FM423

Asset Markets
Lecture 10: Multifactor Models
 0. Overview Page 1

General Overview

Over the next 3 lectures, we will cover these (slightly) applied topics

(1) Multifactor Models (1 lecture)

• The Arbitrage Pricing Theory (APT)

• Three-factor Model

(2) Market Efficiency and Market Anomalies (2 lectures)

• Defining and Characterizing Market Efficiency

• Testing Market Efficiency
• Do Fund Managers Outperform?
• Market Anomalies
• Behavioral Finance

Lecture 10: Multifactor Models Copyright c Zachariadis

 0. Overview Page 2

Overview of Lecture 10

(1) Multifactor models: Motivation and description

(2) Multifactor models: Specification and varieties

(3) The APT: The concept of arbitrage

(4) The APT: Derivation and Intuition

(5) Comparison of APT and CAPM

(6) Practical applications of Multifactor models

Lecture 10: Multifactor Models Copyright c Zachariadis

 1. Multifactor Models: Motivation and Description Page 3

1. Multifactor Models: Motivation and

Description

Starting point: the CAPM

Under the CAPM, an asset’s β is the only influence on that asset’s expected return, and β can be thought
of as the link between systematic risk and expected returns.

To understand cross-sectional variation in average stock returns, all you need to know is the cross-section
of β s.

Modification

What if investors face other sources of risk as well as stock market risk? In this situation, it is likely that an
asset’s exposures to these other risk-factors would affect its expected returns.

Lecture 10: Multifactor Models Copyright c Zachariadis

 1. Multifactor Models: Motivation and Description Page 4

Example

Hypothesize that there are 3 key sources of risk that affect the returns on stocks:

• Business cycle risk: which can be measured by GDP growth

• Oil price risk: which can be measured by changes in the price of oil

• Interest rate risk: measured by looking at returns of long-term government bonds relative to short-term
government bonds

It is reasonable to assume that different stocks will respond differently to these three sources of risk.
Consider, for example, the sensitivity of corporate profits for a financial services firm, an airline, and a
luxury goods firm to these three systematic risks.

Lecture 10: Multifactor Models Copyright c Zachariadis

 1. Multifactor Models: Motivation and Description Page 5

Writing down a Multifactor model

Consider the example on the previous slide. We can formalize its implications in the following model:

ri,t = ai + bi,1 I1,t + bi,2 I2,t + bi,3 I3,t + ei,t

where;

• ri,t is return on stock i at time t

• I1,t , I2,t , and I3,t are the realizations of GDP, oil price and interest rate factors at time t
• bi,1 , bi,2 , and bi,3 are the sensitivities to or loadings on the three factors for stock i
• ei,t is a random return component for stock i at time t
Implications:

• Different stocks can have different sensitivities to the three factors

• As the factors vary over time, so do the returns on the stocks in question
• Some portion of the return on any security cannot be explained by any of our factors, represented by
the stock-specific random return component

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 6

2. Multifactor model: Specifications

and Varieties

Basic setup: N stocks and K risk factors where N >> K

The basic return generating model looks as follows:

K
X
ri,t = ai + bi,j Ij,t + ei,t
j=1

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 7

Restrictions

We impose the following restrictions on the factors,

E(Ii,t ) = 0 ∀i, t

Cov(Ij,t , Ik,t ) = 0 ∀t, j 6= k

on the random return components,

E(ei,t ) = 0 ∀i, t

V ar(ei,t ) = σi2 ∀i, t

Cov(ej,t , ek,t ) = 0 ∀t, j 6= k

and on the relationship between factors and the random return components,

cov(ei,t , Ij,t ) = 0 ∀t, i, j

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 8

Interpreting the model

Given the basic model and the conditions on both the random return component and the factors, we have
the following implications:

• All stock returns are generated as a linear combination of the realizations of K common risk factors
plus a stock-specific constant term and a stock-specific random return component

• Different stocks might have different sensitivities to the K risk factors via the values of the bi,j
coefficients

• The expected return is E(ri,t ) = r̄i,t = ai

• The risk factors are uncorrelated with each other

• The random return components are uncorrelated with each other and also with all of the risk factors

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 9

Factors and the covariance of the stock returns

Consider a world with 3 risk factors. Let’s use our factor model to determine the covariance between
returns on two stocks, X and Y :

Cov(rX,t , rY,t )
 
 
 X3 X3 
 
= Cov aX + bX,j Ij,t + eX,t , aY + bY,j Ij,t + eY,t 
 
 j=1 j=1 
| {z } | {z }
rX,t rY,t

3 3
!
X X
= Cov bX,j Ij,t , bY,j Ij,t ,
j=1 j=1

where the a terms drop out because they are constants, and the e’s disappear because they are
uncorrelated with everything else in the model.

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 10

Factors and the covariance of the stock returns: Cont’d

Finally, as the common factors are all uncorrelated with one another, we get:
3
X
Cov(rX,t , rY,t ) = bX,j bY,j V ar(Ij,t ).
j=1

Thus,

• Stock returns are correlated due to the common risk factors

• Stock-specific risks (e’s) affect the variance of stock returns but not their covariances

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 11

How do we identify the factors?

There are a number of approaches to the problem of choosing the number and identities of the risk factors
in the model:

• Economic factor models: use macroeconomic/financial intuition to decide on the set of risk factors

• Statistical factor models: use a statistical technique called ‘factor analysis’ to identify factors (Connor
and Korajczyk 1988, Bai and Ng, 2002)

• Characteristic-based factor models: use corporate characteristics (e.g. firm size) to measure the
sensitivity of stocks to unobserved risk factors

We will focus on the first and last of these approaches.

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 12

Economic factor models: Chen, Roll and Ross (1986)

From our work on valuation, we know that stock prices can be estimated as the sum of discounted
expected future cashflows/dividends. Thus, stock returns should be driven by unanticipated movements in
expected future cashflows and discount rates.

Chen, Roll and Ross (1986) use that intuition to propose the following broad set of factors:

• Growth in industrial production

• Unexpected inflation
• Change in expected inflation
• Risk premium: realized return on low-grade long-term debt relative to government long-term debt
• Term structure: realized return on long-term government bonds relative to short-term government
bonds

• Market index return: to soak up any remaining common movement in returns

Other candidate factors: exchange rate changes, unexpected unemployment ...

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 13

The Fama-French Three Factor Model

Fama and French proposed a three-factor specification which has become very popular and combines
elements of economic and characteristic-based factor modeling.

ri,t − rf,t = bi (rm,t − rf,t ) + si (SM Bt ) + hi (HM Lt ) + ei,t ,

where ri,t − rf,t is the stock return in excess of the riskfree rate.
Factor return measurements:

• rm,t − rf,t , the excess return on the market, is the value-weight return on all NYSE, AMEX, and
NASDAQ stocks (from CRSP) minus the one-month Treasury bill rate (from Ibbotson Associates)

• SMB (Small minus big) is the size factor and HML (High Minus Low) is the value factor

Extensions: Several other characteristic-based factors have been added to Fama and French’s original
specification. These include factors for momentum, profitability, and investment.

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 14

Implementation

• For each stock in one’s investment universe, compute size (market value of equity) and value (the
book-to-market-equity ratio) once a year at the end of June

• Calculate the median NYSE size as well as the 30th and 70th NYSE percentiles for value
• Using these breakpoints, split the universe into 3 value groups and 2 size groups. The intersection of
these groups generates 6 elementary portfolios
• The return on SMB at t is the equal-weight return on the three small portfolios minus the equal-weight
return on the three big portfolios
SM B = 1/3(SmallV alue + SmallN eutral + SmallGrowth)

−1/3(BigV alue + BigN eutral + BigGrowth)

• The return on HML at t is the equal-weight return on the two value portfolios minus the equal-weight
return on the two growth portfolios
HM L = 1/2(SmallV alue + BigV alue)

−1/2(SmallGrowth + BigGrowth)

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 15

Estimation

With either macroeconomic factor models or characteristic-based factor models, for each one of our N
stocks, we simply estimate the factor sensitivities (i.e., the bi,j coefficients) via a time-series regression of
stock returns on the factors in question:

ri,t = ai + bi,1 I1,t + bi,2 I2,t + bi,3 I3,t + ei,t

Of course, linear regression theory ensures that the error terms are uncorrelated with the factors.

Lecture 10: Multifactor Models Copyright c Zachariadis

2. Multifactor model: Specifications and Varieties Page 16

Fama-French Factors

Lecture 10: Multifactor Models Copyright c Zachariadis

 2. Multifactor model: Specifications and Varieties Page 17

Fama - French Example: MSFT

• Average factor returns 1927-2015:

– E(rm,t − rf,t ) = 8.3%
– E(SM B) = 3.3%
– E(HM L) = 4.9%

• CAPM regression:

rM SF T − rf = 0.994 + 1.708 · (rM − rf ) + ǫ

E(rM SF T ) = 2% + 1.708 · 8.3% = 16.2%

• FF regression:

rM SF T − rf = 0.634 + 1.096 · (rM − rf ) − 1.374 · SM B − 1.389 · HM L + ǫ

E(rM SF T ) = 2% + 1.096 · 8.3% − 1.374 · 3.3% − 1.389 · 4.9% = −0.2%

Lecture 10: Multifactor Models Copyright c Zachariadis

 3. The Arbitrage Pricing Theory Page 18

3. The Arbitrage Pricing Theory

We will now use our generic factor structure to derive the Arbitrage Pricing Theory or APT. As its name
suggests, the theory relies on two main underpinnings:

• A multifactor structure for returns, and

• Absence of arbitrage.
Recall that the CAPM is an equilibrium model (i.e., it imposes demand = supply in the security market).
The APT does not impose equilibrium and instead requires that securities are priced such that investors
cannot find any arbitrage opportunities.

Lecture 10: Multifactor Models Copyright c Zachariadis

 3. The Arbitrage Pricing Theory Page 19

Review: Absence of arbitrage

Recall that in a well-functioning financial market, arbitrage opportunities should not exist. If they did,
investors would pursue arbitrage strategies on a massive scale, and, by doing so, would eliminate the
arbitrage opportunity.

To see why, consider the effect of running the arbitrage strategy on the prices of A and B.

• The arbitrage creates buying pressure for A, which should increase A’s price

• The arbitrage creates selling pressure for B, which should decrease B’s price

As the price of A rises and B falls, the arbitrage opportunity shrinks and eventually disappears.

Lecture 10: Multifactor Models Copyright c Zachariadis

 3. The Arbitrage Pricing Theory Page 20

Review: Formal definition of arbitrage

We can describe two distinct types of arbitrage portfolios

• Portfolios with negative investment costs (i.e. where proceeds of sales exceed costs of purchases) and
zero future payoffs in all states of nature

• Portfolios with zero investment costs, strictly non-negative future payoffs in all states of nature, and at
least one positive future payoff in some state of nature

Assume that we are trying to price a new asset, Z. Our first step is to create a portfolio of existing assets
that has identical payoffs to Z in every state of nature – the replicating portfolio. Then, to ensure no
arbitrage opportunity exists, the price of Z must be exactly the same as the cost of replicating portfolio.

Note: This method of pricing relies only on assuming that investors prefer more to less money and no
further preference restrictions.

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 21

4. The APT: Simple derivation

We now have all of the ingredients to derive the APT.

• We assume that individuals prefer more to less and thus act so as to eliminate any arbitrage
opportunities that arise.

• All investors know that returns on our N assets are generated by a factor model.

Start with a two-factor model for a set of N diversified portfolios (“mutual funds” ) as follows:

ri,t = ai + bi,1 I1,t + bi,2 I2,t , i = 1, . . . , N.

Note: The assumption that asset i is a diversified portfolio means that there is no specific/idiosyncratic risk
(the ei,t term) for this asset.

Example: The 6 elementary portfolios of Fama and French are diversified.

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 22

Factor Characteristics of An Arbitrary Portfolio, P

Consider forming an arbitrary portfolio, P , of our “mutual funds” where the portfolio weights are denoted
wi . The factor loadings of this portfolio are:

N
X
bP,1 = wi bi,1 ,
i=1

N
X
bP,2 = wi bi,2 ,
i=1

N
X
wi = 1.
i=1

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 23

The Factor-replicating Portfolios

Now, we define 3 synthetic securities (by synthetic, we mean portfolios of our original securities) which will
help our derivation. These are called factor-replicating portfolios.

Definition: The factor-replicating portfolio for factor 1, is a portfolio with a loading of 1 on factor 1 and 0 on
factor 2.

Similarly, the factor-replicating portfolio for factor 2 has loading zero on factor 1 and loading 1 on factor 2. A
final portfolio replicates the risk-free asset and has zero loading on both factors.

Portfolio bi,1 bi,2 Mean Return

FRP 1 1 0 r̄1
FRP 2 0 1 r̄2
Risk-free 0 0 rf

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 24

Economic difference between factors and FRP

• The market factor is a traded factor; one can buy or sell the market and earn a return.

• However, factors in general are not necessarily traded

• As a consequence, realizations of these factors are not necessarily returns (e.g. GDP)

• This is why the restriction E(Ii,t ) = 0 is useful

• In contrast, FRP is a portfolio and thus does generate a return

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 25

Example: Fama - French Factors

• I1,t is the size factor, and I2,t is the value factor

• FRP 1 is
SM B = 1/3(SmallV alue + SmallN eutral + SmallGrowth)

−1/3(BigV alue + BigN eutral + BigGrowth).

and FRP 2 is
HM L = 1/2(SmallV alue + BigV alue)

−1/2(SmallGrowth + BigGrowth).

• Since the elementary portfolios are diversified, so are the two resulting factors

• Since SMB and HML are zero-cost portfolios, their returns are excess returns

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 26

Replication

Now comes the application of the absence of arbitrage.

Consider: An asset A with factor loadings bA,1 and bA,2 .

Question: What is its expected return?

Technique: We are going to form a portfolio of the factor-replicating portfolios and the risk-free asset which
has identical factor loadings to A.

Absence of arbitrage: With identical loadings and no specific risk, the expected return on A must be
identical to the expected return on the portfolio we have constructed.

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 27

Calculating the replicating weights

If w1 is the weight on FRP 1 and w2 is the weight on FRP 2 and w3 is the weight on the risk-free asset
then we must have:

bA,1 = w1 × 1 + w2 × 0 + w3 × 0 = w1

bA,2 = w1 × 0 + w2 × 1 + w3 × 0 = w2

1 = w1 + w2 + w3

Hence, given the values for w1 and w2 , we have

w3 = 1 − bA,1 − bA,2

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 28

The expected return on the replicating portfolio is:

w1 r̄1 + w2 r̄2 + w3 rf = bA,1 r̄1 + bA,2 r̄2 + (1 − bA,1 − bA,2 )rf

Example: In the FF case, r̄1 − rf = E(SM B) = 2.7%, and r̄2 − rf = E(HM L) = 6.0%
By absence of arbitrage, this must be equal to the expected return on A. Thus;

aA = r̄A = bA,1 r̄1 + bA,2 r̄2 + (1 − bA,1 − bA,2 )rf

= rf + bA,1 (r̄1 − rf ) + bA,2 (r̄2 − rf )

Bottom Line: This is the key cross-sectional APT equation. Assets have different expected returns because
they have differing sensitivities to the set of risk factors and expected returns are linear in these factor
sensitivities.

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 29

Generalized to k -factor APT model

aA = r̄A = rf + bA,1 (r̄1 − rf ) + bA,2 (r̄2 − rf ) +

... + bA,k (r̄k − rf )

Interpretation: The expected return on any asset or portfolio can be described as the sum of the risk-free
rate and a linear combination of the asset’s factor exposures (i.e., β ’s). The weights in that linear
combinations are given by the excess expected returns of a set of factor-replicating portfolios, also called
factor risk premia (e.g., r̄k − rf for the k th factor).

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 30

Example: Two-factors and three assets

Consider the following three diversified portfolios which exist in a 2-factor world:

Asset ai b1,i b2,i

A 0.05 3 1

B 0.01 0 2

C 0.00 -1 1

The assumption that these portfolios are well-diversified allows us to assume that their residual risks are
zero at all times. Thus, for example, asset A has a return-generating model that looks as follows:

rA = 0.05 + 3I1 + I2

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 31

Technique: We proceed by deriving the weights for 3 portfolios; the first has zero exposure to both common
factors (and is thus risk-free) and the others are pure factor-replicating portfolios for factors 1 and 2
respectively.

Deriving these weights is straightforward – it just involves solving sets of simultaneous equations. For
example, to derive the weights for the risk-free portfolio, we must solve the following set of simultaneous
equations:

3 × w1 + 0 × w2 − 1 × w3 = 0

1 × w1 + 2 × w2 + 1 × w3 = 0

w1 + w2 + w3 = 1

where the final equation is just an adding-up condition for the portfolio weights. Solving these equations
gives w1= 0.5, w2 = −1, and w3 = 1.5. The expected return on this portfolio is
0.5 × 0.05 − 1 × 0.01 + 1.5 × 0.00 = 0.015.

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 32

To derive the weights for the FRP for factor 1, we must solve the following set of simultaneous equations:

3 × w1 + 0 × w2 − 1 × w3 = 1

1 × w1 + 2 × w2 + 1 × w3 = 0

w1 + w2 + w3 = 1

Solving these equations gives w1= 0.75, w2 = −1, and w3 = 1.25. The expected return on this
portfolio is 0.75 × 0.05 − 1 × 0.01 + 1.25 × 0.00 = 0.0275.

Repeating the same exercise to derive the FRP for factor 2, we end up with the following data:

Portfolio w1 w2 w3 r̄
Risk-free 0.5 -1 1.5 0.015

Factor 1 0.75 -1 1.25 0.0275

Factor 2 0.25 0 0.75 0.0125

Lecture 10: Multifactor Models Copyright c Zachariadis

 4. The APT: Simple derivation Page 33

Thus, in this specific context, the APT expected return generating equation for any asset can be written as:

r̄i = rf + (r̄1 − rf )b1,i + (r̄2 − rf )b2,i

= 0.015 + 0.0125b1,i − 0.0025b2,i

All assets in this world must have expected returns and factor exposures that satisfy the APT equation.

Thus, if we were to introduce a new asset, called D, with exposure of 2 and 4 to the first and second factor
respectively, then its expected return would have to be:

r̄D = 0.015 + 0.0125 × 2 − 0.0025 × 4 = 0.03

Lecture 10: Multifactor Models Copyright c Zachariadis

 5. The CAPM and the APT Page 34

5. The CAPM and the APT

If we assume there is only 1 common factor in the APT then the CAPM and the APT equations are very
similar – just substitute r̄1 with r̄m .

However, note the differences in the derivations of the CAPM and the APT.

• The former is an equilibrium model built on the foundation of mean-variance analysis.

• The latter is based on an assumed factor structure for returns along with an application of absence of
arbitrage.

Advantage of the APT over the CAPM: Via absence of arbitrage, it should be (approximately) valid for any
(sufficiently) large subset of risky assets. As we saw previously, testing the CAPM requires one to know the
composition of the market portfolio of assets.

Disadvantage of the APT over the CAPM: The APT factors need to be identified.

Lecture 10: Multifactor Models Copyright c Zachariadis

 6. APT: Applications Page 35

6. APT: Applications

• Valuation
– Derive risk-adjusted discount rates

– Using the discount rates to value firms and investment projects

• Risk measurement

• Portfolio selection

• Performance evaluation

We will elaborate on some of these applications in subsequent lectures.

Lecture 10: Multifactor Models Copyright c Zachariadis

 6. APT: Applications Page 36

Risk Measurement

A key practical problem in portfolio selection and construction is estimating the covariance matrix of
security returns. If we assume that we are trading in a universe consisting of 1000 equities then our return
covariance matrix has about 1,000,000 entries and has the following form:
 
2 2
σ1,1 ... σ1,1000
 
 . .. . 
Σ =  .. . .
. , (1)
 
2 2
σ1000,1 ... σ1000,1000
2 2 2
where σi,j is the covariance between returns on asset i and j . As σi,j = σj,i , estimating this matrix
means estimating 500,500 (= (1000 × 1001)/2) individual variances and covariances.

Lecture 10: Multifactor Models Copyright c Zachariadis

 6. APT: Applications Page 37

Simplifying with APT

The risk matrix can be greatly simplified if we assume that the returns follow a 3 factor model. Assume we
have done the following:

• Identified our 3 factors and their respective variances (V ar (Im,t ) where m = 1, 2, 3.)
• Run 1000 linear regressions to estimate the bi,j ’s (i.e., factor loadings) and σǫ2i ’s (i.e.,the variance of
the stock-specific random component)

Then the return covariances can be calculated as:

3
X
2
σi,j = bi,m bj,m V ar (Im,t )
m=1

and the return variances are:

3
X
σi2 = b2i,m V ar (Im,t ) + σǫ2i
m=1

Such an approach places structure on the covariance matrix that makes estimates of individual
covariances less noisy.

Lecture 10: Multifactor Models Copyright c Zachariadis

 7. Conclusions Page 38

7. Conclusions

We have introduced multifactor generalizations of the risk-return tradeoff by deriving the APT

Multifactor models are useful in several ways:

• Risk adjustment in tests of market efficiency can control for multiple sources of risk

• Portfolio selection and risk measurement for active portfolio management can exploit a multifactor
approach

• Finally, we can use multifactor models to evaluate whether a specific fund manager has performed well
or poorly

The multifactor approach is pervasive in asset management

Lecture 10: Multifactor Models Copyright c Zachariadis

 7. Conclusions Page 39

Key Equations

A time-series factor structure

K
X
ri,t = ai + bi,j Ij,t + ei,t
j=1

and the absence of arbitrage results in the cross-sectional asset pricing restrictions of the APT

ai = r̄i = rf + bi,1 (r̄1 − rf ) + bi,2 (r̄2 − rf ) +

... + bi,K (r̄K − rf )

Lecture 10: Multifactor Models Copyright c Zachariadis