ; 2 ; 12 ; 22 ; /;
t D1

where f ./ is the pdf in (12.8) A numerical optimization method could be used to maximize
this likelihood function. However, this is tricky so an alternative approach is often used.
This is an iterative approach in three steps:

236
 Distribution of daily S&P500,1957:1-2013:5

normal pdf
0.6

0.5

0.4

0.3

0.2

0.1

0
−5 −4 −3 −2 −1 0 1 2 3 4 5
Daily excess return, %

Figure 12.8: Histogram of returns and a fitted normal distribution

(1) Guess values of 1 ; 2 ; 12 ; 22 and . For instance, pick 1 D x1 , 2 D x2 , 12 D
22 D Var.x t / and  D 0:5.
(2) Calculate

.x t I 2 ; 22 /
t D for t D 1; : : : ; T:
.1 /.x t I 1 ; 12 / C .x t I 2 ; 22 /

(3) Calculate (in this order)

PT PT
t D1 .1 t /x t 2 .1 t /.x t 1 /2
1 D PT , 1 D tD1PT ;
t D1 .1 t / t D1 .1 t /
PT PT
t D1 t x t t .x t 2 /2
2 D PT , 2 D t D1PT
2
, and
t D1 t tD1 t
XT
D t =T .
t D1

Iterate over (2) and (3) until the parameter values converge. (This is an example of the
EM algorithm.) Notice that the calculation of i2 uses i from the same (not the previous)

237
 Distribution of daily S&P500, 1957:1-2013:5

Mixture pdf 1
0.6 Mixture pdf 2
pdf 1 pdf 2
mean 0.03 -0.04 Total pdf
0.5 std 0.65 1.99
weight 0.84 0.16

0.4

0.3

0.2

0.1

0
−5 −4 −3 −2 −1 0 1 2 3 4 5
Daily excess return, %

Figure 12.9: Histogram of returns and a fitted mixture normal distribution

iteration.

12.1.6 Kernel Density Estimation

Reference: Silverman (1986)

A histogram is just a count of the relative number of observations that fall in (pre-
specified) non-overlapping intervals. If we also divide by the width of the interval, then
the area under the histogram is unity, so the scaled histogram can be interpreted as a den-
sity function. For instance, if the intervals (“bins”) are h wide, then the scaled histogram
at the point x (say, x D 2:3) can be defined as
T
1 X1
g.x/ D ı.x t is in bini /; where (12.9)
T t D1 h
(
1 if q is true
ı.q/ D
0 else.

Note that the area under g.x/ indeed integrates to unity.

238
 QQ plot of daily S&P 500 returns
6
0.1th to 99.9th percentiles

2
Empirical quantiles

−2

−4

Daily S&P 500 returns, 1957:1-2013:5

−6

−6 −4 −2 0 2 4 6
Quantiles from estimated mixture normal, %

Figure 12.10: Quantiles of daily S&P returns

We can gain efficiency by using a more sophisticated estimator. In particular, using a

pdf instead of the binary function is often both convenient and more efficient. To develop
that method, we first show an alternative way of constructing a histogram. First, let a bin
be defined as symmetric interval around a point x: x h=2 to x C h=2. (We can vary the
value of x to define other bins.) Second, notice that the histogram value at point x can be
written
T
1 X 1 ˇˇ x t x ˇˇ 
g.x/ D ı ˇ ˇ  1=2 : (12.10)
T t D1 h h

In fact, h1 ı.jx t xj  h=2/ is the pdf value of a uniformly distributed variable (over the
interval x h=2 to x Ch=2). This shows that our estimate of the pdf (here: the histogram)
can be thought of as a average of hypothetical pdf values of the data in the neighbourhood
of x. However, we can gain efficiency and get a smoother (across x values) estimate by
using another density function that the uniform. In particular, using a density function
that tapers off continuously instead of suddenly dropping to zero (as the uniform density

239
 Weights for histogram (at x = 4) Weights for kernel density (at x = 4)
2 2
data data
weights weights
1.5 1.5

1 1
0.5 Bin: 4 ± 0.25 0.5

0 0
3 4 5 6 3 4 5 6
x x

The estimate (at x = 4) equals

the average of the weights

Figure 12.11: Calculation of the pdf at x D 4

does) improves the properties. In fact, the N.0; h2 / is often used. The kernel density
estimator of the pdf at some point x is then
 2 
1 XT 1 1 x x
fO .x/ D
t
p exp (12.11)
T t D1 h 2 2 h

Notice that the function in the summation is the density function of a N.x; h2 / distribu-
tion.
The value h D Std.x t /1:06T 1=5 is sometimes recommended, since it can be shown
to be the optimal choice (in MSE sense) if data is normally distributed. The bandwidth h
could also be chosen by a leave-one-out cross-validation technique.
See Figure 12.12 for an example and Figure 12.13 for a QQ plot which is a good way
to visualize the difference between the empirical and a given theoretical distribution.
It can be shown that (with iid data and a Gaussian kernel) the asymptotic distribution
is
p h i 
1

O O
T h f .x/ E f .x/ ! N 0; p f .x/ ;
d
(12.12)
2 
The easiest way to handle a bounded support of x is to transform the variable into one
with an unbounded support, estimate the pdf for this variable, and then use the “change
of variable” technique to transform to the pdf of the original variable.
We can also estimate multivariate pdfs. Let x t be a d 1 matrix and ˝O be the estimated
covariance matrix of x t . We can then estimate the pdf at a point x by using a multivariate

240
 Histogram (scaled: area=1) Kernel density estimation
0.2 0.2
Small h (optimal)
0.15 0.15 Large h

0.1 0.1

0.05 0.05

0 0
0 5 10 15 20 0 5 10 15 20
Interest rate Interest rate

Daily federal funds rates√1954:7-2013:4

K-S (against N(µ, σ2 )) : T D = 14.1
Skewness: 1.1
kurtosis: 4.9
Bera-Jarque: 7244.7

Figure 12.12: Federal funds rate

Gaussian kernel as
 
1 XT 1 1
fO .x/ D exp .x t 0 O
x/ .H ˝/ 2
.x t 1
x/ : (12.13)
T O 1=2
t D1 .2/d=2 jH 2 ˝j 2

Notice that the function in the summation is the (multivariate) density function of a
O distribution. The value H D 1:06T 1=.d C4/ is sometimes recommended.
N.x; H 2 ˝/

Remark 12.6 ((12.13) with d D 1) With just one variable, (12.13) becomes
"  2 #
1 XT 1 1 x x
fO .x/ D
t
p exp ;
T t D1 H Std.x / 2
t 2 H Std.x t/

which is the same as (12.11) if h D H Std.x t /.

12.1.7 “Foundations of Technical Analysis...” by Lo, Mamaysky and Wang (2000)

Reference: Lo, Mamaysky, and Wang (2000)

Topic: is the distribution of the return different after a “signal” (TA). This paper uses
kernel regressions to identify and implement some technical trading rules, and then tests
if the distribution (of the return) after a signal is the same as the unconditional distribution

241
 QQ plot of daily federal funds rates
20 0.1th to 99.9th percentiles

15
Empirical quantiles

10

Daily federal funds rates, 1954:7-2013:4

−5
−5 0 5 10 15 20
Quantiles from estimated normal distribution, %

Figure 12.13: Federal funds rate

(using Pearson’s 2 test and the Kolmogorov-Smirnov test). They reject that hypothesis
in many cases, using daily data (1962–1996) for around 50 (randomly selected) stocks.
See Figures 12.14–12.15 for an illustration.

12.2 Tail Distribution

Reference: McNeil, Frey, and Embrechts (2005) 7, Alexander (2008) 3

In risk control, the focus is the distribution of losses beyond some threshold level.
This has three direct implications. First, the object under study is the loss

XD R; (12.14)

that is, the negative of the return. Second, the attention is on how the distribution looks
like beyond a threshold and also on the the probability of exceeding this threshold. In con-
trast, the exact shape of the distribution below that point is typically disregarded. Third,

242
 Inverted MA rule, S&P 500
1350
MA(3) and MA(25), bandwidth 0.01
Circles at the bottom (top) margin indicates buys (sells)
1300

1250

1200
Long MA (-)
Long MA (+)
Short MA
1150
Jan Feb Mar Apr
1999

Figure 12.14: Examples of trading rules

modelling the tail of the distribution is best done by using a distribution that allows for a
much heavier tail that suggested by a normal distribution. The generalized Pareto (GP)
distribution is often used.

12.2.1 Loss Distribution and the Generalized Pareto Distribution

The generalized Pareto (GP) distribution is often used to model the tail of the loss distri-
bution. See Figure 12.16 for an illustration.

Remark 12.7 (Cdf and pdf of the generalized Pareto distribution) The generalized Pareto
distribution is described by a scale parameter (ˇ > 0) and a shape parameter (). The
cdf (Pr.Z  z/, where Z is the random variable and z is a value) is
(
1 .1 C z=ˇ/ 1= if  ¤ 0
G.z/ D
1 exp. z=ˇ/  D 0;

for 0  z if   0 and z  ˇ= in case  < 0. The pdf is therefore

(
ˇ
1
.1 C z=ˇ/ 1= 1 if  ¤ 0
g.z/ D
ˇ
1
exp. z=ˇ/  D 0:

243
 Distribution of returns for all days Inverted MA rule: after buy signal
Mean Std Mean Std
0.6 0.03 1.17 0.6 0.05 1.72

0.4 0.4

0.2 0.2

0 0
−2 0 2 −2 0 2
Return Return
Daily S&P 500 data 1990:1-2013:4

Inverted MA rule: after neutral signal Inverted MA rule: after sell signal
Mean Std Mean Std
0.6 0.05 0.94 0.6 0.01 0.90

0.4 0.4

0.2 0.2

0 0
−2 0 2 −2 0 2
Return Return

Figure 12.15: Examples of trading rules

unknown shape

90% probability mass (Pu D 0:9) generalized Pareto dist

u Loss

Figure 12.16: Loss distribution

The mean is defined (finite) if  < 1 and is then E.z/ D ˇ=.1 /. Similarly, the variance
is finite if  < 1=2 and is then Var.z/ D ˇ 2 =Œ.1 /2 .1 2/. See Figure 12.17 for an
illustration.

244
 Pdf of generalized Pareto distribution (β = 0.15)
7
ξ =0
6 ξ = 0.25
ξ = 0.45
5

0
0 0.1 0.2 0.3 0.4 0.5
Outcome of random variable

Figure 12.17: Generalized Pareto distributions

Consider the loss X (the negative of the return) and let u be a threshold. Assume
that the threshold exceedance (X u) has a generalized Pareto distribution. Let Pu be
probability of the loss being smaller than the threshold, X  u. Then, the cdf of the loss
for values greater than the threshold (Pr.X  x/ for x > u) can be written

Pr.X  x/ D F .x/ D Pu C G.x u/.1 Pu /, for x > u; (12.15)

where G.z/ is the cdf of the generalized Pareto distribution. Noticed that, the cdf value is
Pu at at x D u (or just slightly above u), and that it becomes one as x goes to infinity.
Clearly, the pdf is

f .x/ D g.x u/.1 Pu /, for x > u; (12.16)

where g.z/ is the pdf of the generalized Pareto distribution. Notice that integrating the
pdf from x D u to infinity shows that the probability mass of X above u is 1 Pu . Since
the probability mass below u is Pu , it adds up to unity (as it should). See Figures 12.16
and 12.19 for illustrations.

Remark 12.8 (Random number from a generalized Pareto distribution ) By inverting the
cdf of the generalized Pareto distribution, we can notice that if u is uniformly distributed

245
on .0; 1, then we can construct random variables with a GP distribution by

z D ˇ Œ.1 u/ 
1 if  ¤ 0
zD ln.1 u/ˇ  D 0:

In addition, if z is the threshold exceedance (z D X u/, then the loss (conditional on

being above the threshold u) can be generated as X D u C z.

It is often useful to calculate the tail probability Pr.X > x/, which in the case of the
cdf in (12.15) is

Pr.X > x/ D 1 F .x/ D .1 Pu /Œ1 G.x u/; (12.17)

where G.z/ is the cdf of the generalized Pareto distribution.

Loss distributions for loss > 12, Pr(loss > 12) = 10%

1 N(0.08, 0.162 )
generalized Pareto (ξ = 0.22, β = 0.16)
0.8
VaR(95%) ES(95%)
0.6
Normal dist 18.2 25.3
GP dist 24.5 48.4
0.4

0.2

0
15 20 25 30 35 40 45 50 55 60
Loss, %

Figure 12.18: Comparison of a normal and a generalized Pareto distribution for the tail of
losses

12.2.2 VaR and Expected Shortfall of a GP Distribution

The value at risk, VaR˛ (say, ˛ D 0:95), is the ˛-th quantile of the loss distribution

VaR˛ D cdfX 1 .˛/; (12.18)

246
where cdfX 1 ./ is the inverse cumulative distribution function of the losses, so cdfX 1 .˛/
is the ˛ quantile of the loss distribution. For instance, VaR95% is the 0:95 quantile of the
loss distribution. This clearly means that the probability of the loss to be less than VaR˛
equals ˛
Pr.X  VaR˛ / D ˛: (12.19)

(Equivalently, the Pr.X >VaR˛ / D 1 ˛:)

Assuming ˛ is higher than Pu (so VaR˛  u), the cdf (12.15) together with the form
of the generalized Pareto distribution give the VaR
8    
ˇ
< uC 1 ˛
1 if  ¤ 0
ˆ
 1 Pu
VaR˛ D   , for ˛  Pu : (12.20)
u ˇ ln 1 Pu 1 ˛
D0
ˆ
:

Proof. (of (12.20)) Set F .x/ D ˛ in (12.15) and use z D x u in the cdf from
Remark 12.7 and solve for x.
If we assume  < 1 (to make sure that the mean is finite), then straightforward inte-
gration using (12.16) shows that the expected shortfall is

ES˛ D E.XjX  VaR˛ /

VaRa ˇ u
D C , for ˛ > Pu and  < 1: (12.21)
1  1 

12.2.3 Expected Exceedance of a GP Distribution

To locate the cut-off level where the tail “starts,” that is, to choose the value of u. It often
helps to study the expected exceedance.
The average exceedance (in data) over some threshold level  is the mean of X t 
for those observations where X t > 
PT
.X t /ı.X t > /
O
e./ D t D1PT ; where (12.22)
t D1 .X t > /
(
1 if q is true
ı.q/ D
0 else.

The expected exceedance of a GD distribution is easily found by letting  DVaR˛

in the expected shortfall (12.21) and then subtract  from both sides to get the expected

247
exceedance of the loss over another threshold  > u

e./ D E .X jX > /

 ˇ u
D C , for  > u and  < 1. (12.23)
1  1 

The expected exceedance of a generalized Pareto distribution (with  > 0) is increasing

with the threshold level . This indicates that the tail of the distribution is very long.
In contrast, a normal distribution would typically show a negative relation (see Figure
12.19for an illustration). This provides a way of assessing which distribution that best
fits the tail of the historical histogram. In addition, if we have decided to use the GP
distribution for the tail, but does not know where the tail starts (the value of u), then it can
be chosen as the lowest value (of ) after which the average exceedance in data (12.22)
appears to be a linear function of the threshold.

Remark 12.9 (Expected exceedance from a normal distribution) If X  N.;  2 /, then

.0 /
E.X jX > / D  C  ;
1 ˚.0 /
with 0 D . /=

where ./ and ˚ are the pdf and cdf of a N.0; 1/ variable respectively.

12.2.4 Estimating a GP Distribution

The estimation of the parameters of the distribution ( and ˇ) is typically done by maxi-
mum likelihood. Alternatively, a comparison of the empirical exceedance (12.22) with the
theoretical (12.23) can help. Suppose we calculate the empirical exceedance for different
values of the threshold level (denoted i —all large enough so the relation looks linear),
then we can estimate (by LS)

O i / D a C bi C "i :
e. (12.24)

248
Then, the theoretical exceedance (12.23) for a given starting point of the GP distribution
(u) is related to this regression according to

ˇ u 
aD and b D , or
1  1 
b
D and ˇ D a.1 / C u: (12.25)
1Cb
See Figure 12.20 for an illustration.

Expected exeedance (loss minus threshold, v)

30

25

20

N(0.08, 0.162 )
15
generalized Pareto (ξ = 0.22, β = 0.16, u = 12)

10

0
15 20 25 30 35 40
Threshold v, %

Figure 12.19: Expected exceedance, normal and generalized Pareto distribution

Remark 12.10 (Log likelihood function of the loss distribution) Since we have assumed
that the threshold exceedance (X u) has a generalized Pareto distribution, Remark 12.7
shows that the log likelihood for the observation of the loss above the threshold (X t > u)
is
X
LD Lt
t st. X t >u
(
ln ˇ .1= C 1/ ln Œ1 C  .X t u/ =ˇ if  ¤ 0
ln L t D
ln ˇ .X t u/ =ˇ  D 0:

249
This allows us to estimate  and ˇ by maximum likelihood. Typically, u is not estimated,
but imposed a priori (based on the expected exceedance).

Expected exceedance Estimated loss distribution (pdf)

Loss minus threshold, v

(50th to 99th percentiles) u = 1.3, Pr(loss > u) = 6.8%

1.2 u = 1.3, ξ = 0.27, β = 0.53 0.1
ξ = 0.23, β = 0.58
1
0.05
0.8

0.6
0
0 1 2 1.5 2 2.5 3 3.5 4
Threshold v, % Loss, %

QQ plot
Daily S&P 500 returns, 1957:1-2012:12
Empirical quantiles

(94th to 99th percentiles)

2.5

1.5
1.5 2 2.5
Quantiles from estimated GPD, %

Figure 12.20: Results from S&P 500 data

Example 12.11 (Estimation of the generalized Pareto distribution on S&P daily returns).
Figure 12.20 (upper left panel) shows that it may be reasonable to fit a GP distribution
with a threshold u D 1:3. The upper right panel illustrates the estimated distribution,
while the lower left panel shows that the highest quantiles are well captured by estimated
distribution.

Bibliography
Alexander, C., 2008, Market Risk Analysis: Value at Risk Models, Wiley.

250
Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,
Oxford University Press, Oxford.

DeGroot, M. H., 1986, Probability and statistics, Addison-Wesley, Reading, Mas-

sachusetts.

Harvey, A. C., 1989, Forecasting, structural time series models and the Kalman filter,
Cambridge University Press.

Hastie, T., R. Tibshirani, and J. Friedman, 2001, The elements of statistical learning: data
mining, inference and prediction, Springer Verlag.

Lo, A. W., H. Mamaysky, and J. Wang, 2000, “Foundations of technical analysis: com-
putational algorithms, statistical inference, and empirical implementation,” Journal of
Finance, 55, 1705–1765.

McNeil, A. J., R. Frey, and P. Embrechts, 2005, Quantitative risk management, Princeton
University Press.

Mittelhammer, R. C., 1996, Mathematical statistics for economics and business, Springer-
Verlag, New York.

Silverman, B. W., 1986, Density estimation for statistics and data analysis, Chapman and
Hall, London.

251
13 Return Distributions (Multivariate)
More advanced material is denoted by a star ( ). It is not required reading.

13.1 Recap of Univariate Distributions

The cdf (cumulative distribution function) measures the probability that the random vari-
able Xi is below or at some numerical value xi ,

ui D Fi .xi / D Pr.Xi  xi /: (13.1)

For instance, with an N.0; 1/ distribution, F . 1:64/ D 0:05. Clearly, the cdf values
are between (and including) 0 and 1. The distribution of Xi is often called the marginal
distribution of Xi —to distinguish it from the joint distribution of Xi and Xj . (See below
for more information on joint distributions.)
The pdf (probability density function) fi .xi / is the “height” of the distribution in the
sense that the cdf F .xi / is the integral of the pdf from minus infinity to xi
Z xi
Fi .xi / D fi .s/ds: (13.2)
sD 1

(Conversely, the pdf is the derivative of the cdf, fi .xi / D @Fi .xi /=@xi .) The Gaussian
pdf (the normal distribution) is bell shaped.

Remark 13.1 (Quantile of a distribution) The ˛ quantile of a distribution (˛ ) is the value
of x such that there is a probability of ˛ of a lower value. We can solve for the quantile by
inverting the cdf, ˛ D F .˛ / as ˛ D F 1 .˛/. For instance, the 5% quantile of a N.0; 1/
distribution is 1:64 D ˚ 1 .0:05/, where ˚ 1 ./ denotes the inverse of an N.0; 1/ cdf.
See Figure 13.1 for an illustration.

13.2 Exceedance Correlations

Reference: Ang and Chen (2002)

252
 Density of N(0,1) Density of N(8, 162 )
5% quantile is c = −1.64 3 5% quantile is µ + cσ = −18
0.4
0.3 2
pdf

pdf
0.2
1
0.1
0 0
−3 c 0 3 −40 0 40
x R

cdf of N(8, 162 ) Inverse of cdf of N(8, 162 )

1
40
cdf

0.5 0

−40
0
−40 0 40 0 0.5 1
R cdf

Figure 13.1: Finding quantiles of a N(, 2 ) distribution

It is often argued that most assets are more strongly correlated in down markets than
in up markets. If so, diversification may not be such a powerful tool as what we would
otherwise believe.
A straightforward way of examining this is to calculate the correlation of two returns(x
and y, say) for specific intervals. For instance, we could specify that x t should be between
h1 and h2 and y t between k1 and k2

Corr.x t ; y t jh1 < x t  h2 ; k1 < y t  k2 /: (13.3)

For instance, by setting the lower boundaries (h1 and k1 ) to 1 and the upper boundaries
(h2 and k2 ) to 0, we get the correlation in down markets.
A (bivariate) normal distribution has very little probability mass at low returns, which
leads to the correlation being squeezed towards zero as we only consider data far out in

253
the tail. In short, the tail correlation of a normal distribution is always closer to zero than
the correlation for all data points. This is illustrated in Figure 13.2.
In contrast, Figures 13.3–13.4 suggest (for two US portfolios) that the correlation in
the lower tail is almost as high as for all the data and considerably higher than for the
upper tail. This suggests that the relation between the two returns in the tails is not well
described by a normal distribution. In particular, we need to use a distribution that allows
for much stronger dependence in the lower tail. Otherwise, the diversification benefits (in
down markets) are likely to be exaggerated.

Correlation in lower tail, bivariate N(0,1) distribution

0.7 ρ = 0.75
ρ = 0.5
0.6 ρ = 0.25

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5
Upper boundary (prob of lower value)

Figure 13.2: Correlation in lower tail when data is drawn from a normal distribution with
correlation 

13.3 Beyond (Linear) Correlations

Reference: Alexander (2008) 6, McNeil, Frey, and Embrechts (2005)

The standard correlation (also called Pearson’s correlation) measures the linear rela-
tion between two variables, that is, to what extent one variable can be explained by a
linear function of the other variable (and a constant). That is adequate for most issues
in finance, but we sometimes need to go beyond the correlation—to capture non-linear

254
 Extreme returns of two portfolios

Daily US data 1979:1-2012:12

10
Lines mark 5th and 95th percentiles

5
Returns of large stocks, %

−5

−10
Corr freq
All 0.69 1.00

−15 Low 0.76 0.02

Mid 0.52 0.84
High 0.50 0.02
−20
−20 −15 −10 −5 0 5 10
Returns of small stocks, %

Figure 13.3: Correlation of two portfolios

Lower tail correlation Upper tail correlation

0.8 0.8

0.7 0.7

0.6 0.6

0.5 small stocks and large stocks 0.5

Daily US data 1979:1-2012:12
0.4 0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Upper boundary (prob of lower value) Lower boundary (prob of lower value)

Figure 13.4: Correlation in the tails for two portfolios

255
relations. It also turns out to be easier to calibrate/estimate copulas (see below) by using
other measures of dependency.
Spearman’s rank correlation (called Spearman’s rho) of two variables measures to
what degree their relation is monotonic: it is the correlation of their respective ranks. It
measures if one variable tends to be high when the other also is—without imposing the
restriction that this relation must be linear.

Corr = 0.90 Corr = 0.03

2 2
1 1
0 0
y

y
−1 −1
−2 ρ = 0.88, τ = 0.69 −2 ρ = 0.03, τ = 0.01

−5 0 5 −5 0 5
x x

Corr = −0.88 Corr = 0.49

2 2
1 1
0 0
y

−1 −1
−2 ρ = −0.84, τ = −0.65 −2 ρ = 1.00, τ = 1.00

−5 0 5 −5 0 5
x x

Figure 13.5: Illustration of correlation and rank correlation

It is computed in two steps. First, the data is ranked from the smallest (rank 1) to
the largest (ranked T , where T is the sample size). Ties (when two or more observations
have the same values) are handled by averaging the ranks. The following illustrates this

256
for two variables
x t rank.x t / y t rank.y t /
2 2:5 7 2
10 4 8 3 (13.4)
3 1 2 1
2 2:5 10 4
In the second step, simply estimate the correlation of the ranks of two variables

Spearman’s  D CorrŒrank.x t /; rank.y t /: (13.5)

Clearly, this correlation is between 1 and 1. (There is an alternative way of calculating

the rank correlation based on the difference of the ranks, d t Drank.x t / rank.y t /,  D
1 6˙ tTD1 d t2 =.T 3 T /. It gives the same result if there are no tied ranks.) See Figure
13.5 for an illustration.
The rank correlation can be tested by using the fact that under the null hypothesis the
rank correlation is zero. We then get
p
T 1O !d N.0; 1/: (13.6)

(For samples of 20 to 40 observations, it is often recommended to use

p
.T 2/=.1 O2 /O
which has an tT 2 distribution.)

Remark 13.2 (Spearman’s  for a distribution ) If we have specified the joint distribu-
tion of the random variables X and Y , then we can also calculate the implied Spearman’s
 (sometimes only numerically) as CorrŒFX .X/; FY .Y / where FX .X/ is the cdf of X and
FY .Y / of Y .

Kendall’s rank correlation (called Kendall’s ) is similar, but is based on compar-

ing changes of x t (compared to x1 ; : : : x t 1 ) with the corresponding changes of y t . For
instance, with three data points (.x1 ; y1 /; .x2 ; y2 /; .x3 ; y3 /) we first calculate

Changes of x Changes of y
x2 x1 y2 y1
(13.7)
x3 x1 y3 y1
x3 x2 y3 y2 ;

257
which gives T .T 1/=2 (here 3) pairs. Then, we investigate if the pairs are concordant
(same sign of the change of x and y) or discordant (different signs) pairs

ij is concordant if .xj xi /.yj yi / > 0 (13.8)

ij is discordant if .xj xi /.yj yi / < 0:

Finally, we count the number of concordant (Tc ) and discordant (Td ) pairs and calculate
Kendall’s tau as
Tc Td
Kendall’s  D : (13.9)
T .T 1/=2
It can be shown that
4T C 10
 
Kendall’s  ! N 0;
d
; (13.10)
9T .T 1/
so it is straightforward to test  by a t-test.

Example 13.3 (Kendall’s tau) Suppose the data is

x y
2 7
10 9
3 10:

We then get the following changes

Changes of x Changes of y
x2 x1 D 10 2 D 8 y2 y1 D 9 7 D 2 concordant
x3 x1 D 3 2 D 5 y3 y1 D 10 7 D 3 discordant
x3 x2 D 3 10 D 13 y3 y2 D 10 9 D 1; discordant.

Kendall’s tau is therefore

1 2 1
D D :
3.3 1/=2 3
If x and y actually has bivariate normal distribution with correlation , then it can be
shown that on average we have
6
Spearman’s rho = arcsin.=2/   (13.11)

2
Kendall’s tau D arcsin./: (13.12)


258
In this case, all three measures give similar messages (although the Kendall’s tau tends to
be lower than the linear correlation and Spearman’s rho). This is illustrated in Figure 13.6.
Clearly, when data is not normally distributed, then these measures can give distinctly
different answers.

ρ and τ as a function of the linear correlation (Gaussian distribution)

1
Spearman’s ρ
Kendall’s τ
0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
Correlation

Figure 13.6: Spearman’s rho and Kendall’s tau if data has a bivariate normal distribution

A joint ˛-quantile exceedance probability measures how often two random variables
(x and y, say) are both above their ˛ quantile. Similarly, we can also define the probability
that they are both below their ˛ quantile

G˛ D Pr.x  x;˛ ; y  y;˛ /; (13.13)

x;˛ and y;˛ are ˛-quantile of the x- and y-distribution respectively.

In practice, this can be estimated from data by first finding the empirical ˛-quantiles
(Ox;˛ and Oy;˛ ) by simply sorting the data and then picking out the value of observation
˛T of this sorted list (do this individually for x and y). Then, calculate the estimate

1 if x t  Ox;˛ and y t  Oy;˛

(
1 XT
GO ˛ D ı t ; where ı t D (13.14)
T t D1 0 otherwise.

See Figure 13.7 for an illustration based on a joint normal distribution.

259
 Pr(x < quantile, y < quantile), Gauss Pr(x < quantile, y < quantile), Gauss
100 10
corr = 0.25
corr = 0.5
%

%
50 5

0 0
0 50 100 0 5 10
Quantile level, % Quatile level, %

Figure 13.7: Probability of joint low returns, bivariate normal distribution

13.4 Copulas

Reference: McNeil, Frey, and Embrechts (2005), Alexander (2008) 6, Jondeau, Poon, and
Rockinger (2007) 6
Portfolio choice and risk analysis depend crucially on the joint distribution of asset
returns. Empirical evidence suggest that many returns have non-normal distribution, es-
pecially when we focus on the tails. There are several ways of estimating complicated
(non-normal) distributions: using copulas is one. This approach has the advantage that
it proceeds in two steps: first we estimate the marginal distribution of each returns sepa-
rately, then we model the comovements by a copula.

13.4.1 Multivariate Distributions and Copulas

Any pdf can also be written as

f1;2 .x1 ; x2 / D c.u1 ; u2 /f1 .x1 /f2 .x2 /; with ui D Fi .xi /; (13.15)

where c./ is a copula density function and ui D Fi .xi / is the cdf value as in (13.1). The
extension to three or more random variables is straightforward.
Equation (13.15) means that if we know the joint pdf f1;2 .x1 ; x2 /—and thus also the
cdfs F1 .x1 / and F2 .x2 /—then we can figure out what the copula density function must
be. Alternatively, if we know the pdfs f1 .x1 / and f2 .x2 /—and thus also the cdfs F1 .x1 /
and F2 .x2 /—and the copula function, then we can construct the joint distribution. (This
is called Sklar’s theorem.) This latter approach will turn out to be useful.

260
 The correlation of x1 and x2 depends on both the copula and the marginal distribu-
tions. In contrast, both Spearman’s rho and Kendall’s tau are determined by the copula
only. They therefore provide a way of calibrating/estimating the copula without having to
involve the marginal distributions directly.

Example 13.4 (Independent X and Y ) If X and Y are independent, then we know that
f1;2 .x1 ; x2 / D f1 .x1 /f2 .x2 /, so the copula density function is just a constant equal to
one.

Remark 13.5 (Joint cdf) A joint cdf of two random variables (X1 and X2 ) is defined as

F1;2 .x1 ; x2 / D Pr.X1  x1 and X2  x2 /:

This cdf is obtained by integrating the joint pdf f1;2 .x1 ; x2 / over both variables
Z x1 Z x2
F1;2 .x1 ; x2 / D f1;2 .s; t/dsdt:
sD 1 tD 1

(Conversely, the pdf is the mixed derivative of the cdf, f1;2 .x1 ; x2 / D @2 F1;2 .x1 ; x2 /=@x1 @x2 .)
See Figure 13.8 for an illustration.

Remark 13.6 (From joint to univariate pdf) The pdf of x1 (also called the marginal pdf
R1
of x1 ) can be calculate from the joint pdf as f1 .x1 / D x2 D 1 f1;2 .x1 ; x2 /dx2 .

pdf of bivariate N() distribution, corr = 0.8 cdf of bivariate N() distribution, corr = 0.8

0.2 1

0.1 0.5
2 2
0 0
2 0 2 0
0 x 0 x
y −2 −2 y −2 −2

Figure 13.8: Bivariate normal distributions

261
Remark 13.7 (Joint pdf and copula density, n variables) For n variables (13.15) gener-
alizes to

f1;2;:::;n .x1 ; x2 ; : : : ; xn / D c.u1 ; u2 ; : : : ; un /f1 .x1 /f2 .x2 / : : : fn .xn /; with ui D Fi .xi /;

Remark 13.8 (Cdfs and copulas ) The joint cdf can be written as

F1;2 .x1 ; x2 / D C ŒF1 .x1 /; F2 .x2 /;

where C./ is the unique copula function. Taking derivatives gives (13.15) where

@2 C.u1 ; u2 /
c.u1 ; u2 / D :
@u1 @u2
Notice the derivatives are with respect to ui D Fi .xi /, not xi . Conversely, integrating the
density over both u1 and u2 gives the copula function C./.

13.4.2 The Gaussian and Other Copula Densities

The Gaussian copula density function is

 2 2
 1 21 2 C 2 22

1
c.u1 ; u2 / D p exp , with i D ˚ 1
.ui /; (13.16)
1 2 2.1 2 /

where ˚ 1 ./ is the inverse of an N.0; 1/ distribution. Notice that when using this function
in (13.15) to construct the joint pdf, we have to first calculate the cdf values ui D Fi .xi /
from the univariate distribution of xi (which may be non-normal) and then calculate
the quantiles of those according to a standard normal distribution i D ˚ 1 .ui / D
˚ 1 ŒFi .xi /.
It can be shown that assuming that the marginal pdfs (f1 .x1 / and f2 .x2 /) are normal
and then combining with the Gaussian copula density recovers a bivariate normal dis-
tribution. However, the way we typically use copulas is to assume (and estimate) some
other type of univariate distribution, for instance, with fat tails—and then combine with a
(Gaussian) copula density to create the joint distribution. See Figure 13.9 for an illustra-
tion.
A zero correlation ( D 0) makes the copula density (13.16) equal to unity—so the
joint density is just the product of the marginal densities. A positive correlation makes the
copula density high when both x1 and x2 deviate from their means in the same direction.

262
The easiest way to calibrate a Gaussian copula is therefore to set

 D Spearman’s rho, (13.17)

as suggested by (13.11).
Alternatively, the  parameter can calibrated to give a joint probability of both x1
and x2 being lower than some quantile as to match data: see (13.14). The values of this
probability (according to a copula) is easily calculated by finding the copula function
(essentially the cdf) corresponding to a copula density. Some results are given in remarks
below. See Figure 13.7 for results from a Gaussian copula. This figure shows that a
higher correlation implies a larger probability that both variables are very low—but that
the probabilities quickly become very small as we move towards lower quantiles (lower
returns).

Remark 13.9 (The Gaussian copula function ) The distribution function corresponding
to the Gaussian copula density (13.16) is obtained by integrating over both u1 and u2 and
the value is C.u1 ; u2 I / D" ˚# ."1 ; 2 /#!
where i is defined in (13.16) and ˚ is the bi-
0 1 
variate normal cdf for N ; . Most statistical software contains numerical
0  1
returns for calculating this cdf.

Remark 13.10 (Multivariate Gaussian copula density ) The Gaussian copula density for
n variables is  
1 1 0
c.u/ D p exp  .R 1
In / ;
jRj 2
where R is the correlation matrix with determinant jRj and  is a column vector with
i D ˚ 1 .ui / as the i th element.

The Gaussian copula is useful, but it has the drawback that it is symmetric—so the
downside and the upside look the same. This is at odds with evidence from many financial
markets that show higher correlations across assets in down markets. The Clayton copula
density is therefore an interesting alternative

c.u1 ; u2 / D . 1 C u1 ˛ C u2 ˛ / 2 1=˛
.u1 u2 / ˛ 1
.1 C ˛/; (13.18)

where ˛ ¤ 0. When ˛ > 0, then correlation on the downside is much higher than on the
upside (where it goes to zero as we move further out the tail).

263
 See Figure 13.9 for an illustration.
For the Clayton copula we have
˛
Kendall’s  D , so (13.19)
˛C2
2
˛D : (13.20)
1 
The easiest way to calibrate a Clayton copula is therefore to set the parameter ˛ according
to (13.20).
Figure 13.10 illustrates how the probability of both variables to be below their respec-
tive quantiles depend on the ˛ parameter. These parameters are comparable to the those
for the correlations in Figure 13.7 for the Gaussian copula, see (13.11)–(13.12). The figure
are therefore comparable—and the main point is that Clayton’s copula gives probabilities
of joint low values (both variables being low) that do not decay as quickly as according to
the Gaussian copulas. Intuitively, this means that the Clayton copula exhibits much higher
“correlations” in the lower tail than the Gaussian copula does—although they imply the
same overall correlation. That is, according to the Clayton copula more of the overall
correlation of data is driven by synchronized movements in the left tail. This could be
interpreted as if the correlation is higher in market crashes than during normal times.

Remark 13.11 (Multivariate Clayton copula density ) The Clayton copula density for n
variables is
Pn ˛ n 1=˛ Qn ˛ 1 Qn
c.u/ D 1 nC C .i




i D1 ui i D1 ui i D1 Œ1 1/˛ :

Remark 13.12 (Clayton copula function ) The copula function (the cdf) corresponding
to (13.18) is
C.u1 ; u2 / D . 1 C u1 ˛ C u2 ˛ / 1=˛ :

The following steps summarize how the copula is used to construct the multivariate
distribution.

1. Construct the marginal pdfs fi .xi / and thus also the marginal cdfs Fi .xi /. For in-
stance, this could be done by fitting a distribution with a fat tail. With this, calculate
the cdf values for the data ui D Fi .xi / as in (13.1).

264
 Gaussian copula density, corr = -0.5 Gaussian copula density, corr = 0

5 5

0 0
2 2
0 0
x2 −2 0 2 x2 −2 2
−2 x1 −2 x1 0

Gaussian copula density, corr = 0.5 Clayton copula density, α = 0.5(τ = 0.2)

5 5

0 0
2 2
0 0
x2 −2 0 2 x2 −2 2
−2 x1 −2 x1 0

Figure 13.9: Copula densities (as functions of xi )

Pr(x < quantile, y < quantile), Clayton Pr(x < quantile, y < quantile), Clayton
100 10
α = 0.16
α = 0.33
%

50 5

0 0
0 50 100 0 5 10
Quantile level, % Quantile level, %

Figure 13.10: Probability of joint low returns, Clayton copula

2. Calculate the copula density as follows (for the Gaussian or Clayton copulas, re-
spectively):

265
 (a) for the Gaussian copula (13.16)
i. assume (or estimate/calibrate) a correlation  to use in the Gaussian cop-
ula
ii. calculate i D ˚ 1
.ui /, where ˚ 1
./ is the inverse of a N.0; 1/ distribu-
tion
iii. combine to get the copula density value c.u1 ; u2 /
(b) for the Clayton copula (13.18)
i. assume (or estimate/calibrate) an ˛ to use in the Clayton copula (typically
based on Kendall’s  as in (13.20))
ii. calculate the copula density value c.u1 ; u2 /

3. Combine the marginal pdfs and the copula density as in (13.15), f1;2 .x1 ; x2 / D
c.u1 ; u2 /f1 .x1 /f2 .x2 /, where ui D Fi .xi / is the cdf value according to the marginal
distribution of variable i .

See Figures 13.11–13.12 for illustrations.

Remark 13.13 (Tail Dependence ) The measure of lower tail dependence starts by find-
ing the probability that X1 is lower than its qth quantile (X1  F1 1 .q/) given that X2 is
lower than its qth quantile (X2  F2 1 .q/)

l D PrŒX1  F1 1 .q/jX2  F2 1 .q/;

and then takes the limit as the quantile goes to zero

l D limq!0 PrŒX1  F1 1 .q/jX2  F2 1 .q/:

It can be shown that a Gaussian copula gives zero or very weak tail dependence,
unless the correlation is 1. It can also be shown that the lower tail dependence of the
Clayton copula is
l D 2 1=˛ if ˛ > 0

and zero otherwise.

266
 Joint pdf, Gaussian copula, corr = -0.5 Joint pdf, Gaussian copula, corr = 0
2 2
1 1
x2

x2
0 0
−1 −1
−2 −2
−2 −1 0 1 2 −2 −1 0 1 2
x1 x1
Notice: marginal distributions are N(0,1)

Joint pdf, Gaussian copula, corr = 0.5 Joint pdf, Clayton copula, α = 0.5
2 2
1 1
x2

x2

0 0
−1 −1
−2 −2
−2 −1 0 1 2 −2 −1 0 1 2
x1 x1

Figure 13.11: Contours of bivariate pdfs

13.5 Joint Tail Distribution

The methods for estimating the (marginal, that is, for one variable at a time) distribution
of the lower tail can be combined with a copula to model the joint tail distribution. In
particular, combining the generalized Pareto distribution (GPD) with the Clayton copula
provides a flexible way.
This can be done by first modelling the loss (X t D R t ) beyond some threshold (u),
that is, the variable X t u with the GDP. To get a distribution of the return, we simply use
the fact that pdfR . z/ D pdfX .z/ for any value z. Then, in a second step we calibrate the
copula by using Kendall’s  for the subsample when both returns are less than u. Figures
13.13–13.15 provide an illustration.

Remark 13.14 Figure 13.13 suggests that the joint occurrence (of these two assets) of
really negative returns happens more often than the estimated normal distribution would

267
 Joint pdf, Gaussian copula, corr = -0.5 Joint pdf, Gaussian copula, corr = 0
2 2
1 1
x2

x2
0 0
−1 −1
−2 −2
−2 −1 0 1 2 −2 −1 0 1 2
x1 x1
Notice: marginal distributions are t5

Joint pdf, Gaussian copula, corr = 0.5 Joint pdf, Clayton copula, α = 0.5
2 2
1 1
x2

x2

0 0
−1 −1
−2 −2
−2 −1 0 1 2 −2 −1 0 1 2
x1 x1

Figure 13.12: Contours of bivariate pdfs

suggest. For that reason, the joint distribution is estimated by first fitting generalized
Pareto distributions to each of the series and then these are combined with a copula as in
(13.15) to generate the joint distribution. In particular, the Clayton copula seems to give
a long joint negative tail.

To find the implication for a portfolio of several assets with a given joint tail distribu-
tion, we often resort to simulations. That is, we draw random numbers (returns for each
of the assets) from the joint tail distribution and then study the properties of the portfolio
(with say, equal weights or whatever). The reason we simulate is that it is very hard to
actually calculate the distribution of the portfolio by using mathematics, so we have to
rely on raw number crunching.
The approach proceeds in two steps. First, draw n values for the copula (ui ; i D
1; : : : ; n). Second, calculate the random number (“return”) by inverting the cdf ui D

268
 Prob(both returns < quantile) Prob(both returns < quantile), zoomed in
100 5
Data
estimated N() 4
3
%

%
50
2
1
0 0
0 50 100 0 5 10
Quantile level, % Quantile level, %

Daily US data 1979:1-2012:12

small stocks and large stocks

Figure 13.13: Probability of joint low returns

Fi .xi / in (13.15) as
xi D Fi 1 .ui /; (13.21)

where Fi 1 ./ is the inverse of the cdf.

Remark 13.15 (To draw n random numbers from a Gaussian copula) First, draw n num-
bers from an N.0; R/ distribution, where R is the correlations matrix. Second, calculate
ui D ˚.xi /, where ˚ is the cdf of a standard normal distribution.

Remark 13.16 (To draw n random numbers from a Clayton copula) First, draw xi for
i D 1; : : : ; n from a uniform distribution (between 0 and 1). Second, draw v from a
gamma(1=˛; 1) distribution. Third, calculate ui D Œ1 ln.xi /=v 1=˛ for i D 1; : : : ; n.
These ui values are the marginal cdf values.

Remark 13.17 (Inverting a normal and a generalised Pareto cdf) Must numerical soft-
ware packages contain a routine for investing a normal cdf. My lecture notes on the
Generalised Pareto distribution shows how to invert that distribution.

Such simulations can be used to quickly calculate the VaR and other risk measures
for different portfolios. A Clayton copula with a high ˛ parameter (and hence a high
Kendall’s ) has long lower tail with highly correlated returns: when asset takes a dive,

269
 Loss distribution, small stocks Loss distribution, large stocks
u = 0.5, Pr(loss > u) = 17.8% u = 0.5, Pr(loss > u) = 24.9%
0.2 ξ = 0.20, β = 0.61 0.2 ξ = 0.12, β = 0.68
0.15 Daily US data 1979:1-2012:12 0.15
0.1 0.1
0.05 0.05
0 0
1 2 3 4 5 1 2 3 4 5
Loss, % Loss, %

Return distribution, small stocks Return distribution, large stocks

(Only lower tail is shown) (Only lower tail is shown)
0.2 0.2
0.15 GP 0.15
0.1 Normal 0.1
0.05 0.05
0 0
−5 −4 −3 −2 −1 −5 −4 −3 −2 −1
Return, % Return, %

Figure 13.14: Estimation of marginal loss distributions

other assets are also likely to decrease. That is, the correlation in the lower tail of the
return distribution is high, which will make the VaR high.
Figures 13.16–13.17 give an illustration of how the movements in the lower get more
synchronised as the ˛ parameter in the Clayton copula increases.

Bibliography
Alexander, C., 2008, Market Risk Analysis: Practical Financial Econometrics, Wiley.

Ang, A., and J. Chen, 2002, “Asymmetric correlations of equity portfolios,” Journal of
Financial Economics, 63, 443–494.

Jondeau, E., S.-H. Poon, and M. Rockinger, 2007, Financial Modeling under Non-
Gaussian Distributions, Springer.

270
 Joint pdf, independent copula Joint pdf, Gaussian copula

−1 −1
large stocks

large stocks
−2 −2

−3 −3
Spearman’s ρ = 0.43
−4 −4
−4 −3 −2 −1 −4 −3 −2 −1
small stocks small stocks

Joint pdf, Clayton copula

Daily US data 1979:1-2012:12
−1 The marginal distributions of the losses
large stocks

are estimated GP distributions

−2

−3
Kendall’s τ = 0.30, α = 0.87
−4
−4 −3 −2 −1
small stocks

Figure 13.15: Joint pdfs with different copulas

McNeil, A. J., R. Frey, and P. Embrechts, 2005, Quantitative risk management, Princeton
University Press.

271
 Gaussian copula, ρ = 0.49 Clayton copula, α = 0.56
4 4
τ = 0.22

2 2
Asset 2

Asset 2
0 0

−2 −2
marginal pdf:
normal
−4 −4
−4 −2 0 2 4 −4 −2 0 2 4
Asset 1 Asset 1
Clayton copula, α = 1.06 Clayton copula, α = 2.06
4 4
τ = 0.35 τ = 0.51

2 2
Asset 2

Asset 2

0 0

−2 −2

−4 −4
−4 −2 0 2 4 −4 −2 0 2 4
Asset 1 Asset 1

Figure 13.16: Example of scatter plots of two asset returns drawn from different copulas

272
 Quantile of equally weighted portfolio, different copulas
−0.8
Notice: VaR95% = − (the 5% quantile)
−1

−1.2

−1.4

−1.6

−1.8
N
−2 Clayton, α = 0.56
Clayton, α = 1.06
Clayton, α = 2.06
−2.2
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Prob of lower outcome

Figure 13.17: Quantiles of an equally weighted portfolio of two asset returns drawn from
different copulas

273
14 Option Pricing and Estimation of Continuous Time
Processes
Reference: Hull (2006) 19, Elton, Gruber, Brown, and Goetzmann (2003) 22 or Bodie,
Kane, and Marcus (2005) 21
Reference (advanced): Taylor (2005) 13–14; Campbell, Lo, and MacKinlay (1997) 9;
Gourieroux and Jasiak (2001) 12–13
More advanced material is denoted by a star ( ). It is not required reading.

14.1 The Black-Scholes Model

14.1.1 The Black-Scholes Option Price Model

A European call option contract traded (contracted and paid) in t may stipulate that the
buyer of the contract has the right (not the obligation) to buy one unit of the underlying
asset (from the issuer of the option) in t C m at the strike price K. The option payoff
(in t C m) is clearly max .0; S t Cm K/ ;where S t Cm is the asset price, and K the strike
price. See Figure 14.1 for the timing convention.

t t Cm
buy option: if S > K: pay
agree on K, pay C K and get asset,
otherwise: do nothing

Figure 14.1: Timing convention of option contract

274
 Pdf of N(0,1) Pdf of N(0,1)
0.4 Pr(x ≤ −1) = 0.4 Pr(x ≤ 0.5) =
0.16 0.69
0.3 0.3

0.2 0.2

0.1 0.1

0 0
−2 0 2 −2 0 2
x x

Cdf of N(0,1)
1

0.5

0
−2 0 2
x

Figure 14.2: Pdf and cdf of N(0,1)

The Black-Scholes formula for a European call option price is

rm
p
C t D S t ˚.d1 / Ke ˚.d1  m/, where (14.1)
ln.S t =K/ C r C  2 =2 m


d1 D p :
 m

where ˚./ is the cumulative distribution function of a standard normal, N.0; 1/, variable.
For instance, ˚.2/ is the probability that the variable is less or equal to two, see Figure
14.2. In this equation, S0 is the price of the underlying asset in period t , and r is the
continuously compounded interest rate (on an annualized basis).
Some basic properties of the model are illustrated in Figure 14.3. In particular, the
call option price is increasing in the volöatility and decreasing in the strike price.
The B-S formula can be derived from several stochastic processes of the underlying
asset price (discussed below), but they all imply that the distribution of log asset price in

275
 BS call, S varies BS call, K varies BS call, σ varies
15 15
6
10 10
4

5 5 2

0 0 0
30 40 50 60 30 40 50 60 0 0.2 0.4
Asset price Strike price Standard deviation

BS call, m varies BS call, y varies The base case has

S, K, m, y, and σ:
6 6 42 42 0.5 0.05 0.2

4 4

2 2

0 0
0 0.2 0.4 0 0.05 0.1
Expiration Interest rate

Figure 14.3: Call option price, Black-Scholes model

t C m (conditional on the information in t) is normal with some mean ˛ (not important

for the option price) and the variance m 2

ln S t Cm  N.˛; m 2 /: (14.2)

Option pricing is basically about forecasting the volatility (until expiration of the op-
tion) of the underlying asset. This is clear from the Black-Scholes model where the only
unknown parameter is the volatility. It is also true more generally—which can be seen in
at least two ways. First, a higher volatility is good for an owner of a call option since it
increases the upside potential (higher probability of a really good outcome), at the same
time as the down side is protected. Second, a many option portfolios highlight how volatil-
ity matters for the potential profits. For instance, a straddle (a long position in both a call
and a put at the same strike price) pays off if the price of the underlying asset moves a

276
 Distribution of future asset price Distribution of future asset price

0.06 K = 42 0.06 K = 42
Std = 0.14 Std = 0.21
0.04 0.04

0.02 0.02

0 0
20 30 40 50 60 70 20 30 40 50 60 70
Asset price Asset price

Figure 14.4: Distribution of future stock price

Profit of straddle, call + put

Call(K)
Put(K)
Straddle

Asset price (at expiration)

Figure 14.5: Profit of straddle portfolio

lot (in either direction) from the strike price, that is, when volatility is high. See Figures
14.4–14.5 for illustrations.

277
14.1.2 Implied Volatility

The pricing formula (14.1) contains only one unknown parameter: the standard deviation
 in the distribution of ln S t Cm , see (14.2). With data on the option price, spot price,
the interest rate, and the strike price, we can solve for standard deviation: the implied
volatility. This should not be thought of as an estimation of an unknown parameter—
rather as just a transformation of the option price. Notice that we can solve (by trial-and-
error or some numerical routine) for one implied volatility for each available strike price.
See Figure 14.3 for an illustration.
If the Black-Scholes formula is correct, that is, if the assumption in (14.2) is cor-
rect, then these volatilities should be the same across strike prices—and it should also be
constant over time.
In contrast, it is often found that the implied volatility is a “smirk” (equity markets)
or “smile” (FX markets) shaped function of the strike price. One possible explanation for
a smirk shape is that market participants assign a higher probability to a dramatic drop in
share prices than a normal distribution suggests. A possible explanation for a smile shape
is that the (perceived) distribution of the future asset price has relatively more probability
mass in the tails (“fat tails”) than a normal distribution has. See Figures 14.6–14.7 for
illustrations. In addition, the implied volatilities seems to move considerably over time—
see Figure 14.8 for a time series of implied volatilities

14.1.3 Brownian Motion without Mean Reversion: The Random Walk

The basic assumption behind the B-S formula (14.1) is that the log price of the underlying
asset, ln S t , follows a geometric Brownian motion—with or without mean reversion.
This section discusses the standard geometric Brownian motion without mean rever-
sion
d ln S t D dt C d W t ; (14.3)

where d ln S t is the change in the log price (the return) over a very short time interval. On
the right hand side,  is the drift (typically expressed on annual basis), dt just indicates
the change in time,  is the standard deviation (per year), and d W t is a random compo-
nent (Wiener process) that has an N.0; 1/ distribution if we cumulate d W t over a year
R1
( 0 d W t  N.0; 1/). By comparing (14.1) and (14.3) we notice that only the volatility
( ), not the drift (), show up in the option pricing formula. In essence, the drift is al-

278
 Implied volatility, SMI option
1
2009-07-14
2009-08-14
0.8

0.6

0.4

0.2

0
3000 4000 5000 6000 7000 8000
Strike price

Figure 14.6: Implied volatilities of SMI options, selected dates

ready accounted for by the current spot price in the option pricing formula (as the spot
price certainly depends on the expected drift of the asset price).

Remark 14.1 (Alternative stock price process ) If we instead of (14.3) assume the pro-
cess dS t D SQ t dt C S t d W t , then we get the same option price. The reason is that Itô’s
lemma tells us that (14.3) implies this second process with Q D  C  2 =2. The difference
is only in terms of the drift, which does not show up (directly, at least) in the B-S formula.

Remark 14.2 ((14.3) as a limit of a discrete time process ) (14.3) can be thought of as
p
the limit of the discrete time process ln S t Ch ln S t D h C  h" t Ch (where " t is iid
N.0; 1/) as the time interval h becomes very small.

We can only observe the value of the asset price at a limited number of times, so we
need to understand what (14.3) implies for discrete time intervals. It is straightforward to
show that (14.3) implies that we have normally distributed changes (returns) and that the
changes (returns) are uncorrelated (for non-overlapping data)

ln.S t Ch =S t /  N.h;  2 h/ (14.4)

CovŒln.S t =S t h /; ln.S tCh =S t / D 0: (14.5)

279
 SMI implied volatility atm SMI, average iv - iv(atm)

0.3
0.1
0.25
0
0.2

0.15 −0.1
Jan Apr Jul −10 −5 0 5 10
distance of strike and forward price, %

SMI, std(iv - iv(atm))

0.2 SMI options (daily), 2009:1-2009:9

0.15

0.1

0.05

0
−10 −5 0 5 10
distance of strike and forward price, %

Figure 14.7: Implied volatilities

Notice that both the drift and the variance scale linearly with the horizon h. The reason
is, or course, that the growth rates (even for the infinitesimal horizon) are iid.

Remark 14.3 (iid random variable in discrete time) Suppose x t has the constant mean 
and a variance  2 . Then E.x t Cx t 1 / D 2 and Var.x t Cx t 1 / D 2 2 C2 Cov.x t ; x t 1 /.
If x t is iid, then the covariance is zero, so Var.x t C x t 1 / D 2 2 . In this case, both mean
and variance scale linearly with the horizon.

14.1.4 Brownian Motion with Mean Reversion

The mean reverting Ornstein-Uhlenbeck process is

d ln S t D . ln S t /dt C d W t , with  > 0: (14.6)

280
 CBOE volatility index (VIX)
Kuwait Stock market crash LTCM/Russia pharma Higher rates Lehman Fukushima
80 Gulf war Asia 9/11 Iraq sub prime Greece

70

60

50

40

30

20

10
1990 1995 2000 2005 2010

Figure 14.8: CBOE VIX, summary measure of implied volatities (30 days) on US stock
markets

This process makes ln S t revert back to the mean , and the mean reversion is faster if 
is large. It is used in, for instance, the Vasicek model of interest rates.
To estimate the parameters in (14.6) on real life data, we (once again) have to under-
stand what the model implies for discretely sampled data. It can be shown that it implies
a discrete time AR(1)

ln S t D ˛ C  ln S t h C " t , with (14.7)

h
De ; ˛ D .1 /, and " t  N 0;  2 .1 2 /=.2/ : (14.8)
 

We know that the maximum likelihood estimator (MLE) of the discrete AR(1) is least
squares combined with the traditional estimator of the residual variance. MLE has the
further advantage of being invariant to parameter transformations, which here means that
the MLE of ;  and  2 can be backed out from the LS estimates of ; ˛ and Var." t ) by
using (14.8).

Example 14.4 Suppose ;  and  2 are 2, 0, and 0.25 respectively—and the periods are
years (so one unit of time corresponds to a year) . Equations (14.7)–(14.8) then gives the

281
 0 h 2h T D nh

0 h 2h T D nh

Figure 14.9: Two different samplings with same time span T

following AR(1) for weekly (h D 1=52) data

ln S t D 0:96 ln S t h C " t with Var." t /  0:24:

14.2 Estimation of the Volatility of a Random Walk Process

This section discusses different ways of estimating the volatility. We will assume that we
have data for observations in  D 1; 2; ::; n. This could be 5-minute intervals, days or
weeks or whatever. Let the time between  and  C 1 be h (years). The sample therefore
stretches over T D nh periods (years). For instance, for daily data h D 1=365 (or
possibly something like 1=252 if only trading days are counted). Instead, with weekly
data h D 1=52. See Figure 14.9 for an illustration.

14.2.1 Standard Approach

We first estimate the variance for the sampling frequency we have, and then convert to the
annual frequency.
According to (14.4) the growth rates, ln.S t =S t h /, are iid over any sampling fre-
quency. To simplify the notation, let y D ln.S =S 1 / be the observed growth rates.
The classical estimator of the variance of an iid data series is
Xn
sO 2 D .y N 2 =n, where
y/ (14.9)
 D1
Xn
yN D y =n: (14.10)
 D1

(This is also the maximum likelihood estimator.) To annualize these numbers, use

O 2 D sO 2 = h, and O D y=
N h: (14.11)

282
Example 14.5 If .y;N sO 2 / D .0:001; 0:03/ on daily data, then the annualized values are
.;  2 / D .0:001  250; 0:03  250/ D .0:25; 7:5/ if we assume 250 trading days per
year.

Notice that is can be quite important to subtract the mean drift, y.

N Recall that for any
random variable, we have
 2 D E.x 2 / 2 ; (14.12)

so a non-zero mean drives a wedge between the variance (which we want) and the second
moment (which we estimate if we assume yN D 0).

Example 14.6 (US stock market volatility) For the US stock market index excess return
since WWII we have approximately a variance of 0:162 and a mean of 0:08. In this case,
(14.12) becomes
0:162 D E.x 2 / 0:082 , so E.x 2 /  0:182 :

Assuming that the drift is zero gives an estimate of the variance equal to 0:182 which is
25% too high.

Remark 14.7 ( Variance vs second moment, the effect of the maturity) Suppose we are
interested in the variance over an m-period horizon, for instance, because we want to
price an option that matures in t C m. How important is it then to use the variance (m 2 )
rather than the second moment? The relative error is
Second moment - variance m2 2 m2
D D 2 ;
variance m 2 
where we have used the fact that the second moment equals the variance plus the squared
mean (cf (14.12)). Clearly, this relative exaggeration is zero if the mean is zero. The
relative exaggeration is small if the maturity is small.

If we have high frequency data on the asset price or the return, then we can choose
which sampling frequency to use in (14.9)–(14.10). Recall that a sample with n obser-
vations (where the length of time between the observations is h) covers T D nh periods
(years). It can be shown that the asymptotic variances (that is, the variances in a very
large sample) of the estimators of  and  2 in (14.9)–(14.11) are

Var./
O D  2 =T and Var.O 2 / D 2 4 =n: (14.13)

283
Therefore, to get a precise estimator of the mean drift, , a sample that stretches over
a long period is crucial: it does not help to just sample more frequently. However, the
sampling frequency is crucial for getting a precise estimator of  2 , while a sample that
stretches over a long period is unimportant. For estimating the volatility (to use in the B-S
model) we should therefore use high frequency data.

14.2.2 Exponentially Weighted Moving Average

The traditional estimator is based on the assumption that volatility is constant—which is

consistent with the assumptions of the B-S model. In reality, volatility is time varying.
A practical ad hoc approach to estimate time varying volatility is to modify (14.9)–
(14.10) so that recent observations carry larger weight. The exponentially weighted mov-
ing average (EWMA) model lets the weight for lag s be .1 /s where 0 <  < 1. If
we assume that yN is the same in all periods, then we have

sO2 D Os2 1 C .1 / .y 1 N 2;

y/ (14.14)

where  is the current period and  1 the pervious period (say, today and yesterday).
Clearly, a higher  means that old data plays a larger role—and at the limit as  goes
towards one, we have the traditional estimator. See Figure 14.11 for a comparison using
daily US equity returns. This method is commonly used by practitioners. For instance,
the RISK Metrics is based on  D 0:94 on daily data. Alternatively,  can be chosen to
minimize some criterion function.

Remark 14.8 (EWMA with time-variation in the mean ) If we want also the mean to be
time-varying, then we can use the estimator

sO2 D .1 yN /2 C  .y yN /2 C 2 .y yN /2 C : : :

 
/ .y 1 2 3

yN D Œy 1 C y 2 C y 3 C : : : =. 1/:

Notice that the mean is estimated as a traditional sample mean, using observations 1 to
 1. This guarantees that the variance will always be a non-negative number.

It should be noted, however, that the B-S formula does not allow for random volatility.

284
 GARCH std, annualized EWMA std, annualized, λ = 0.99
50 50
40 40
30 30
20 20
10 10
0 0
1980 1990 2000 2010 1980 1990 2000 2010

EWMA std, annualized, λ = 0.9

S&P 500 (daily) 1954:1-2013:4 50
40
AR(1) of excess returns
with GARCH(1,1) errors 30
AR(1) coef: 0.09 20
GARCH coefs: 0.08 0.91 10
0
1980 1990 2000 2010

Figure 14.10: Different estimates of US equity market volatility

Std, EWMA estimate, λ = 0.9 CBOE volatility index (VIX)

50 50
40 40
30 30
20 20
10 10
0 0
1990 1995 2000 2005 2010 1990 1995 2000 2005 2010
S&P 500, daily data 1954:1-2013:4

Figure 14.11: Different estimates of US equity market volatility

14.2.3 Autoregressive Conditional Heteroskedasticity

The model with Autoregressive Conditional Heteroskedasticity (ARCH) is a useful tool

for estimating the properties of volatility clustering. The first-order ARCH expresses

285
volatility as a function of the latest squared shock

s2 D ˛0 C ˛1 u2 1 ; (14.15)

where u is a zero-mean variable. The model requires ˛0 > 0 and 0  ˛1 < 1 to

guarantee that the volatility stays positive and finite. The variance reverts back to an
average variance (˛0 =.1 ˛1 /). The rate of mean reversion is ˛1 , that is, the variance
behaves much like an AR(1) model with an autocorrelation parameter of ˛1 . The model
parameters are typically estimated by maximum likelihood. Higher-order ARCH models
include further lags of the squared shocks (for instance, u2 2 ).
Instead of using a high-order ARCH model, it is often convenient to use a first-order
generalized ARCH model, the GARCH(1,1) model. It adds a term that captures direct
autoregression of the volatility

s2 D ˛0 C ˛1 u2 1 C ˇ1 s2 1 : (14.16)

We require that ˛0 > 0, ˛1  0, ˇ1  0, and ˛1 C ˇ1 < 1 to guarantee that the

volatility stays positive and finite. This is very similar to the EMA in (14.14), except that
the variance reverts back to the mean (˛0 =.1 ˛1 ˇ1 /). The rate of mean reversion is
˛1 C ˇ1 , that is, the variance behaves much like an AR(1) model with an autocorrelation
parameter of ˛1 C ˇ1 .

14.2.4 Time-Variation in Volatility and the B-S Formula

The ARCH and GARCH models imply that volatility is random, so they are (strictly
speaking) not consistent with the B-S model. However, they are often combined with the
B-S model to provide an approximate option price. See Figure 14.12 for a comparison
of the actual distribution of the log asset price (actually, cumulated returns, so assuming
that the intial log asset price is zero) at different horizons (1 and 10 days) when the daily
returns are generated by a GARCH model—and a normal distribution with the same mean
and variance. To be specific, the figure shows the distribution of the futurelog asset price
calculated as

ln S t C r tCh , or (14.17)
ln S t C 10 (14.18)
P
i D1 r t Cih ;

286
where each of the returns (r t Cih ) is drawn from an N.0; s t2Cih ) distribution where the
variance follows the GARCH(1,1) process like in (14.16).
It is clear the normal distribution is a good approximation unless the ARCH com-
ponent (˛1 lagged squared shock) dominates the GARCH component (ˇ1 lagged vari-
ance).
Intuitively, we get (almost) a normal distribution when the random part of the volatility
(the ARCH component) is relatively small compared to the non-random part (the GARCH
component). For instance, if there is no random part at all, then we get exactly a normal
distribution (the sum of normally distributed variables is normally distributed—if all the
variances are deterministic).
However, to get an option price that is perfectly consistent with a GARCH process,
we need to go beyond the B-S model (see, for instance, Heston and Nandi (2000)).

Remark 14.9 (Time-varying, but deterministic volatility ) A time-varying, but non-random

volatility could be consistent with (14.2): if ln S tCm is the sum (integral) of normally dis-
tributed changes with known (but time-varying variances), then this sum has a normal
distribution (recall: if the random variables x and y are normally distributed, so is xCy).
A random variance does not fit this case, since a variable with a random variance is not
normally distributed.

Bibliography
Bodie, Z., A. Kane, and A. J. Marcus, 2005, Investments, McGraw-Hill, Boston, 6th edn.

Campbell, J. Y., A. W. Lo, and A. C. MacKinlay, 1997, The econometrics of financial

markets, Princeton University Press, Princeton, New Jersey.

Elton, E. J., M. J. Gruber, S. J. Brown, and W. N. Goetzmann, 2003, Modern portfolio

theory and investment analysis, John Wiley and Sons, 6th edn.

Gourieroux, C., and J. Jasiak, 2001, Financial econometrics: problems, models, and
methods, Princeton University Press.

Heston, S. L., and S. Nandi, 2000, “A closed-form GARCH option valuation model,”
Review of Financial Studies, 13, 585–625.

287
 Distribution from GARCH, T = 1
0.4
Normal
0.3 Simulated

0.2

0.1

0
−5 0 5
Return over 1 day

Distribution of cumulated returns, T = 10 Distribution from GARCH, T = 10

0.2 GARCH parameters: 0.2 GARCH parameters:
(α, β) = (0.8, 0.09) (α, β) = (0.09, 0.8)
0.15 0.15

0.1 0.1

0.05 0.05

0 0
−10 −5 0 5 10 −10 −5 0 5 10
Cumulated return over 10 days Cumulated return over 10 days

Figure 14.12: Comparison of normal and simulated distribution of m-period returns

Hull, J. C., 2006, Options, futures, and other derivatives, Prentice-Hall, Upper Saddle
River, NJ, 6th edn.

Taylor, S. J., 2005, Asset price dynamics, volatility, and prediction, Princeton University
Press.

288
15 Event Studies
Reference: Bodie, Kane, and Marcus (2005) 12.3 or Copeland, Weston, and Shastri
(2005) 11
Reference (advanced): Campbell, Lo, and MacKinlay (1997) 4
More advanced material is denoted by a star ( ). It is not required reading.

15.1 Basic Structure of Event Studies

The idea of an event study is to study the effect (on stock prices or returns) of a special
event by using a cross-section of such events. For instance, what is the effect of a stock
split announcement on the share price? Other events could be debt issues, mergers and
acquisitions, earnings announcements, or monetary policy moves.
The event is typically assumed to be a discrete variable. For instance, it could be a
merger or not or if the monetary policy surprise was positive (lower interest than expected)
or not. The basic approach is then to study what happens to the returns of those assets
that have such an event.
Only news should move the asset price, so it is often necessary to explicitly model
the previous expectations to define the event. For earnings, the event is typically taken to
be the earnings announcement minus (some average of) analysts’ forecast. Similarly, for
monetary policy moves, the event could be specified as the interest rate decision minus
previous forward rates (as a measure of previous expectations).
The abnormal return of asset i in period t is

ui t D R i t Rinormal
t ; (15.1)

where Ri t is the actual return and the last term is the normal return (which may differ
across assets and time). The definition of the normal return is discussed in detail in Section
15.2. These returns could be nominal returns, but more likely (at least for slightly longer
horizons) real returns or excess returns.
Suppose we have a sample of n such events (“assets”). To keep the notation (reason-

289
 -1 0 1 -1 0 1 time

firm 1 firm 2

Figure 15.1: Event days and windows

ably) simple, we “normalize” the time so period 0 is the time of the event. Clearly the
actual calendar time of the events for assets i and j are likely to differ, but we shift the
time line for each asset individually so the time of the event is normalized to zero for
every asset. See Figure 15.1 for an illustration.
To control for information leakage and slow price adjustment, the abnormal return is
often calculated for some time before and after the event: the “event window” (often ˙20
days or so). For day s (that is, s days after the event time 0), the cross sectional average
abnormal return is
uN s D niD1 ui s =n: (15.2)
P

For instance, uN 2 is the average abnormal return two days after the event, and uN 1 is for
one day before the event.
The cumulative abnormal return (CAR) of asset i is simply the sum of the abnormal
return in (15.1) over some period around the event. It is often calculated from the be-
ginning of the event window. For instance, if the event window starts at w, then the
q-period (day?) car for firm i is

cari q D ui; w C ui; wC1 C : : : C ui; wCq 1 : (15.3)

The cross sectional average of the q-period car is

Pn
carq D i D1 cari q =n: (15.4)

See Figure 15.2 for an empirical example.

Example 15.1 (Abnormal returns for ˙ day around event, two firms) Suppose there are
two firms and the event window contains ˙1 day around the event day, and that the

290
 Cumulative excess return (average) with 90% conf band
100

80
Returns, %

60

40

20

Sample: 196 IPOs on the Shanghai Stock Exchange, 2001-2004

0
0 5 10 15 20 25
Days after IPO

Figure 15.2: Event study of IPOs in Shanghai 2001–2004. (Data from Nou Lai.)

abnormal returns (in percent) are

Time Firm 1 Firm 2 Cross-sectional Average

1 0:2 0:1 0:05
0 1:0 2:0 1:5
1 0:1 0:3 0:2

We have the following cumulative returns

Time Firm 1 Firm 2 Cross-sectional Average

1 0:2 0:1 0:05
0 1:2 1:9 1:55
1 1:3 2:2 1:75

15.2 Models of Normal Returns

This section summarizes the most common ways of calculating the normal return in
(15.1). The parameters in these models are typically estimated on a recent sample, the
“estimation window,” that ends before the event window. See Figure 15.3 for an illustra-

291
tion. (When there is no return data before the event window (for instance, when the event
is an IPO), then the estimation window can be after the event window.)
In this way, the estimated behaviour of the normal return should be unaffected by the
event. It is almost always assumed that the event is exogenous in the sense that it is not
due to the movements in the asset price during either the estimation window or the event
window. This allows us to get a clean estimate of the normal return.
The constant mean return model assumes that the return of asset i fluctuates randomly
around some mean i

Ri t D i C "i t with (15.5)

E."i t / D Cov."i t ; "i;t s / D 0:

This mean is estimated by the sample average (during the estimation window). The nor-
mal return in (15.1) is then the estimated mean. O i so the abnormal return (in the estima-
tion window) becomes "Oi t . During the event window, we calculate the abnormal return
as
ui t D Ri t O i : (15.6)

The standard error of this is estimated by the standard error of "Oi t (in the estimation
window).
The market model is a linear regression of the return of asset i on the market return

Ri t D ˛i C ˇi Rmt C "i t with (15.7)

E."i t / D Cov."i t ; "i;t s / D Cov."i t ; Rmt / D 0:

Notice that we typically do not impose the CAPM restrictions on the intercept in (15.7).
The normal return in (15.1) is then calculated by combining the regression coefficients
with the actual market return as ˛O i C ˇOi Rmt , so the the abnormal return in the estimation
window is "Oi t . For the event window we calculate the abnormal return as

ui t D R i t ˛O i ˇOi Rmt : (15.8)

The standard error of this is estimated by the standard error of "Oi t (in the estimation
window).
When we restrict ˛i D 0 and ˇi D 1, then this approach is called the market-adjusted-

292
 0 time

estimation window event window

(for normal return)

Figure 15.3: Event and estimation windows

return model. This is a particularly useful approach when there is no return data before
the event, for instance, with an IPO. For the event window we calculate the abnormal
return as
ui t D Ri t Rmt (15.9)

and the standard error of it is estimated by Std.Ri t Rmt / in the estimation window.
Recently, the market model has increasingly been replaced by a multi-factor model
which uses several regressors instead of only the market return. For instance, Fama and
French (1993) argue that (15.7) needs to be augmented by a portfolio that captures the
different returns of small and large firms and also by a portfolio that captures the different
returns of firms with high and low book-to-market ratios.
Finally, another approach is to construct a normal return as the actual return on assets
which are very similar to the asset with an event. For instance, if asset i is a small man-
ufacturing firm (with an event), then the normal return could be calculated as the actual
return for other small manufacturing firms (without events). In this case, the abnormal
return becomes the difference between the actual return and the return on the matching
portfolio. This type of matching portfolio is becoming increasingly popular. For the event
window we calculate the abnormal return as

ui t D R i t Rpt ; (15.10)

where Rpt is the return of the matching portfolio. The standard error of it is estimated by
Std.Ri t Rpt / in the estimation window.
All the methods discussed here try to take into account the risk premium on the asset.
It is captured by the mean in the constant mean mode, the beta in the market model, and

293
by the way the matching portfolio is constructed. However, sometimes there is no data in
the estimation window. The typical approach is then to use the actual market return as the
normal return—that is, to use (15.7) but assuming that ˛i D 0 and ˇi D 1. Clearly, this
does not account for the risk premium on asset i , and is therefore a fairly rough guide.
Apart from accounting for the risk premium, does the choice of the model of the
normal return matter a lot? Yes, but only if the model produces a higher coefficient of
determination (R2 ) than competing models. In that case, the variance of the abnormal
return is smaller for the market model which the test more precise (see Section 15.3 for
a discussion of how the variance of the abnormal return affects the variance of the test
statistic).
To illustrate the importance of the model for normal returns, consider the market
model (15.7). Under the null hypothesis that the event has no effect on the return, the
abnormal return would be just the residual in the regression (15.7). It has the variance
(assuming we know the model parameters)

Var.ui t / D Var."i t / D .1 R2 / Var.Ri t /; (15.11)

where R2 is the coefficient of determination of the regression (15.7).

Proof. (of (15.11)) From (15.7) we have (dropping the time subscripts)

Var.Ri / D ˇi2 Var.Rm / C Var."i /:

We therefore get

Var."i / D Var.Ri / ˇi2 Var.Rm /

D Var.Ri / Cov.Ri ; Rm /2 = Var.Rm /
D Var.Ri / Corr.Ri ; Rm /2 Var.Ri /
D .1 R2 / Var.Ri /:

The second equality follows from the fact that ˇi D Cov.Ri ; Rm /= Var.Rm /, the third
equality from multiplying and dividing the last term by Var.Ri / and using the definition of
the correlation, and the fourth equality from the fact that the coefficient of determination
in a simple regression equals the squared correlation of the dependent variable and the
regressor.
This variance is crucial for testing the hypothesis of no abnormal returns: the smaller

294
is the variance, the easier it is to reject a false null hypothesis (see Section 15.3). The
constant mean model has R2 D 0, so the market model could potentially give a much
smaller variance. If the market model has R2 D 0:75, then the standard deviation of
the abnormal return is only half that of the constant mean model. More realistically,
R2 might be 0.43 (or less), so the market model gives a 25% decrease in the standard
deviation, which is not a whole lot. Experience with multi-factor models also suggest that
they give relatively small improvements of the R2 compared to the market model. For
these reasons, and for reasons of convenience, the market model is still the dominating
model of normal returns.
High frequency data can be very helpful, provided the time of the event is known.
High frequency data effectively allows us to decrease the volatility of the abnormal return
since it filters out irrelevant (for the event study) shocks to the return while still capturing
the effect of the event.

15.3 Testing the Abnormal Return

In testing if the abnormal return is different from zero, there are two sources of sampling
uncertainty. First, the parameters of the normal return are uncertain. Second, even if
we knew the normal return for sure, the actual returns are random variables—and they
will always deviate from their population mean in any finite sample. The first source
of uncertainty is likely to be much smaller than the second—provided the estimation
window is much longer than the event window. This is the typical situation, so the rest of
the discussion will focus on the second source of uncertainty.
It is typically assumed that the abnormal returns are uncorrelated across time and
across assets. The first assumption is motivated by the very low autocorrelation of returns.
The second assumption makes a lot of sense if the events are not overlapping in time, so
that the event of assets i and j happen at different (calendar) times. It can also be argued
that the model for the normal return (for instance, a market model) should capture all
common movements by the regressors — leaving the abnormal returns (the residuals)
uncorrelated across firms. In contrast, if the events happen at the same time, the cross-
correlation must be handled somehow. This is, for instance, the case if the events are
macroeconomic announcements or monetary policy moves. An easy way to handle such
synchronized (clustered) events is to form portfolios of those assets that share the event

295
time—and then only use portfolios with non-overlapping events in the cross-sectional
study. For the rest of this section we assume no autocorrelation or cross correlation.
Let i2 D Var.ui t / be the variance of the abnormal return of asset i . The variance of
the cross-sectional (across the n assets) average, uN s in (15.2), is then

Var.uN s / D 12 C 22 C ::: C n2 =n2 D niD1 i2 =n2 ; (15.12)

P

since all covariances are assumed to be zero. In a large sample (where the asymptotic
normality of a sample average starts to kick in), we can therefore use a t-test since

uN s = Std.uN s / !d N.0; 1/: (15.13)

The cumulative abnormal return over q period, cari;q , can also be tested with a t-test.
Since the returns are assumed to have no autocorrelation the variance of the cari;q

Var.cari q / D qi2 : (15.14)

This variance is increasing in q since we are considering cumulative returns (not the time
average of returns).
The cross-sectional average cari;q is then (similarly to (15.12))

Var.carq / D q12 C q22 C ::: C qn2 =n2 D q niD1 i2 =n2 ; (15.15)

P

if the abnormal returns are uncorrelated across time and assets.

Figures 4.2a–b in Campbell, Lo, and MacKinlay (1997) provide a nice example of an
event study (based on the effect of earnings announcements).

Example 15.2 (Variances of abnormal returns) If the standard deviations of the daily
abnormal returns of the two firms in Example 15.1 are 1 D 0:1 and and 2 D 0:2, then
we have the following variances for the abnormal returns at different days

Time Firm 1 Firm 2 Cross-sectional Average

0:12 0:22 0:12 C 0:22 =4


1
0:12 0:22 0:12 C 0:22 =4


0
0:12 0:22 0:12 C 0:22 =4


1

296
Similarly, the variances for the cumulative abnormal returns are

Time Firm 1 Firm 2 Cross-sectional Average

0:12 0:22 0:12 C 0:22 =4


1
0 2  0:12 2  0:22 2  0:12 C 0:22 =4

1 3  0:12 3  0:22 3  0:12 C 0:22 =4

Example 15.3 (Tests of abnormal returns) By dividing the numbers in Example 15.1 by
the square root of the numbers in Example 15.2 (that is, the standard deviations) we get
the test statistics for the abnormal returns

Time Firm 1 Firm 2 Cross-sectional Average

1 2 0:5 0:4
0 10 10 13:4
1 1 1:5 1:8

Similarly, the variances for the cumulative abnormal returns we have

Time Firm 1 Firm 2 Cross-sectional Average

1 2 0:5 0:4
0 8:5 6:7 9:8
1 7:5 6:4 9:0

15.4 Quantitative Events

Some events are not easily classified as discrete variables. For instance, the effect of
positive earnings surprise is likely to depend on how large the surprise is—not just if there
was a positive surprise. This can be studied by regressing the abnormal return (typically
the cumulative abnormal return) on the value of the event (xi )

cari q D a C bxi C i : (15.16)

The slope coefficient is then a measure of how much the cumulative abnormal return
reacts to a change of one unit of xi .

297
Bibliography
Bodie, Z., A. Kane, and A. J. Marcus, 2005, Investments, McGraw-Hill, Boston, 6th edn.

Campbell, J. Y., A. W. Lo, and A. C. MacKinlay, 1997, The econometrics of financial

markets, Princeton University Press, Princeton, New Jersey.

Copeland, T. E., J. F. Weston, and K. Shastri, 2005, Financial theory and corporate policy,
Pearson Education, 4 edn.

Fama, E. F., and K. R. French, 1993, “Common risk factors in the returns on stocks and
bonds,” Journal of Financial Economics, 33, 3–56.

298
16 Kernel Density Estimation and Regression

16.1 Non-Parametric Regression

Reference: Campbell, Lo, and MacKinlay (1997) 12.3; Härdle (1990); Pagan and Ullah
(1999); Mittelhammer, Judge, and Miller (2000) 21

16.1.1 Simple Kernel Regression

Non-parametric regressions are used when we are unwilling to impose a parametric form
on the regression equation—and we have a lot of data.
Let the scalars y t and x t be related as

y t D b.x t / C " t ; " t is iid and E " t D Cov Œb.x t /; " t  D 0; (16.1)

where b./ is an unknown, possibly non-linear, function.

One possibility of estimating such a function is to approximate b.x t / by a polynomial
(or some other basis). This will give quick estimates, but the results are “global” in the
sense that the value of b.x/ at a particular x value (x D 1:9, say) will depend on all
the data points—and potentially very strongly so. The approach in this section is more
“local” by down weighting information from data points where xs is far from x t .
Suppose the sample had 3 observations (say, t D 3, 27, and 99) with exactly the same
value of x t , say 1:9. A natural way of estimating b.x/ at x D 1:9 would then be to
average over these 3 observations as we can expect average of the error terms to be close
to zero (iid and zero mean).
Unfortunately, we seldom have repeated observations of this type. Instead, we may
try to approximate the value of b.x/ (x is a single value, 1.9, say) by averaging over

299
observations where x t is close to x. The general form of this type of estimator is

O w1 .x/y1 C w2 .x/y2 C : : : C wT .x/yT

b.x/ D
w1 .x/ C w2 .x/ C : : : C wT .x/
PT
w t .x/y t
D PtD1T
; (16.2)
t D1 w t .x/

where w t .x/=˙ tTD1 w t .x/ is the weight given to observation t . The function w t .x/ is
positive and (weakly) increasing in the distance between x t and x. Note that denominator
makes the weights sum to unity. The basic assumption behind (16.2) is that the b.x/
function is smooth so local (around x) averaging makes sense.
As an example of a w.:/ function, it could give equal weight to the k values of x t
which are closest to x and zero weight to all other observations (this is the “k-nearest
neighbor” estimator, see Härdle (1990) 3.2). As another example, the weight function
could be defined so that it trades off the expected squared errors, EŒy t b.x/O 2
, and the
2O
expected squared acceleration, EŒd b.x/=dx  . This defines a cubic spline (and is often
2 2

used in macroeconomics, where x t D t and is then called the Hodrick-Prescott filter).

A Kernel regression uses a probability density function (pdf) as the weight function
w.:/.
The perhaps simplest choice is a uniform density function over x h=2 to x C h=2
(and zero outside this interval). In this case, the weighting function is
(
1 ˇˇ x t x ˇˇ  1 if q is true
w t .x/ D ı ˇ ˇ  1=2 ; where ı.q/ D (16.3)
h h 0 else.

This weighting function puts the weight 1= h on all data point in the interval x ˙ h=2 and
zero on all other data points.
However, we can gain efficiency and get a smoother (across x values) estimate by
using another density function that the uniform. In particular, using a density function
that tapers off continuously instead of suddenly dropping to zero (as the uniform density
does) improves the properties. The pdf of N.x; h2 / is commonly used as a kernel, where
the choice of h allows us to easily vary the relative weights of different observations. This
weighting function is positive so all observations get a positive weight, but the weights
are highest for observations close to x and then tapers of in a bell-shaped way.
See Figure 16.1 for an illustration.

300
 A low value of h means that the weights taper off fast—the weight function is then a
normal pdf with a low variance. With this particular kernel, we get the following weights
t a point x h i
xt x 2
exp


h
=2
w t .x/ D p : (16.4)
h 2
When h ! 0, then b.x/O evaluated at x D x t becomes just y t , so no averaging is
O
done. In contrast, as h ! 1, b.x/ becomes the sample average of y t , so we have global
averaging. Clearly, some value of h in between is needed.
O
In practice we have to estimate b.x/ at a finite number of points x. This could, for
instance, be 100 evenly spread points in the interval between the minimum and maximum
values observed in the sample. See Figure 16.2 for an illustration. Special corrections
might be needed if there are a lot of observations stacked close to the boundary of the
support of x (see Härdle (1990) 4.4).
See Figures 16.3–16.4 for an example. Note that the volatility is defined as the square
of the drift minus expected drift (from the same estimation method).
A rule of thumb value of h is

1=5 2=5
hDT j j "2=5 .xmax xmin /1=5  0:6; (16.5)

where is a from the regression y D ˛ C ˇx C x 2 C " and "2 is the variance of those
fitted residuals. In practice, replace xmax xmin by the difference between the 90th and
10th percentiles of x.
A good (but computationally intensive) approach to choose h is by the leave-one-out
cross-validation technique. This approach would, for instance, choose h to minimize the
expected (or average) prediction error
XT h i2
EPE.h/ D yt bO t .x t ; h/ =T; (16.6)
t D1

where bO t .x t ; h/ is the fitted value at x t when we use a regression function estimated on

a sample that excludes observation t, and a bandwidth h. This means that each prediction
is out-of-sample. To calculate (16.6) we clearly need to make T estimations (for each
x t )—and then repeat this for different values of h to find the minimum.
See Figure 16.5 for an example.

301
 Data and weights for b(1.7) Data and weights for b(1.9)

5 5
w t (1.7) w t (1.9)
⊗

weights

weights
1 1
yt

yt
⊗
4 4

0 0
1.5 2 2.5 1.5 2 2.5
xt xt

Data on xt : 1.5 2.0 2.5

Data and weights for b(2.1) Data on yt : 5.0 4.0 3.5

5 N(x, h2 ) kernel, h = 0.25

w t (2.1) . denotes the data
⊗ denotes the fitted b(x)
weights

1
Left y-axis: data; right y-axis: weights
yt

4 ⊗

0
1.5 2 2.5
xt

Figure 16.1: Example of kernel regression with three data points

Remark 16.1 (EPE calculations) Step 1: pick a value for h

Step 2: estimate the b.x/ function on all data, but exclude t D 1, then calculate bO 1 .x1 /
and the error y1 bO 1 .x1 /
Step 3: redo Step 2, but now exclude t D 2 and. calculate the error y2 bO 2 .x2 /. Repeat
this for t D 3; 4; :::; T . Calculate the EPE as in (16.6).
Step 4: redo Steps 2–3, but for another value of h. Keep doing this until you find the best
h (the one that gives the lowest EPE)

If the observations are independent, then it can be shown (see Härdle (1990) 4.2 and
Pagan and Ullah (1999) 3.3–6) that, with a Gaussian kernel, the estimator at point x is
asymptotically normally distributed
p h i 
1  2
.x/

O
T h b.x/ O
E b.x/ d
! N 0; p ; (16.7)
2  f .x/

302
 Kernel regression, effect of bandwidth (h)

Data
5 h = 0.25
h = 0.2

4.5
y

3.5
1.4 1.6 1.8 2 2.2 2.4
x

Figure 16.2: Example of kernel regression with three data points

where  2 .x/ is the variance of the residuals in (16.1) and f .x/ the marginal density of x.
Clearly, this means that we have (with sloppy notation)

1  2 .x/ 1
 
Ob.x/ “ !d ” N E b.x/;
O p ; (16.8)
2  f .x/ T h
To estimate the density function needed in (16.7), we can use a kernel density estima-
tor of the pdf at some point x
 
O 1 XT xt x
f .x/ D K , where (16.9)
T hx t D1 hx
exp u2 =2


K.u/ D p :
2

The value hx D Std.x t /1:06T 1=5 is sometimes recommended for estimating the density
function, since it can be shown to be the optimal choice (in MSE sense) if data is normally
distributed and the N.0; 1/ kernel is used. (Clearly, using K Œ.x t x/= hx  = hx is the
same as using pdf of N.x; h2x /.) Notice that hx need not be the same as the bandwidth
(h) used in the kernel regression.
See Figure 16.6 for an example where the width of the confidence band varies across
x values—mostly because the sample contains few observations close to some x values.

303
(However, the assumption of independent observations can be questioned in this case.)
To estimate the function  2 .x/ in (16.7), we use a non-parametric regression of the
squared fitted residuals on x t

"O2t D  2 .x t /, where "O t D y t O t /;

b.x (16.10)

O t / are the fitted values from the non-parametric regression (16.1). Notice that
where b.x
this approach allows the variance to depend on the x value.

Example 16.2 Suppose the sample has three data points Œx1 ; x2 ; x3  D Œ1:5; 2; 2:5 and
Œy1 ; y2 ; y3  D Œ5; 4; 3:5. Consider the estimation of b.x/ at x D 1:9. With h D 1, the
numerator in (16.4) is
XT   p
.1:5 1:9/2 =2 .2 1:9/2 =2 .2:5 1:9/2 =2
w t .x/y t D e 5Ce 4Ce  3:5 = 2
t D1
p
 .0:92  5 C 1:0  4 C 0:84  3:5/ = 2
p
D 11:52= 2:

The denominator is
XT   p
.1:5 1:9/2 =2 .2 1:9/2 =2 .2:5 1:9/2 =2
w t .x/ D e Ce Ce = 2
t D1
p
 2:75= 2:

The estimate at x D 1:9 is therefore

O
b.1:9/  11:52=2:75  4:19:

Kernel regressions are typically consistent, provided longer samples are accompanied
by smaller values of h, so the weighting function becomes more and more local as the
sample size increases.

16.1.2 Multivariate Kernel Regression

Suppose that y t depends on two variables (x t and z t )

y t D b.x t ; z t / C " t ; " t is iid and E " t D 0: (16.11)

304
 Return vs lagged return, in bins
6
Daily S&P 500 returns 1979:1-2013:4
4

2
Return

−2

−4
−10 −5 0 5 10
Lagged return

Figure 16.3: Crude non-parametric regression

Return vs lagged return, kernel regression

5
Daily S&P 500 returns 1979:1-2013:4
Return

Small bandwidth
Optimal bandwitdh
Large bandwidth
−5
−10 −5 0 5 10
Lagged return

Figure 16.4: Non-parametric regression, importance of bandwidth

This makes the estimation problem much harder since there are typically few observations
in every bivariate bin (rectangle) of x and z. For instance, with as little as a 20 intervals of
each of x and z, we get 400 bins, so we need a large sample to have a reasonable number

305
 Cross validation simulations, kernel regression

Avg prediction error, relative to min 1.01

Daily S&P 500 returns 1979:1-2013:4

1.008

1.006

1.004

1.002

1
0.5 1 1.5 2 2.5
Bandwidth

Figure 16.5: Cross-validation

Return vs lagged return, kernel regression with 90% conf band

5
Daily S&P 500 returns 1979:1-2013:4
Return

−5
−10 −5 0 5 10
Lagged return

Figure 16.6: Non-parametric regression with confidence bands

of observations in every bin.

306
 Fitted return as a function of 2 lags of returns

−2

−4
Daily S&P 500 returns 1979:1-2013:4
−6 −10
−5
−10 0
−5
0 5
5 10
10
Return lagged twice
Return lagged once

Figure 16.7: Non-parametric regression with two regressors

In any case, the most common way to implement the kernel regressor is to let
PT
O z/ D Pt D1 w t .x/w t .z/y t ;
b.x; (16.12)
T
tD1 w t .x/w t .z/

where w t .x/ and w t .z/ are two kernels like in (16.4) and where we may allow the band-
width (h) to be different for x t and z t (and depend on the variance of x t and y t ). In this
case. the weight of the observation (x t ; z t ) is proportional to w t .x/w t .z/, which is high
if both x t and z t are close to x and z respectively.
See Figure 16.7 for an example.

16.2 Examples of Non-Parametric Estimation

16.2.1 A Model for the Short Interest Rate

Interest rate models are typically designed to describe the movements of the entire yield
curve in terms of a small number of factors. For instance, the model assumes that the

307
 Drift vs level, in bins Volatility vs level, in bins
0.1 1.5
∆ interest rate

Volatility
0 1

−0.1 0.5

−0.2 0
0 5 10 15 20 0 5 10 15 20
Interest rate Interest rate
Daily federal funds rates 1954:7-2013:4 Volatility = (actual − fitted ∆ interest rate)2

Figure 16.8: Crude non-parametric estimation

(de-meaned) short interest rate, r t , is a mean-reverting AR(1) process

r t D r t 1 C " t ; where " t  N.0;  2 /, so (16.13)

rt rt 1 D . 1/r t 1 C "t ; (16.14)

and that all term premia are constant. This means that the drift is decreasing in the interest
rate, but that the volatility is constant. For instance, if  D 0:95 (a very peristent interest
rate), then (16.14) is
r t r t 1 D 0:05r t 1 C " t ; (16.15)

so the reversion to the mean (here zero) is very slow.

(The usual assumption is that the short interest rate follows an Ornstein-Uhlenbeck
diffusion process, which implies the discrete time model in (16.13).) It can then be shown
that all interest rates (for different maturities) are linear functions of short interest rates.
To capture more movements in the yield curve, models with richer dynamics are used.
For instance, Cox, Ingersoll, and Ross (1985) construct a model which implies that the
short interest rate follows an AR(1) as in (16.13) except that the variance is proportional
to the interest rate level, so " t  N.0; r t 1  2 /.
Non-parametric methods have been used to estimate how the drift and volatility are
related to the interest rate level (see, for instance, Ait-Sahalia (1996)). Figures 16.8–
16.11 give an example. Note that the volatility is defined as the square of the drift minus
expected drift (from the same estimation method).

308
 Drift vs level, kernel regression Vol vs level, kernel regression
0.1 1.5
∆ interest rate

Volatility
0 1

−0.1 0.5
Optimal h
Larger h(4×)
−0.2 0
0 5 10 15 20 0 5 10 15 20
Interest rate Interest rate
Daily federal funds rates 1954:7-2013:4 Volatility = (actual − fitted ∆ interest rate)2

Figure 16.9: Kernel regression, importance of bandwidth

Cross validation simulations, kernel regression

Avg prediction error, relative to min

1.0007 Daily federal funds rates 1954:7-2013:4

1.0006

1.0005

1.0004

1.0003

1.0002

1.0001

1
0.4 0.5 0.6 0.7 0.8 0.9 1
h

Figure 16.10: Cross-validation

16.2.2 Non-Parametric Option Pricing

There seems to be systematic deviations from the Black-Scholes model. For instance,
implied volatilities are often higher for options far from the current spot (or forward)
price—the volatility smile. This is sometimes interpreted as if the beliefs about the future
log asset price put larger probabilities on very large movements than what is compatible
with the normal distribution (“fat tails”).

309
 Drift vs level, kernel regression Vol vs level, kernel regression
0.1 1.5
∆ interest rate

Volatility
0 1

−0.1 0.5
Point estimate and
90% confidence band
−0.2 0
0 5 10 15 20 0 5 10 15 20
Interest rate Interest rate
Volatility = (actual − fitted ∆ interest rate)2

Daily federal funds rates 1954:7-2013:4

The bandwith is from cross-validation

Figure 16.11: Kernel regression, confidence band

This has spurred many efforts to both describe the distribution of the underlying asset
price and to amend the Black-Scholes formula by adding various adjustment terms. One
strand of this literature uses non-parametric regressions to fit observed option prices to
the variables that also show up in the Black-Scholes formula (spot price of underlying
asset, strike price, time to expiry, interest rate, and dividends). For instance, Ait-Sahalia
and Lo (1998) applies this to daily data for Jan 1993 to Dec 1993 on S&P 500 index
options (14,000 observations). They find interesting patterns of the implied moments
(mean, volatility, skewness, and kurtosis) as the time to expiry changes. In particular, the
non-parametric estimates suggest that distributions for longer horizons have increasingly
larger skewness and kurtosis. Whereas the distributions for short horizons are not too
different from normal distributions, this is not true for longer horizons.

Bibliography
Ait-Sahalia, Y., 1996, “Testing Continuous-Time Models of the Spot Interest Rate,” Re-
view of Financial Studies, 9, 385–426.

Ait-Sahalia, Y., and A. W. Lo, 1998, “Nonparametric Estimation of State-Price Densities

Implicit in Financial Asset Prices,” Journal of Finance, 53, 499–547.

310
Campbell, J. Y., A. W. Lo, and A. C. MacKinlay, 1997, The Econometrics of Financial
Markets, Princeton University Press, Princeton, New Jersey.

Cox, J. C., J. E. Ingersoll, and S. A. Ross, 1985, “A Theory of the Term Structure of
Interest Rates,” Econometrica, 53, 385–407.

Härdle, W., 1990, Applied Nonparametric Regression, Cambridge University Press, Cam-
bridge.

Mittelhammer, R. C., G. J. Judge, and D. J. Miller, 2000, Econometric Foundations, Cam-

bridge University Press, Cambridge.

Pagan, A., and A. Ullah, 1999, Nonparametric Econometrics, Cambridge University

Press.

311
17 Simulating the Finite Sample Properties
Reference: Greene (2000) 5.3 and Horowitz (2001)
Additional references: Cochrane (2001) 15.2; Davidson and MacKinnon (1993) 21; Davi-
son and Hinkley (1997); Efron and Tibshirani (1993) (bootstrapping, chap 9 in particular);
and Berkowitz and Kilian (2000) (bootstrapping in time series models)
We know the small sample properties of regression coefficients in linear models with
fixed regressors and iid normal error terms. Monte Carlo simulations and bootstrapping
are two common techniques used to understand the small sample properties when these
conditions are not satisfied.
How they should be implemented depends crucially on the properties of the model
and data: if the residuals are autocorrelated, heteroskedastic, or perhaps correlated across
regressions equations. These notes summarize a few typical cases.
The need for using Monte Carlos or bootstraps varies across applications and data
sets. For a case where it is not needed, see Figure 17.1.

alpha t LS t NW t boot
US industry portfolios, 1970:1-2012:12 all NaN NaN NaN NaN
15 A (NoDur) 3.62 2.72 2.71 2.70
Mean excess return

B (Durbl) -1.21 -0.58 -0.59 -0.59

C (Manuf ) 0.70 0.72 0.71 0.69
10 D (Enrgy) 4.06 1.80 1.81 1.81
D
A E (HiTec) -1.82 -1.00 -1.00 -0.98
GC
I FH JB E F (Telcm) 1.82 1.07 1.06 1.05
5 G (Shops) 1.37 0.94 0.94 0.94
H (Hlth ) 2.13 1.22 1.24 1.24
0 I (Utils) 2.87 1.61 1.58 1.56
0 0.5 1 1.5 J (Other) -0.65 -0.61 -0.60 -0.61
β (against the market)
NW uses 1 lag
The bootstrap samples pairs of (yt , xt )
3000 simulations

Figure 17.1: CAPM, US industry portfolios, different t-stats

312
17.1 Monte Carlo Simulations

17.1.1 Monte Carlo Simulations in the Simplest Case

Monte Carlo simulations is essentially a way to generate many artificial (small) samples
from a parameterized model and then estimating the statistic on each of those samples.
The distribution of the statistic is then used as the small sample distribution of the estima-
tor.
The following is an example of how Monte Carlo simulations could be done in the
special case of a linear model with a scalar dependent variable

y t D x t0 ˇ C u t ; (17.1)

where u t is iid N.0;  2 / and x t is stochastic but independent of u t ˙s for all s. This means
that x t cannot include lags of y t .
Suppose we want to find the small sample distribution of a function of the estimate,
O
g.ˇ/. To do a Monte Carlo experiment, we need information on (i) the coefficients ˇ; (ii)
the variance of u t ;  2 ; (iii) and a process for x t .
The process for x t is typically estimated from the data on x t (for instance, a VAR
system x t D A1 x t 1 C A2 x t 2 C e t ). Alternatively, we could simply use the actual
sample of x t and repeat it.
The values of ˇ and  2 are often a mix of estimation results and theory. In some
case, we simply take the point estimates. In other cases, we adjust the point estimates
so that g.ˇ/ D 0 holds, that is, so you simulate the model under the null hypothesis
in order to study the size of asymptotic tests and to find valid critical values for small
samples. Alternatively, you may simulate the model under an alternative hypothesis in
order to study the power of the test using either critical values from either the asymptotic
distribution or from a (perhaps simulated) small sample distribution.
To make it a bit concrete, suppose you want to use these simulations to get a 5%
critical value for testing the null hypothesis g.ˇ/ D 0. The Monte Carlo experiment
follows these steps.

1. Construct an artificial sample of the regressors (see above), xQ t for t D 1; : : : ; T .

Draw random numbers uQ t for t D 1; : : : ; T and use those together with the artificial

313
 sample of xQ t to calculate an artificial sample yQ t for t D 1; : : : ; T from

yQ t D xQ t0 ˇ C uQ t ; (17.2)

by using the prespecified values of the coefficients ˇ.

2. Calculate an estimate ˇO and record it along with the value of g.ˇ/

O and perhaps also
the test statistic of the hypothesis that g.ˇ/ D 0.

3. Repeat the previous steps N (3000, say) times. The more times you repeat, the
better is the approximation of the small sample distribution.

4. Sort your simulated ˇ, O g.ˇ/,

O and the test statistic in ascending order. For a one-
sided test (for instance, a chi-square test), take the (0:95N )th observations in these
sorted vector as your 5% critical values. For a two-sided test (for instance, a t-
test), take the (0:025N )th and (0:975N )th observations as the 5% critical values.
You may also record how many times the 5% critical values from the asymptotic
distribution would reject a true null hypothesis.

O g.ˇ/,
5. You may also want to plot a histogram of ˇ, O and the test statistic to see if there
is a small sample bias, and how the distribution looks like. Is it close to normal?
How wide is it?

See Figures 17.2–17.3 for an example.

We have the same basic procedure when y t is a vector, except that we might have
to consider correlations across the elements of the vector of residuals u t . For instance,
we might want to generate the vector uQ t from a N.0; ˙/ distribution—where ˙ is the
variance-covariance matrix of u t .

Remark 17.1 (Generating N.; ˙/ random numbers) Suppose you want to draw an n1
vector " t of N.; ˙/ variables. Use the Cholesky decomposition to calculate the lower
triangular P such that ˙ D PP 0 (note that Gauss and MatLab returns P 0 instead of
P ). Draw u t from an N.0; I / distribution (randn in MatLab, rndn in Gauss), and define
" t D  C P u t . Note that Cov." t / D E P u t u0t P 0 D PIP 0 D ˙ .

314
 Average LS estimate of ρ Std of LS estimate of ρ

simulation
0.9 asymptotic
0.1
0.85
0.8
0.05
0.75
0.7 0
0 100 200 300 400 500 0 100 200 300 400 500
Sample size, T Sample size, T

√
T × Std of LS estimate of ρ True model: yt = 0.9yt−1 + ǫt ,
0.8 where ǫt is iid N(0,2)

0.7 Estimated model: yt = a + ρyt−1 + ut

0.6 Number of simulations: 25000

0.5

0.4
0 100 200 300 400 500
Sample size, T

Figure 17.2: Results from a Monte Carlo experiment of LS estimation of the AR coeffi-
cient.

17.1.2 Monte Carlo Simulations with more Complicated Errors

It is straightforward to sample the errors from other distributions than the normal, for in-
stance, a student-t distribution. Equipped with uniformly distributed random numbers,
you can always (numerically) invert the cumulative distribution function (cdf) of any
distribution to generate random variables from any distribution by using the probability
transformation method. See Figure 17.4 for an example.

Remark 17.2 Let X  U.0; 1/ and consider the transformation Y D F 1 .X/, where
F 1 ./ is the inverse of a strictly increasing cumulative distribution function F , then Y
has the cdf F .

Example 17.3 The exponential cdf is x D 1 exp. y/ with inverse y D ln .1 x/ = .

Draw x from U.0:1/ and transform to y to get an exponentially distributed variable.

315
 Distribution of LS estimator, T = 25 Distribution of LS estimator, T = 100
Mean and std: Mean and std:
0.1 0.74 0.16 0.1 0.86 0.06

0.05 0.05

0 0
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

True model: yt = 0.9yt−1 + ǫt , ǫt is iid N(0,2)

Estimated model: yt = a + ρyt−1 + ut
Number of simulations: 25000

Figure 17.3: Results from a Monte Carlo experiment of LS estimation of the AR coeffi-
cient.

It is more difficult to handle non-iid errors, like those with autocorrelation and het-
eroskedasticity. We then need to model the error process and generate the errors from that
model.
If the errors are autocorrelated, then we could estimate that process from the fitted
errors and then generate artificial samples of errors (here by an AR(2))

uQ t D a1 uQ t 1 C a2 uQ t 2 C "Q t : (17.3)

Alternatively, heteroskedastic errors can be generated by, for instance, a GARCH(1,1)

model
u t  N.0;  t2 /, where  t2 D ! C ˛u2t 1 C ˇ t2 1 : (17.4)

However, this specification does not account for any link between the volatility and the
regressors (squared)—as tested for by White’s test. This would invalidate the usual OLS
standard errors and therefore deserves to be taken seriously. A simple, but crude, approach
is to generate residuals from a N.0;  t2 ) process, but where  t2 is approximated by the
fitted values from
"2t D c 0 w t C  t ; (17.5)

where w t include the squares and cross product of all the regressors.

316
 Distribution of LS t-stat, T = 5 Distribution of LS t-stat, T = 100
t = (b̂ − 0.9)/Std(b̂)
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−4 −2 0 2 4 −4 −2 0 2 4

Model: Rt = 0.9ft + ǫt , ǫt = vt − 2,
where vt has a χ22 distribution Probability density functions

Estimated model: yt = a + bft + ut N(0,1)

0.4
Number of simulations: 25000 χ22 − 2
0.3
T = 5 T = 100 0.2
Kurtosis of t-stat: 46.753 3.049
Frequency of |t-stat| > 1.65 0.294 0.105 0.1
Frequency of |t-stat| > 1.96 0.227 0.054 0
−4 −2 0 2 4

Figure 17.4: Results from a Monte Carlo experiment with thick-tailed errors.

17.2 Bootstrapping

17.2.1 Bootstrapping in the Simplest Case

Bootstrapping is another way to do simulations, where we construct artificial samples by

sampling from the actual data. The advantage of the bootstrap is then that we do not
have to try to estimate the process of the errors and regressors (as we do in a Monte Carlo
experiment). The real benefit of this is that we do not have to make any strong assumption
about the distribution of the errors.
The bootstrap approach works particularly well when the errors are iid and indepen-
dent of x t s for all s. This means that x t cannot include lags of y t . We here consider
bootstrapping the linear model (17.1), for which we have point estimates (perhaps from
LS) and fitted residuals. The procedure is similar to the Monte Carlo approach, except
that the artificial sample is generated differently. In particular, Step 1 in the Monte Carlo
simulation is replaced by the following:

317
 1. Construct an artificial sample yQ t for t D 1; : : : ; T by

yQ t D x t0 ˇ C uQ t ; (17.6)

where uQ t is drawn (with replacement) from the fitted residual and where ˇ is the
point estimate.

Example 17.4 With T D 3, the artificial sample could be

2 3 2 0 3
.yQ1 ; xQ 1 / .x1 ˇ0 C u2 ; x1 /
4 .yQ2 ; xQ 2 / 5 D 4 .x20 ˇ0 C u1 ; x2 / 5 :
6 7 6 7

.yQ3 ; xQ 3 / .x30 ˇ0 C u2 ; x3 /

The approach in (17.6) works also when y t is a vector of dependent variables—and

will then help retain the cross-sectional correlation of the residuals.

17.2.2 Bootstrapping when Errors Are Heteroskedastic

Suppose now that the errors are heteroskedastic, but serially uncorrelated. If the het-
eroskedasticity is unrelated to the regressors, then we can still use (17.6).
On contrast, if the heteroskedasticity is related to the regressors, then the traditional LS
covariance matrix is not correct (this is the case that White’s test for heteroskedasticity
tries to identify). It would then be wrong to pair x t with just any uQ t D us since that
destroys the relation between x t and the variance of the residual.
An alternative way of bootstrapping can then be used: generate the artificial sample
by drawing (with replacement) pairs .ys ; xs /, that is, we let the artificial pair in t be
.yQ t ; xQ t / D .xs0 ˇ0 C us ; xs / for some random draw of s so we are always pairing the
residual, us , with the contemporaneous regressors, xs . Note that we are always sampling
with replacement—otherwise the approach of drawing pairs would be to just re-create the
original data set.
This approach works also when y t is a vector of dependent variables.

Example 17.5 With T D 3, the artificial sample could be

2 3 2 0 3
.yQ1 ; xQ 1 / .x2 ˇ0 C u2 ; x2 /
4 .yQ2 ; xQ 2 / 5 D 4 .x30 ˇ0 C u3 ; x3 / 5
6 7 6 7

.yQ3 ; xQ 3 / .x30 ˇ0 C u3 ; x3 /

318
 It could be argued (see, for instance, Davidson and MacKinnon (1993)) that bootstrap-
ping the pairs .ys ; xs / makes little sense when xs contains lags of ys , since the random
sampling of the pair .ys ; xs / destroys the autocorrelation pattern on the regressors.

17.2.3 Autocorrelated Errors

It is quite hard to handle the case when the errors are serially dependent, since we must
the sample in such a way that we do not destroy the autocorrelation structure of the data.
A common approach is to fit a model for the residuals, for instance, an AR(1), and then
bootstrap the (hopefully iid) innovations to that process.
Another approach amounts to resampling blocks of data. For instance, suppose the
sample has 10 observations, and we decide to create blocks of 3 observations. The first
block is .uO 1 ; uO 2 ; uO 3 /, the second block is .uO 2 ; uO 3 ; uO 4 /, and so forth until the last block,
.uO 8 ; uO 9 ; uO 10 /. If we need a sample of length 3 , say, then we simply draw  of those
block randomly (with replacement) and stack them to form a longer series. To handle
end point effects (so that all data points have the same probability to be drawn), we also
create blocks by “wrapping” the data around a circle. In practice, this means that we add
a the following blocks: .uO 10 ; uO 1 ; uO 2 / and .uO 9 ; uO 10 ; uO 1 /. The length of the blocks should
clearly depend on the degree of autocorrelation, but T 1=3 is sometimes recommended as
a rough guide. An alternative approach is to have non-overlapping blocks. See Berkowitz
and Kilian (2000) for some other approaches.
See Figures 17.5–17.6 for an illustration.

Bibliography
Berkowitz, J., and L. Kilian, 2000, “Recent developments in bootstrapping time series,”
Econometric-Reviews, 19, 1–48.

Cochrane, J. H., 2001, Asset pricing, Princeton University Press, Princeton, New Jersey.

Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,

Oxford University Press, Oxford.

Davison, A. C., and D. V. Hinkley, 1997, Bootstrap methods and their applications, Cam-
bridge University Press.

319
 Std of LS slope under autocorrelation Std of LS slope under autocorrelation
0.1 κ = −0.9 0.1 κ = 0
OLS formula
Newey-West
Simulated
Bootstrapped
0.05 0.05

0 0
−0.5 0 0.5 −0.5 0 0.5
ρ (autocorrelation of residual) ρ (autocorrelation of residual)

Model: yt = 0.9xt + ǫt ,
Std of LS slope under autocorrelation where ǫt = ρǫt−1 + ξt , ξt is iid N
0.1 κ = 0.9 xt = κxt−1 + ηt , ηt is iid N

ut is the residual from LS estimate of

yt = a + bxt + ut
0.05
NW uses 15 lags
The block bootstrap uses blocks of size 20
Number of simulations: 25000
0
−0.5 0 0.5
ρ (autocorrelation of residual)

Figure 17.5: Standard error of OLS estimator, autocorrelated errors

Efron, B., and R. J. Tibshirani, 1993, An introduction to the bootstrap, Chapman and Hall,
New York.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Horowitz, J. L., 2001, “The Bootstrap,” in J.J. Heckman, and E. Leamer (ed.), Handbook
of Econometrics . , vol. 5, Elsevier.

320
 Std of LS intercept under autocorrelation Std of LS intercept under autocorrelation
0.1 κ = −0.9 0.1 κ = 0

0.05 0.05
OLS formula
Newey-West
Simulated
Bootstrapped
0 0
−0.5 0 0.5 −0.5 0 0.5
ρ (autocorrelation of residual) ρ (autocorrelation of residual)

Model: yt = 0.9xt + ǫt ,
Std of LS intercept under autocorrelation where ǫt = ρǫt−1 + ξt , ξt is iid N
0.1 κ = 0.9 xt = κxt−1 + ηt , ηt is iid N

ut is the residual from LS estimate of

yt = a + bxt + ut
0.05
NW uses 15 lags
The block bootstrap uses blocks of size 20
Number of simulations: 25000
0
−0.5 0 0.5
ρ (autocorrelation of residual)

Figure 17.6: Standard error of OLS estimator, autocorrelated errors

321
18 Panel Data
References: Verbeek (2012) 10 and Baltagi (2008)

18.1 Introduction to Panel Data

A panel data set has data on a cross-section (i D 1; 2; : : : ; N , individuals or firms) over

many time periods (t D 1; 2; : : : ; T ). The aim is to estimate a linear relation between the
dependent variable and the regressors

yi t D xi0 t ˇ C "i t : (18.1)

For instance, data on the dependent variable might have this structure
2 3
i D 1 i D 2


i D N
6 t D 1 W y11
6 7
6 y21 yN1 7 7
(18.2)
6t D 2 W y y y 7
6 12 22 N2 7
6 : 7
6 :: 7
4 5
t D T W y1T y2T yN T

The structure for each of the regressors is similar.

The most basic estimation approach is to just run LS (on all observations “stacked”).
This is not that bad (although GLS might be more efficient), especially since there is
typically lots of data points. However, we may want to allow for individual (i) effects.

18.2 Fixed Effects Model

In the fixed effects model, we allow for different individual intercepts

yi t D i C xi0 t ˇ C "i t ; "i t is iidN.0; "2 /; (18.3)

and maintain the basic assumption that "jt is uncorrelated with xi t (across all i and j ).

322
 There are several ways to estimate this model. The conceptually most straightforward
is to include individual dummies (N ) where dummy i takes the value of one if the data
refers to individual i and zero otherwise. (Clearly, the regression can then not include any
intercept. Alternatively, include an intercept but only N 1 dummies—for i D 2 N .)
However, this approach can be difficult to implement since it may involve a very large
number of regressors.
As an alternative (that gives the same point estimates as OLS with dummies) consider
the following approach. First, take average across time (for a given i ) of yi t and xi t
in (18.3. That is, think (but d not run any estimation...) of forming the cross-sectional
regression

yNi D i C xN i0 ˇ C "Ni , where (18.4)

1 PT 1 PT
yNi D t D1 yi t and x Ni D xi t : (18.5)
T T t D1
Second, subtract from (18.3) to get

yi t yNi D .xi t xN i /0 ˇ C ."i t "Ni / : (18.6)

At this stage, estimate ˇ by running LS on all observations of (18.6) “stacked.” We denote

this estimate ˇOFE (FE stands for fixed effects) and it is also often called the within estima-
tor. The interpretation of this approach is that we estimate the slope coefficients by using
the movements around individual means (not how the individual means differ). Notice
that it gives the same results as OLS with dummies. Third and finally, get estimates of
individual intercepts as
i D yNi xN i0 ˇOFE : (18.7)

Clearly, the within estimator wipes out all regressors that are constant across time for a
given individual (say, gender and schooling) : they are effectly merged with the individual
means (i ).
We can apply the usual tests (possibly, small-sample adjustment of standard errors).
In particular, we can estimate the standard error of the residual as
1 PT PN 2
u2 D uO , where (18.8)
T N t D1 i D1 i t
uO 2it D yi t D O i xi0 t ˇOFE ;

323
and the covariance matrix of the slope coefficients as

Var.ˇOFE / D  2 Sxx1 where Sxx D

PT PN
t D1 i D1 .xi t xN i /.xi t xN i /0 : (18.9)

Notice that these results (on the variance of the slope coefficients) rely on the assumption
that the residuals are uncorrelated across time and individuals.

Example 18.1 N D 2; T D 2. If we stack data for t D T 1 (i D 1 and N ) first and

for t D T second, then we have the following covariance matrix of the residuals ui t
0 1 2 3
u1;T 1 u2 0 0 0
B uN;T 1 C 6 0 u2 0 0 7
B C 6 7
Cov B
B u
CD6
C 6 0 0 2 0 7 :
7
@ 1T A 4 u 5
2
uN T 0 0 0 u

This is a diagonal matrix.

Remark 18.2 (Lagged dependent variable as regressor.) If yi;t 1 is among the regres-
sors xi t , then the within estimator (18.6) is biased in small samples (that is, when T is
small)—and increasing the cross-section (that is, N ) does not help. To see the problem,
suppose that the lagged dependent variable is the only regressor (xi t D yi;t 1 ). The
within estimator (18.6) is then
PT  PT   PT 
yi t t D1 yi t =T D yi;t 1 t D2 yi;t 1 =.T 1/ ˇ C "i t t D1 "i t =T :

The problem is that yi;t 1 is correlated with TtD1 "i t =T since the latter contains "i;t 1
P

which affects yi;t 1 directly. In addition, TtD2 yi;t 1 =T contains yi;t which is correlated
P

with "i t . It can be shown that this bias can be substantial for panels with small T .

An another way of estimating the fixed effects model is to difference away the i by
taking first-differences (in time)

yi t yi;t 1 D .xi t xi;t 1/

0
ˇ C "i t "i;t 1 : (18.10)
„ ƒ‚ …
residual uit

This can be estimated by OLS, but we could adjust the covariance matrix of the slope
coefficients, since the residuals are now (negatively) autocorrelated ("i;t 1 shows up both

324
 LS Fixed eff Between GLS
exper/100 7:84 4:11 10:64 4:57
.8:25/ .6:21/ .4:05/ .7:12/
exper2/100 0:20 0:04 0:32 0:06
. 5:04/ . 1:50/ . 2:83/ . 2:37/
tenure/100 1:21 1:39 1:25 1:38
.2:47/ .4:25/ .0:90/ .4:32/
tenure2/100 0:02 0:09 0:02 0:07
. 0:85/ . 4:36/ . 0:20/ . 3:77/
south 0:20 0:02 0:20 0:13
. 13:51/ . 0:45/ . 6:67/ . 5:70/
union 0:11 0:06 0:12 0:07
.6:72/ .4:47/ .3:09/ .5:57/

Table 18.1: Panel estimation of log wages for women, T D 5 and N D 716, from NLS
(1982,1983,1985,1987,1988). Example of fixed and random effects models, Hill et al
(2008), Table 15.9. Numbers in parentheses are t-stats.

in t and t 1, but with different signs). Becuase of the negative autocorrelation, un-
adjusted standard errors are likely to overstate the uncertainty—and carn therefore be
used as a conservative approach. Notice that the first-difference approach focuses on
how changes in the regressors (over time, for the same individual) affect changes in the
dependent variable. Also this method wipes out all regressors that are constant across
time (for a given individual).

Example 18.3 N D 2; T D 2. Stack the data for individual i D 1 first and those for
individual i D N second
0 1 0 1 2 3
u1;T 1 "1;T 1 "1;T 2 2"2 "2 0 0
B "1;T "1;T 1 C 6 "2 2"2
B C B C 6 7
B u1T C 0 0
Cov B C D Cov B
7
CD6 7:
B u B " 2 27
@ N;T 1 A
C
@ N;T 1 " C 6 0
N;T 2 A 4 0 2 "  "5
2 2
uN T "N;T "N;T 1 0 0 " 2"

Remark 18.4 (Difference-in-difference estimator) Suppose there are only two periods
(T D 2) and that one of the regerssors is a dummy that equals one for a some individ-
uals who got a “treatment” (say, extra schooling) between the two periods and zero for
the other individuals. Running the first-difference method (18.10) and studying the coef-
ficient of that dummy variable is then the so called “difference-in-difference” method. It

325
measures how much the dependent variable changed for those with treatment compared
to the change for those without the treatment.

Remark 18.5 (Lagged dependent variable as regressor) If yi;t 1 is among the regressors
xi t , then the first-difference method (18.10) does not work (OLS is inconsistent). The
reason is that the (autocorrelated) residual is then correlated with the lagged dependent
variable. This model cannot be estimated by OLS (the instrumental variable method might
work).

18.3 Random Effects Model

The random effects model allows for random individual “intercepts” (i )

yi t D ˇ0 C xi0 t ˇ C i C "i t ; where (18.11)

"i t is iidN.0; "2 / and i is iidN.0; 2 /: (18.12)

Notice that i is random (across agents) but constant across time, while "i t is just random
noise. Hence, i can be interpreted as the permanent “luck” of individual i . For instance,
suppose the panel is drawn for a large sample so as to be representative. This means,
effectively, that the sample contains values of (yi t ; xi t ) that match those of the population.
An example could be that one of the xi t variables measure age of individuals—and the
sample is drawn so that it has the same agre distribution as the population. In this setting,
a random i makes sense as a proxy for whatever information we leave out. Clearly, if
the we regard i as non-random, then we are back in the fixed-effects model. (The choice
between the two models is not always easy, so it may be wise to try both—and compare
the results.)
We could therefore write the regression as

yi t D ˇ0 C xi0 t ˇ C ui t , where ui t D i C "i t ; (18.13)

and we typically assume that ui t is uncorrelated across individuals, but correlated across
time (only because of i ). In addition, we assume that "jt and i are not correlated with
each other or with xi t .
There are several ways to estimate the random effects model. First, the methods for
fixed effects (the within and first-difference estimators) all work—so the “fixed effect”

326
can actually be a random effect. Second, the between estimator using only individual
time averages
yNi D ˇ0 C xN i0 ˇ C i C "Ni ; (18.14)
„ ƒ‚ …
residuali

is also consistent, but discards all time-series information. Third, LS on

yi t D ˇ0 C xi0 t ˇ C i C "i t (18.15)

„ ƒ‚ …
residualit

O
is consistent (but not really efficient). However, in this case we may need to adjust Cov.ˇ/
since the covariance matrix of the residuals is not diagonal.
Example 18.6 N D 2; T D 2. If we stack the data for individual i D 1 first and those
for individual i D N second
0 1 2 3
u1;T 1 2 C "2 2 0 0
B u1T C 6 2 2
C 2
B C 6 7
 " 0 0
Cov B
7
CD6 7;
B u C 6 0
@ N;T 1 A 4 0 2 C "2 2 7 5
2 2 2
uN T 0 0   C "

which has elements off the main diagonal.

Remark 18.7 (Generalized least squares ) GLS is an alternative estimation method that
exploits correlation structure of residuals to increase the efficiency. In this case, it can be
implemented by running OLS on

yi t # yNi D ˇ0 .1 #/ C .xi t # xN i /0 ˇ C i t , where

q
# D1 u2 =.u2 C T 2 /:

In this equation, u2 is the variance of the residuals in the “within regression” as estimated
in (18.8) and 2 D B2 u2 =T , where B2 is the variance of the residuals in the “between
regression” (18.14).Here, 2 can be interpreted as the variance of the random effect i .
However, watch out for negative values of 2 and notice that when #  1, then GLS is
similar to the “within estimator” from (18.6). This happens when 2 >> u2 or when T
is large. The intuition is that when 2 is large, then it is smart to get rid of that source
of noise by using the within estimator, which disregards the information in the differences
between individual means.

327
 In the random effects model, the i variable can be thought of as an excluded variable.
Excluded variables typically give a bias in the coefficients of all included variables—
unless the excluded variable is uncorrelated with all of them. This is the assumption in
the random effects model (recall: we assumed that i is uncorrelated with xjt ). If this
assumption is wrong, then we cannot estimate the RE model by either OLS or GLS, but
the within-estimator (cf. the FE model) works, since it effectively eliminates the excluded
variable from the system.

Bibliography
Baltagi, D. H., 2008, Econometric Analysis of Panel Data, Wiley, 4th edn.

Verbeek, M., 2012, A guide to modern econometrics, Wiley, 4th edn.

328
19 Binary Choice Models
Reference: Verbeek (2012) 7

19.1 Binary Choice Model

19.1.1 Basic Model Setup

A binary variable (
0 firm i doesn’t pay dividends
yi D (19.1)
1 firm i pays dividends
We know a few things about firm i: xi (industry, size, profitability...)
Model: the probability that firm i pays dividends is some function of xi

Pr.yi D 1jxi / D F .xi0 ˇ/ (19.2)

Idea: if xi contains profitability, then (presumably) most firms with high profits will have
dividends. What you estimate is (effectively) how the typical pattern changes with profits.
What function F ./ should we use in (19.2)? Mostly a matter of convenience. A
probit model assumed that F ./ is a standard normal cumulative distribution function, see
Figure 19.1. Other choices of F ./ give logit model (F ./ is a logistic function) or linear
probability model (F .xi0 ˇ/ D xi0 ˇ). See Figure 19.2 for an illustration.
How to interpret the results? Mostly by looking at the marginal effects

@F .xi0 ˇ/
(19.3)
@xi
For instance, how does the probability of having dividends change when profits changes?

Example 19.1 Constant plus two more regressors (w and z): xi0 ˇ D ˇ0 C ˇ1 wi C ˇ2 zi ,
then
@F .xi0 ˇ/
D f .ˇ0 C ˇ1 wi C ˇ2 zi /ˇ1 ;
@wi

329
 Pdf of N(0,1) Pdf of N(0,1)
0.4 Pr(x ≤ −1) = 0.4 Pr(x ≤ 0.5) =
0.16 0.69
0.3 0.3

0.2 0.2

0.1 0.1

0 0
−2 0 2 −2 0 2
x x

Cdf of N(0,1)
1

0.5

0
−2 0 2
x

Figure 19.1: Pdf and cdf of N(0,1)

where f ./ is the derivative of F ./. Sign(derivative)Dsign(ˇ1 ) Calculated at some typical

value of ˇ0 C ˇ1 wi C ˇ2 zi .

Example 19.2 If a regressor is a dummy variable, then use a simple difference instead of
attempting a derivative. For instance, if zi is either 0 or 1, then we can use

F .ˇ0 C ˇ1 wi C ˇ2 / F .ˇ0 C ˇ1 wi /:

This is calculated at some typical value of ˇ0 C ˇ1 wi .

Notice: the ratio of two coefficients equals the ratio of their marginal effect on the
probability
@F .xi0 ˇ/ @F .xi0 ˇ/
ˇk =ˇm D =
@xk;i @xm;i

330
 OLS, linear probability model Probit, MLE

1 1
Pr(y = 1)

Pr(y = 1)
0.5 Estimate 0.5
Data

0 0

−10 −5 0 5 10 −10 −5 0 5 10
x x

y = 1 if y ∗ > 0 and zero otherwise

y ∗ = 2 + 1.3x + ǫ, ǫ ∼ N (0, 4)
x ∼ N (0, 5)

Figure 19.2: Example of probit model

Probit estimation of HasAuto (0 or 1)

Probit estimation of HasAuto
coeff t-stat
Pr(HasAuto = 1)

1 const -0.06 -0.16

dtime 0.03 2.92

0.5 data
est McFadden’s R2 : 0.58

−100 −50 0 50 100

dtime (bus time - auto time to work)

Figure 19.3: Example of probit model, Hill et al (2008), Table 16.1

19.1.2 Estimation

The model is typically estimated with MLE. To do that we need to construct the likelihood
function.

Remark 19.3 Bernoulli distribution. Pr.yi D 1/ D pi , Pr.yi D 0/ D 1 pi .

331
 Assume independent observations (firm 1 and 2). Then, the probabilities (likelihoods)
for the different outcomes are

Pr.y1 D 1 and y2 D 1/ D p1 p2 (19.4)

Pr.y1 D 1 and y2 D 0/ D p1 .1 p2 /
Pr.y1 D 0 and y2 D 1/ D .1 p1 / p2
Pr.y1 D 0 and y2 D 0/ D .1 p1 / .1 p2 /

This list will be long (and messy to program) when there are many observations
(firms). We therefore use an alternative way of writing the same thing as in (19.4). First,
notice that (
y1 p1 if y1 D 1
p1 .1 p1 /1 y1 D (19.5)
1 p1 if y1 D 0:
For the sample wit two data points, the probability (likelihood) can be written

L D p1y1 .1 p1 /1 y1
 p2y2 .1 p2 /1 y2
: (19.6)

Let pi D F .xi0 ˇ/ from (19.2) and use in (19.6) to get a likelihood function for two data
points
1 y1 1 y2
L D F .x10 ˇ/y1 1 F .x10 ˇ/  F .x20 ˇ/y2 1 F .x20 ˇ/
 
:

or as log (after slight rearranging)

ln L D y1 ln F .x10 ˇ/ C y2 ln F .x20 ˇ/ (19.7)

C .1 y1 / ln 1 F .x10 ˇ/ C .1 y2 / ln 1 F .x20 ˇ/ :
   

For N data points, this generalizes to

PN
ln L D ln F .xi0 ˇ/ C .1 yi / ln 1 F .xi0 ˇ/ : (19.8)
 
i D1 yi

We find the ML estimate by maximizing this log likelihood function with respect to the
parameters ˇ.
See Figure 19.3 for an empirical example.

332
19.1.3 Goodness of Fit

To measure of fit, we use several different approaches—since a traditional R2 is not ap-

propriate for a non-linear model.
First, McFadden’s R2 is a commonly applied measure that has many features in com-
mon with a traditional R2 . It is
log likelihood value (at max)
McFadden’s R2 D 1 : (19.9)
log likelihood value (all coeffs=0, except constant)
Notice: ln L < 0 since it is a log of a probability (the likelihood function value), but gets
closer to zero as the model improves. McFadden’s R2 (19.9) is therefore between 0 (as
bad as a model with only a constant) and 1 (perfect model).

Example 19.4 If ln L D ln 0:9 (at max) and the model with only a constant has ln L D
ln 0:5
ln 0:9
McFadden’s R2 D 1  0:84
ln 0:5
If instead, the model has ln L D ln 0:8 (at max), then
ln 0:8
McFadden’s R2 D 1  0:68
ln 0:5
An alternative measure of the goodness of fit is an “R2 ” for the predicted probabilities.
To compare predictions to data, let the predictions be

O > 0:5
(
1 if F .xi0 ˇ/
yOi D (19.10)
0 otherwise.

O is higher than 50%, then we define the

This says that if the fitted probability F .xi0 ˇ/
fitted binary variable to be one, otherwise zero. We now cross-tabulate the actual (yi ) and
predicted (yOi ) values.
yOi D 0 yOi D 1 Total
yi D 0: n00 n01 N0
(19.11)
yi D 1: n10 n11 N1
Total: NO 0 NO 1 N
For instance, n01 is the number of data points for which yi D 0 but yOi D 1. Similarly, N0
is the number of observations for which yi D 0 (and it clearly equals n00 C n01 ). Define

333
an “Rpred
2
” for the prediction as

number of incorrect predictions

“Rpred
2
”D1 : (19.12)
number of incorrect predictions, constant probabilities

This is somewhat reminiscent of a traditional R2 since it measures the errors as the num-
ber of incorrect predictions—and compare the model with a very static benchmark (con-
stant probability).
The constant probability is just the fraction of data where the binary variable equals
one

pO D Fraction of .y1 D 1/
D N1 =N: (19.13)

If pO  0:5, so the (naive) constant probability model predicts yi D 0, then the number of
incorrect predictions is N1 . Otherwise it is N0 . For the estimated model, the number of
incorrect predictions (when yOi ¤ yi ) is n10 C n01 . This gives the “Rpred
2
” in (19.12) as
(
n10 Cn01
1 if pO  0:5
“Rpred
2
”D N1
n10 Cn01
1 N0
if pO > 0:5:

Example 19.5 Let xi be a scalar. Suppose we have the following data

h i h i
y1 y2 y3 y4 D 1 0 1 1 and
h i h i
x1 x2 x3 x4 D 1:5 1:2 0:5 0:7

See Figure 19.4

Suppose ˇ D 0, then we get the following values
h i
0
F .xi ˇ/ D 0:5 0:5 0:5 0:5

yi log F .xi0 ˇ/ C .1 yi / log 1 F .xi0 ˇ/

 
h i
 0:69 0:69 0:69 0:69
log L  2:77

334
Now, suppose instead that ˇ D 1
h i
F .xi0 ˇ/  0:93 0:12 0:69 0:24

yi log F .xi0 ˇ/ C .1 yi / log 1 F .xi0 ˇ/

 
h i
 0:07 0:12 0:37 1:42
log L  1:98;

which is higher than at ˇ D 0. If ˇ D 1 happened to maximize the likelihood function (it

almost does...), then
1:98
McFadden’s R2 D 1  0:29
2:77
and the predicted would be h i
yOi  1 0 1 0 :

Cross-tabulation of actual (yi ) and predicted (yOi ) values

yOi D 0 yOi D 1 Total

yi D 0: 1 0 1
yi D 1: 1 2 3
Total: 2 2 4

Since the constant probability is

pO D 3=4;

the constant probability model always predicts yi D 1. We therefore get

2 1
“Rpred ”D1 D 0:
1C0

19.1.4 Related Models

Multi-response models answers questions like “a little, more, most?” (ordered logit or
probit) or “Red, blue or yellow car?” (unordered models: multinomial logit or probit).
Models for count data are useful for answer questions like: “how many visits to the
supermarket this week?” They are like a standard model, but yi can only take on integer
values (0; 1; 2; 3; ::).

335
 Likelihood function of probit model
0
y: 1.0 0.0 1.0 1.0
−2 x: 1.5 -1.2 0.5 -0.7

−4

−6

−8

−10

−12

−2 −1 0 1 2 3
β

Figure 19.4: Example of ML estimation of probit model

19.2 Truncated Regression Model

19.2.1 Basic Model Setup

Suppose the correct model is linear

yi D xi0 ˇ C "i , "i  iidN.0;  2 /; (19.14)

but that data (also regressors) are completely missing if yi  0

" #
yi D yi if yi > 0
(19.15)
.yi ; xi / not observed otherwise.

The problem with this is that the sample is no longer random. For instance, if yi is
dividends, xi is profits—and it so happens that firms with low dividends are not in the
sample. See Figure 19.7 for an illustration.
In fact, running OLS of
yi D xi0 ˇ C "i (19.16)

on the available data will give biased (and inconsistent) estimates.

The reason is that we only use those data points where yi is unusually high (for a

336
given value of xi0 ˇ/. To be precise, the expected value of yi , conditional on xi0 ˇ and that
we observe the data, is

E .yi jyi > 0; xi / D xi0 ˇ C E "i jyi > 0 (19.17)

D xi0 ˇ C E "i j"i > xi0 ˇ : (19.18)

The second line follows from the fact that yi > 0 happens when xi0 ˇ C "i > 0 (see
(19.15)) which happens when "i > xi0 ˇ. The key result is that last term is positive (recall
E "i D E ."i j"i > 1/ D 0), which make the OLS estimates inconsistent. The result in
(19.18) means that our data has a higher mean that the corresponding xi0 ˇ would motivate.
Since OLS creates the estimates to make sure that the residual has a zero mean, so OLS
will tilt the coefficient estimates away from ˇ. The basic reason is that E "i j"i > xi0 ˇ

varies if xi , so it acts like an extra error term that is correlated with the regressor—which
is a classical reason for why OLS is inconsistent. The following examples illustrate how.

Remark 19.6 (Truncated normal distribution) Let "  N.;  2 /, then

.a0 /
E."j" > a/ D  C  and a0 D .a /=
1 ˚.a0 /
See Figure 19.5.

Example 19.7 As a trivial example, suppose the model is yi D 0C"i with "i iidN.0; 1/.
Then

E .yi jyi > 0; xi / D 0 C E ."i j"i > 0/

.0/ p
D0C D 2=  0:80;
1 ˚.0/
which is far from the true mean (0). OLS will therefore estimate an intercept of around
0.8 instead of 0.

Example 19.8 Suppose the model is yi D 2xi C "i with "i iidN.0; 1/ and where xi a
scalar random variable. Then

E .yi jyi > 0; xi / D 2xi C E ."i j"i > 2xi /

. 2xi /
D 2xi C
1 ˚. 2xi /

337
 Conditional expectation of residual
4
E(ǫ|ǫ > a), ǫ ∼ N(0, 1)
E(ǫ|ǫ > −a), ǫ ∼ N(0, 1)
3

−1
−3 −2 −1 0 1 2 3
a

Figure 19.5: Expectations of a truncated variable

For some selected values of xi we have

E .yi jyi > 0; xi / D

D 2xi C E ."i j"i > 2xi /
8
< 2  . 1/ C E ."i j"i > 2/
ˆ xD 1
D 2  0 C E ."i j"i > 0/ xi D 0
ˆ
2  1 C E ."i j"i > 2/ xi D 1
:
8
< 2  . 1/ C 2:37 D 0:37
ˆ xD 1
D 2  0 C 0:8 D 0:80 xi D 0
ˆ
2  1 C 0:06 D 2:06 xi D 1
:

so the slope is lower than 2: OLS will therefor fit a slope coefficient that is lower than
2. See Figure 19.6. The basic point is that E ."i j"i > 2xi / is much higher for low than
for high values of xi (compare xi D 1 and xi D 1), making the regression line look
flatter. (Notice that 1 .˚.2x2x
i/
i/
.2xi /
can also be written ˚.2x i/
since the N(0,1) distribution is
symmetric around zero.)

338
 E(y|y > 0), y = 2x + ǫ, ǫ ∼ N(0, 1)
4

−2
E(ǫ|ǫ > −2x)
2x
E(y|y > 0) = 2x + E(ǫ|ǫ > −2x)
−4
−1.5 −1 −0.5 0 0.5 1 1.5
x

Figure 19.6: Expectations of a truncated variable

19.2.2 Estimation

Remark 19.9 (Distribution of truncated a random variable) Let the density function of
a random variable X be pdf.x/. The density function, conditional on a < X  b is
pdf.xja < X  b/ D pdf.x/= Pr.a < X  b/. Clearly, it could be that a D 1 and/or
b D 1..

We need the density function of yi conditional on yi > 0, or equivalently of "i ,

conditional on yi D xi0 ˇ C "i > 0 (so "i > xi0 ˇ)

pdf."i /
pdf."i j"i > xi0 ˇ/ D : (19.19)
Pr."i > xi0 ˇ/

If "i  N.0;  2 /, the denominator is

Pr."i > xi0 ˇ/ D Pr "i = > xi0 ˇ= (19.20)

D 1 ˚ xi0 ˇ= (19.21)

D ˚ xi0 ˇ= : (19.22)

The second line follows from N.0; 1/ being symmetric around 0, so ˚ .z/ D 1 ˚ . z/.
Combine (19.22) with a N.0;  2 ) distribution for "i , replace "i by yi xi0 ˇ and take

339
logs to get
X
ln L D ln Li , where (19.23)
i

Li D pdf."i j"i >

xi0 ˇ/  Pr.yi > 0/ (19.24)

2 !
1 1 yi xi0 ˇ
Dp exp 0
(19.25)


=˚ xi ˇ= :
2 2 2 2

We maximize this likelihood function with respect to ˇ and  2 (numerical optimization).

Notice: ˚ xi0 ˇ= is the new part compared with OLS. See Figure 19.7 for an illustration.

19.3 Censored Regression Model (Tobit Model)

The censored regression model is similar to truncated model, but we are fortunate to
always observe the regressors. xi . We have a bit more information than in truncated case,
and we should try to use it. In short, the model and data are

yi D xi0 ˇ C "i , "i  iidN.0;  2 / (19.26)

(
.yi ; xi / if yi > 0
Data: yi D
.0; xi / otherwise.

Values yi  0 are said to be censored (and assigned the value 0—which is just a normal-
ization). This is the classical Tobit model.
If we estimate yi D xi0 ˇ C "i (with LS), using all data with yi > 0, then we are in
same situation as in truncated model: LS is not consistent. See Figure 19.7.
Example: yi is dividends, xi is profits—firms with low dividends are assigned a
common value (normalized to yi D 0) in the survey.

19.3.1 Estimation of Censored Regression Model

Remark 19.10 (Likelihood function with different states) The likelihood contribution of
observation i is pdf.yi / which can also be written pdf .yi jstate K/  Pr.state K/. See
Remark 19.9

There are two states: yi  0 and yi > 0.

340
 Full sample Truncated sample
20 y = 2 + 1.3x + ǫ
∗
20 y = y if y ∗ > 0
∗

ǫ ∼ N (0, 4) no data otherwise

10 10

0 0
y

y
−10 −10 data
data LS
LS MLE
−20 −20
−10 −5 0 5 10 −10 −5 0 5 10
x x

Notice:
using only y > 0 data from a censored
Censored sample sample gives a truncated sample
20 y = y ∗ if y ∗ > 0
y = 0 otherwise
10 x is always observed

0
y

−10
data
MLE
−20
−10 −5 0 5 10
x

Figure 19.7: Estimation on full, truncated and censored sample

State yi  0 (that is, no data on yi but on xi ) happens when yi D xi0 ˇ C "i  0, that
is, when "i  xi0 ˇ. The probability of this is

Pr."i  xi0 ˇ/ D Pr."i =  xi0 ˇ=/

D ˚. xi0 ˇ=/: (19.27)

(By symmetry of the normal distribution, this also equals 1 ˚.xi0 ˇ=/.) The conditional
density function in this state has the constant value of one, so the likelihood contribution
(see Remark 19.10) is

Li .if yi  0/ D pdf.yi jyi  0/  Pr.yi  0/

D 1  ˚. xi0 ˇ=/: (19.28)

State yi > 0 happens in the same way as in the truncated model (19.19), but the dif-

341
 Hours worked, married women, PSID 1975, sample size: 753
0.5

0.4
frequency

0.3

0.2

0.1

0
0 1000 2000 3000 4000 5000
hours worked

Figure 19.8: Example of probit model, Hill et al (2008), Table 16.1

ference here is that the contribution to the likelihood function (again, see Remark 19.10)
is

Li .if yi > 0/ D pdf."i j"i > xi0 ˇ/  Pr."i > xi0 ˇ/
D pdf."i /:

2 !
1 1 yi xi0 ˇ
Dp exp i
: (19.29)
2 2 2 2

The likelihood function is defined by (19.27) and (19.29). Maximize with respect to
ˇ and  2 (numerical optimization). Compared to OLS, the new part is that we have a way
of calculating the probability of censored data (19.28)—since we know all xi values.

19.3.2 Interpretation of the Tobit Model

We could be interested in several things. First, how is probability of yi D 0 affected by a

change in regressor k? The derivative provides an answer

@ Pr.yi D 0/
D .xi0 ˇ=/ˇk =: (19.30)
@xi k

342
 OLS MLE
const 1335:3 1349:9
.5:7/ 3:5
educ 27:1 73:3
.2:2/ 3:5
exper 48:0 80:5
.13:2/ 12:2
age 31:3 60:8
. 7:9/ 8:4
kids16 447:9 918:9
. 7:7/ 8:1
sigma 753:0 1133:7
26:8
Nobs 753:0

Table 19.1: Tobit estimation of hours worked. Example of a tobit model, Hill et al (2008),
Table 16.8. Numbers in parentheses are t-stats.

This derivative has a absolute value when xi0 ˇ  0, since a small change in xk can then
tip the balance towards yi D 0. In contrast, when xi0 ˇ is very small or very large, then a
small change in xk does not matter much (as we are already safely in yi D 0 or yi D 1
territory). Second, wow is the expected value of yi affected by a change in regressor k?
Once again,we can calculate a derivative

@ E yi
D ˚ xi0 ˇ= ˇk : (19.31)


@xi k
Notice that this derivative depends on xi0 ˇ. For low values of xi0 ˇ, the derivative is close
to zero (since ˚ xi0 ˇ=  0). In contrast, for high values of xi0 ˇ, the derivative is close

to ˇk .

19.4 Heckit: Sample Selection Model

Recall that in a Tobit model, xi0 ˇ C "i decide both the probability of observing yi and its
value. “Heckit” models relax that.

343
 A sample selection model is a two equation model

wi D x1i
0
ˇ1 C "1i (19.32)
hi D x2i
0
ˇ2 C "2i : (19.33)

For instance, wi could be individual productivity and hi could be labour supply, and x1i
and x2i could contain information about education, age, etc. In this special case where
Corr.hi ; wi / D 1, then we are back in standard Tobit model.
It is typical to assume that the residuals in the two equations could be correlated
" # " # " #!
"1i 0 12 12
N ; : (19.34)
"2i 0 12 1

Notice that Var."2i / D 1 is a normalization. A correlation, 12 ¤ 0, means that some un-
observed characteristics (part of the residuals) are affecting both equations. For instance,
“ability” may be hard to measure but is likely to affect both productivity and the labour
supply choice.
The data on wi only observed for people who work (their hourly wage), and hi is
only observed as 0/1 (doesn’t work/works)
(
wi D wi ; hi D 1 if hi > 0
Data: (19.35)
wi not observed, hi D 0 otherwise

To understand the properties of this model, notice that the expected value of wi , con-
ditional on hi D 1, is

E .wi jhi D 1/ D x1i

0
ˇ1 C E ."1i jhi D 1/
„ ƒ‚ …
E."1i j"2i > x2i ˇ2 /
0

0
D x1i ˇ1 C E "1i j"2i > 0


x2i ˇ2
0
.x2i ˇ2 /
0
D x1i ˇ1 C 12 i , where i D 0
; (19.36)
˚.x2i ˇ2 /

where ./ and ˚./ are the standard normal pdf and cdf (i is called the inverse Mill’s
ratio or Heckman’s lambda). Showing this is straightforward, but a bit tedious. The point
of (19.36) that the correlation of the residuals in the two equations (19.32)–(19.33) is
crucial. In fact, when 12 D 0, then we can estimate (19.32) with OLS. Otherwise, it is

344
biased (and inconsistent).
Another way to see this: for the observable data (when hi D 1)

0
wi D x1i ˇ1 C "1i

and the issue is: E.x1i "1i / D 0 for this data? To keep it simple, suppose x2i includes
just a constant: wi observed only when "2i > 0. If Corr."1i ; "2i / > 0, our sample of wi
actually contains mostly observations when "1i > 0 (so "1i isn’t zero on average in the
sample). This gives a sample selection bias.
Is 12 ¤ 0? Must think about the economics. In wage and labour supply equations:
"1t and "2t may capture some unobservable factor that makes a person more productive
at the same time as more prone to supply more labour.
What if Cov.x1i ; i / D 0 (although 12 ¤ 0)? Well, then OLS on

0
wi D x1i ˇ C "1i

is consistent (recall the case of uncorrelated regressors: can then estimate one slope coef-
ficient at a time). The conclusion is that the bias of OLS comes from Cov.x1i ; x2i / ¤ 0
since then Cov.x1i ; i / ¤ 0: extreme case x1i D x2i .

19.4.1 Estimation

Use MLE or Heckman’s 2-step approach, which is as follows:

1. Estimate (19.33) with Probit method (recall hi D 0 or 1). We are then estimating
0 O
ˇ2 in Pr.hi D 1/ D F .x2i
0
ˇ2 /. Extract x2i ˇ2 and create O i as in (19.36).

2. Estimate (ˇ1 and 12 ) with LS

0
wi D x1i ˇ1 C 12 O i C i (19.37)

on the data where wi is observed (and not artificial set to zero or some other value).

Properties: consistent, may need to adjust standard errors (unless you test under the
null hypothesis that 12 D 0).

345
 w = w ∗ , h = 1 if h∗ > 0 Regression lines
15 otherwise, w not observed, h = 0

w ∗ = 1 + 2x1 + ǫ1 , ǫ1 ∼ N (0, 5)
10 h∗ = 0.3 + 1x1 + 4z + ǫ2 , ǫ2 ∼ N (0, 1)
x1 ∼ N (0, 2), z ∼ N (0, 0.1)
Cov(ǫ1 , ǫ2 ) = 2
5

0
w

−5

−10 OLS
Heckman
Observed data
Unobserved data
−15
−5 −4 −3 −2 −1 0 1 2 3 4 5
x1

Figure 19.9: Sample selection model

LS Heckman
const 0:40 1:21
. 2:10/ .2:11/
educ 0:11 0:05
7:73 2:23
exper 0:02 0:02
3:90 4:18
lambda 1:56
2:98
R2 0:15
Nobs 753:00

Table 19.2: OLS and Heckman estimation of log wages, married women, PSID 1975.
Example of a Heckman model, Hill et al (2008), Table 16.8. Numbers in parentheses are
t-stats.

Bibliography
Verbeek, M., 2012, A guide to modern econometrics, Wiley, 4th edn.

346
 Tobit
const 1:05
.2:20/
age 0:01
. 2:98/
educ 0:05
.3:54/
kids 0:19
. 2:45/
mtr 0:96
. 2:33/

Table 19.3: Tobit estimation of labour market participation, hours>0. 1st step of Heckman
estimation. Example of a Heckman model, Hill et al (2008), Table 16.8. Numbers in
parentheses are t-stats.

347