ECMT3150: The Econometrics of Financial

Markets
1a. Linear Time Series Analysis

Simon Kwok
University of Sydney

Semester 1, 2022

1 / 19
Outline

1. AR Model
1.1 Model Properties
1.2 Autocorrelation Function (ACF)
1.3 Model Estimation
1.4 Model Selection
1.5 Goodness-of-…t
1.6 Forecasting

2 / 19
Autoregressive (AR) Model

fyt g follows an AR (p ) model if

yt = φ0 + φ1 yt 1 + φ2 yt 2 + + φp yt p + εt ,

where fεt g is wn(0, σ2ε ).

Suppose fyt g is covariance stationary. The unconditional (long
run) mean is:
φ0
µ E (yt ) = . (1)
1 ∑pi=1 φi

3 / 19
AR Model

Write the AR (p ) model in the demeaned form:

yt = µ + ut ,
p
ut = ∑ φi ut i + εt . (2)
i =1

De…ne the p th -order polynomial function

φ (x ) = φ1 x + φ2 x 2 + + φp x p .
Let L be the lag operator such that Lut = ut 1, Li ut = ut i , and
Lc = c for any constant c.
We can rewrite (2) as

[1 φ(L)] ut = εt .

4 / 19
Autocovariance and Autocorrelation Functions
Suppose fyt g is covariance stationary.
I For a given integer j, the lag-j autocovariance function is
de…ned as
γj = Cov (yt , yt j ).
In particular, the variance is γ0 = Var (yt ).
I For a given integer j, the lag-j autocorrelation function (ACF)
is de…ned as
Cov (yt , yt j ) γj
ρj = Corr (yt , yt j) = = .
Var (yt ) γ0

By convention, ρ0 1.
Both the autocovariance function and ACF are symmetric
functions: γj = γ j and ρj = ρ j .

5 / 19
AR(1) Model
AR (1) model:

yt = µ + ut ,
ut = φ1 ut 1 + εt . (3)

Q: What are the autocovariance and autocorrelation functions of

yt ?
Multiply both sides of (3) by u t j for j 1, and take expectations:

E (u t u t j) = φ1 E (u t 1 u t j ) + E ( εt u t j )
γj = φ1 γj 1. (4)

Note that E (εt u t j) = E [E (εt jFs )u t j] = 0. (4) is the Yule-Walker equation of AR(1).

σ2ε σ2ε φ1
Set j = 1: γ1 = φ1 γ0 . As γ0 = Var (u t ) = 1 φ21
, we can solve for γ1 : γ1 = 1 φ21
.
j
σ2ε φ1
For j > 1, γj = φ1 γj 1. By recursive substitution, we have γj = 1 φ21
.
jj j
σ2ε φ1
By symmetry, the lag-j autocovariance function is γj = 1 φ21
for all integers j .
6 / 19
ACF of AR(1)

γj
Recall that the ACF is ρj = γ0 .
The ACF of AR (1) is
jj j
,
σ2ε φ1 σ2ε
ρj =
1 φ21 1 φ21
jj j
= φ1

for all integers j.

I When 0 < φ1 < 1, the ACF decays smoothly to zero.
I When 1 < φ1 < 0, the ACF alternates between positive and
negative values, but its magnitude decays smoothly to zero.

7 / 19
ACF of AR(1)

8 / 19
ACF of AR(2)
Let yt AR (2): yt = µ + ut and ut = φ1 ut 1 + φ2 ut 2 + εt .
Ex: Show that the autocovariance function is
σ2ε
γ0 = (1 φ2 ),
D
σ2ε
γ1 = φ ,
D 1
γj = φ1 γj 1 + φ2 γj 2 for j 2,
γ j = γj ,

where D (1 + φ2 )(1 + φ1 φ2 )(1 φ1 φ2 ), and that the

ACF is
γ1 φ1
ρ1 = = ,
γ0 1 φ2
ρj = φ1 ρj 1 + φ2 ρj 2 for j 2,
ρ j = ρj .

9 / 19
ACF of AR(2)
The polynomial function of AR (2) is φ (x ) = φ1 x + φ2 x 2 .
Consider the polynomial equation

1 φ (x ) = 0
φ2 x 2 = 0.
1 φ1 x
p 2
φ1 φ1 +4φ2
The equation has two roots x = 2φ .
2

I When the roots are real (φ21 + 4φ2 0), the ACF decays
smoothly.
I When the roots are complex (φ21 + 4φ2 < 0), the ACF is
2π p
oscillating with an average period k = 1
.
cos (φ1 /2 φ2 )
Let the complex roots be a ib. The period kpcan be solved from
φ21 4φ2
cos ( 2π p a
φ1
k ) = a 2 +b 2
, where a = 2φ
2
and b = 2φ 2
.

10 / 19
ACF of AR(2)

11 / 19
PACF
Consider the following regressions:

yt = φ0,1 + φ1,1 yt 1 + e1t

yt = φ0,2 + φ1,2 yt 1 + φ2,2 yt 2 + e2t
yt = φ0,3 + φ1,3 yt 1 + φ2,3 yt 2 + φ3,3 yt 3 + e3t
..
.

The lag-j population partial autocorrelation function (PACF) is

de…ned as φj ,j for each j = 1, 2, . . ..
The lag-j sample PACF is the ordinary least squares (OLS)
estimate φ̂j ,j .
If fyt g follows an AR(p) process, then φ̂p,p ! φp,p as T ! ∞. In
particular, φ̂`,` ! 0 as T ! ∞ for ` > p. The asymptotic variance
of φ̂`,` is T1 for ` > p.
12 / 19
AR Model Estimation
Given the time series data fyt gTt=1 , we can estimate an AR(p)
model by conditional least squares method. Conditional on the
…rst p values: y1 , . . . , yp , we run the regression
yt = φ0 + φ1 yt 1 + + φp yt p + εt
for t = p + 1, p + 2, . . . , T .
Number of observations = T p.
Number of parameters = p + 1.
Degrees of freedom = (T p ) (p + 1) = T 2p 1.
Let φ̂0 , . . . , φ̂p be the OLS coe¢ cient estimates. The residual
series is fε̂t gTt=p +1 , where
ε̂t = yt φ̂0 φ̂1 yt 1 ... φ̂p yt p.

If fεt g is homoskedastic and serially uncorrelated, the variance of

εt is consistently estimated by the sample variance of the residuals
T
1
σ̂2ε =
T 2p ∑
1 t =p +1
ε̂2t .
13 / 19
AR Model Selection

I Akaike information criterion:

AIC = T2 ln(likelihood) + T2 (no. of parameters).
I The Schwarz–Bayesian information criterion:
ln (T )
BIC = T2 ln(likelihood) + T (no. of parameters).
For an AR (p ) model with Gaussian errors, AIC and BIC become

2p
AIC = ln(σ̃2p ) + + constant,
T
p ln(T )
BIC = ln(σ̃2p ) + + constant,
T
where σ̃2p is the maximum likelihood estimate of the error variance:
σ̃2p = T1 ∑Tt=p +1 ε̂2t .

14 / 19
Goodness-of-…t

SSE ∑T 2
t =p +1 ε̂t
I R2 = 1 SST =1 T .
∑t =p +1 t ȳ )2
( y
By SST = SSR + SSE, we have 0 R2 1.
1 T 2
T 2p 1 ∑t =p +1 ε̂t σ̂2ε
I Adjusted R 2 = 1 1 T =1 σ̂2y
, where σ̂2y is
T p 1 ∑t =p +1 (y t ȳ )
2

the sample variance of yt .

Adjusted R 2 may fall outside the [0,1] interval.

15 / 19
Time Series Forecasting
The `-step ahead forecast ŷt (`) is obtained by minimizing the
conditional mean squared error:

ŷt (`) = arg min E [(yt +` g )2 jFt ].

g

The solution is ŷt (`) = E [yt +` jFt ].

Interpretation: the projection of yt +` on the information set at
time t.
The forecast error is

et (`) = yt +` ŷt (`).

Two types of forecasts in practice:

I Conditional forecast: compute ŷt (`) using estimated
parameters, without accounting for parameter uncertainty.
I Unconditional forecast: explicitly account for parameter
uncertainty. Wider con…dence interval around ŷt (`).
16 / 19
Forecasting with AR model
Suppose fyt g follows a stationary AR (p ) with errors fεt g
wn(0, σ2ε ).
For linear models, we consider linear projection. Here, Ft
represents the history of fyt g (or equivalently the history of fεt g)
up to time t, i.e., Ft = fyt , yt 1 , . . .g = fεt , εt 1 , . . .g.
For ` = 1,

ŷt (1) = E [yt +1 jFt ]

= E [φ0 + φ1 yt + + φp yt p +1 + εt +1 jFt ]
= φ0 + φ1 yt + + φp yt p +1 ,

as yt , . . . , yt p +1 are known given Ft , and

E [εt +1 jFt ] = E [εt +1 jεt , εt 1 , . . .] = 0.1
The forecast error is et (1) = yt +1 ŷt (1) = εt +1 , with variance
Var [et (1)] = σ2ε .
1 This would be invalid if the condition expectation is a nonlinear projection,
which is typically the case for a nonlinear model. A stronger mds assumption is
required to ensure E (εt +1 jFt ) = 0.
17 / 19
Forecasting with AR model
For ` = 2,

ŷt (2) = E [yt +2 jFt ]

= E [φ0 + φ1 yt +1 + φ2 yt + + φp yt p +2 + εt +2 jFt ]
= φ0 + φ1 E [yt +1 jFt ] + φ2 yt + + φ p yt p + 2
= φ0 + φ1 ŷt (1) + φ2 yt + + φp yt p +2 .

The forecast error is

et (2) = yt +2 ŷt (2) = φ1 et (1) + εt +2 = φ1 εt +1 + εt +2 , with
variance

Var [et (2)] = φ21 Var (εt +1 ) + Var (εt +2 ) + 2φ1 Cov (εt +1 , εt +2 )
= φ21 σ2ε + σ2ε + 0
= (1 + φ21 )σ2ε .

18 / 19
Forecasting with AR model

Q: Let ` > p. What is the `-step ahead forecast of a stationary

AR (p ) model? What happens to the forecast when ` ! ∞?

A: With ` > p, the `-step ahead forecast is

ŷt (`) = φ0 + φ1 ŷt (` 1) + + φp ŷt (` p ). (5)

By stationarity, ŷt (`) converges to a limit b as ` ! ∞. The limit

must satisfy (5), yielding the solution
φ0
b= ,
1 φ1 φp

which is µ = E (yt ). This is the mean-reverting property.

19 / 19