Lectures 6 and 7

Factor Models and Arbitrage Pricing Theory (APT)

Alexander K. Koch∗
Department of Economics, Royal Holloway, University of London

November 23/26, 2007

In addition to learning the material covered in the reading and the lecture, students should

• be able to distinguish between market-related and non-market-related, as well common-

factor and firm-specific components in decomposing the variance of an asset’s return;

• understand the concept of arbitrage;

• be confident in designing hedging portfolios when given the factor betas for a set of
assets;

• understand how the APT exploits the postulate of no-arbitrage to reach its conclusions.

Required reading:
Bodie, Kane, and Marcus (2008) (Chapters 10 and 11)

Supplementary reading:
Grinblatt and Titman (2002) (Chapter 6), Levy and Post (2005) (Chapter 11)

∗
E-mail: [email protected].

1
1 Factor Models

In lectures 3 and 4 we learned about Markowitz’ portfolio selection theory. As we saw, this
approach proves to be very demanding in terms of the amount of data input needed. We
also saw that, if we are willing to adopt the assumptions of the CAPM, we can reduce the
complexity of the problem. In this framework security prices then only depend on one factor,
namely the market portfolio’s returns.
Let’s step back once again and take a statistician’s view on the problem, rather than the
theorist’s view reflected in the CAPM. What we can observe from the data is that covariances
between securities tend to be positive if the same economic factors affect them. For exam-
ple, stocks in one country are likely to all be affected by the business cycle, interest rates,
changes in technologies, movements in the cost of labor and raw materials... The idea behind
factor models is to summarize all relevant economic variables in a set of indicators that are
the main factors explaining the security market’s returns. The assumption is that few such
factors are sufficient to characterize the common (or systematic) component of securities’
returns, and any remaining uncertainty in security returns can be attributed to firm-specific
(or idiosyncratic) shocks.

1.1 Single factor models

If we believe that one factor is sufficient to capture the systematic component of securities
returns we obtain a single-factor model , which decomposes the holding period return on
security i as follows:

r̃i = E[r̃i ] + βi F̃ + ẽi , (1)

E[F̃ ] = 0, E[ẽi ] = 0, Cov[F̃ , ẽi ] = 0 and Cov[ẽi , ẽj ] = 0 for i 6= j.

Deviations from the expected return on security i (r̃i 6= E[r̃i ]) are explained by two things:
first, an unexpected movement in the common factor F̃ (F̃ 6= E[F̃ ]) – this is called a common
shock ; second, an idiosyncratic (or asset-specific) shock ẽi . Security i’s sensitivity to
common shocks is captured by βi . The idiosyncratic shock is uncorrelated with the common
shock (otherwise the common shock would explain part of the idiosyncratic shock, and one
could not call it idiosyncratic anymore), and idiosyncratic shocks are uncorrelated across se-
curities (which is why we call them idiosyncratic). These assumptions mean that (1) can be
estimated using an Ordinary Least Squares regression.
If we assert that the common component of unanticipated movements in individual security
returns is explained by unanticipated movements in the return on a broad index of securities,

2
such as the S&P500 index, the have a variant of the single-factor model known as the single-
index model . Let r̃M be the return on this market index, then equation (1) becomes

r̃i = r̄i + βi [r̃M − r̄M ] + ẽi , (2)

E[ẽi ] = 0 and Cov[r̃M , ẽi ] = 0.

A convenient transformation of the above consists of looking at excess returns. To derive this,
subtract rf from both sides of equation (2):

r̃i − rf = r̄i + βi [r̃M − rf ] − βi r̄M − (1 − βi ) rf + ẽi ,

⇔ r̃i − rf = r̄i − βi r̄M − (1 − βi ) rf + βi [r̃M − rf ] +ẽi
| {z } | {z } | {z }
⇔ R̃i = αi + βi R̃M +ẽi .

Computing the variance of the excess return on security i, we get

V ar[R̃i ] = βi2 V ar[R̃M ] + V ar[ẽi ]. (3)

Using this formulation, we can decompose the variance of security i’s excess return into

• a systematic component, that captures the variance attributable to uncertainty in

the factor M: βi2 V ar[R̃M ]

• an idiosyncratic component, that captures the firm-specific uncertainty: V ar[ẽi ].

Now, if we look at the covariance between any two securities i and j, we have

Cov[R̃i , R̃j ] = Cov[αi + βi R̃M + ẽi , αj + βj R̃M + ẽj ]

hn on oi
= E αi + βi R̃M + ẽi − E[αi + βi R̃M + ẽi ] αj + βj R̃M + ẽj − E[αj + βj R̃M + ẽj ]
hn on oi
= E βi [R̃M − R̄M ] + ẽi βj [R̃M − R̄M ] + ẽj
h i
= E βi [R̃M − R̄M ] βj [R̃M − R̄M ] + βi [R̃M − R̄M ] ẽj + ẽi βj [R̃M − R̄M ] + ẽi ẽj

= βi βj Cov[R̃M , R̃M ] + βi Cov[R̃M , ẽj ] +βj Cov[R̃M , ẽi ] + Cov[ẽi , ẽj ]

| {z } | {z } | {z }
=0 =0 =0
= βi βj V ar[R̃M ]. (4)

Security covariances are explained entirely by the variance of the market index’s excess returns
and the securities’ different sensitivities to this factor. Thus, if we adopt a single-index model
we require a lot less data than if we implemented Markowitz’ portfolio selection theory. For
an n-security universe both approaches require n estimated expected returns. Instead of
n(n − 1)/2 estimated covariances and n estimated variances for securities, the single-factor
model only needs n estimated sensitivity coefficients, βi , n estimated idiosyncratic variances,

3
V ar[ẽi ], and one estimate for the variance of the common factor, V ar[R̃M ]. In our previous
example, related to the 2,600 companies listed on the London Stock Exchange, we thus cut
data requirements from n (n + 3)/2 = 3, 383, 900 to 3 n + 1 = 7, 801 estimates.

1.2 The CAPM and the single-index model

The CAPM provides a theory that predicts an exact relation to hold for all securities’ expected
returns in relation to the expected return on the market portfolio:

r̄i − rf = βi [r̄M − rf ]. (5)

Realized returns of course depart from their ex ante expected values. Thus, using realized
returns and translating this into an ex post version of the CAPM would yield a single-index
model (following the same steps as we did above):

R̃i = αi + βi R̃M + ẽi , (6)

with the difference that it is predicted by the CAPM that αi = 0, since from (6) we get

αi = r̄i − βi r̄M − (1 − βi ) rf = [r̄i − rf ] − βi [r̄M − rf ] = 0 [using (5)].

R-squared

When running a regression on a sample of T periods of historical excess returns for security
i, Ri1 , . . . , RiT , using the empirical specification given in (6) almost every statistical program
will compute the regression R2 . The statistical measure has a nice economic interpretation in
this context. It quantifies the fraction of variance in excess returns of the individual security
i that can be attributed to unanticipated movements in the common factor:

variance explained
by common factor
z }| {
2
β i V ar[R̃ M]
R2 = (7)
V ar[R̃i ]
| {z }
total variance

1.3 Multi-factor models

Following the logic of the single-factor model we looked at, one can assert that several different
factors affect the rates of return on securities. For example, one might think that both the
state of the economy as measured by the unexpected change in GDP, ∆GDP
^ , and interest

4
rate surprises, ∆IR,
] capture “common” movements in security prices. This would lead to the

following (linear) two-factor model:

r̃i = E[r̃i ] + βi,GDP ∆GDP

^ + βi,IR ∆IR
] + ẽi , (8)

where E[∆GDP
^ ] = 0, E[∆IR]
] = 0. Moreover, the factors are mutually uncorrelated and also

uncorrelated with ei . βi,GDP and βi,IR are called the factor sensitivities or factor loadings.

Example 1 Consider the returns on the two following types of companies:

• a regulated utility company U that mainly supplies electric power to residential areas.

• an airline A.

The economy is doing unexpectedly well and the central bank reacts by hiking up interest rates to
hold off inflation. What signs do the realizations for the factors ∆GDP
^ and ∆IR,
] respectively,

take? What do you expect the impact of this to be on the two companies A and U?

• Utility company U

Residential demand for electricity is not very sensitive to business cycles, therefore
βU,GDP can be expected to be low.

Since the company faces a very stable demand it will likely have a very stable cash flow.
As a first approximation, think of the company generating a cash flow of X every year.
This makes U look like a perpetuity , the value of which is very sensitive to the interest
rate.1 If there is an unexpected change in the interest rate, all these future cash flows will
be discounted differently, having a big impact on the net present value of the company,
which the stock price should reflect. Therefore, one can expect βU,IR to be high.

⇒ utility companies are interest rate sensitive

(other examples: banks and other financial companies).

• Airline A

Airlines face a demand that is very sensitive to the business cycle, therefore one can
expect βA,GDP to be high. Also, A can be expected to be less sensitive to interest rates
than U.
⇒ airlines are sensitive to the business cycle (cyclical )
(other examples: automobile manufacturers, furniture, steel, paper,
heavy machinery, hotels and expensive restaurants).
1 P∞ −t X
The net present value of a perpetuity as a function of the interest rate r is N P V (r) = t=0 (1+r) X= r
.
We will cover this in more depth later in the course.

5
What we can conclude for our example is that the upsurge in economic activity linked with
an interest rate increase is going to be

• bad news for U since it’s returns react strongly to interest rates;

• good news for A since it reacts strongly to the business cycle.

Note, a single-factor model could not capture this distinction.

2 Risky Arbitrage: An example

Factor models are useful for constructing arbitrage portfolios that are (mainly) exposed to risk
from the relative fortunes of two securities that one thinks are not correctly priced in relation
to each other, as the following example illustrates.
Imagine McDonald’s created a new fat free burger that you just love. You find that Burger
Kings new fat free burger, on the other hand, is dry and tasteless. Sensing that others will
feel the same about these products as you do, you ponder how to bet on the relative success of
McDonald’s compared to Burger King, without exposing yourself to the risk of general market
movements.
If you believe that the two fast food chains are affected in the same way by common factors,
then buying, say $10,000 worth of McDonald’s stock and selling short $10,000 of Burger King
stock will leave you exposed only to the difference between the two companies’ firm-specific
movements. However, the strategy involves considerable risk from other factors affecting the
fortunes of both companies, that go beyond the relative success of the fat free burgers (e.g
in April 2004 Jim Cantalupo, then CEO of McDonald’s, died unexpectedly of a heart attack,
following which shares of the company plunged almost three percent - down 71 cents to $26.75
on the NYSE).
The above strategy, which is often labelled “market neutral”, is an example for risky ar-
bitrage. It is often employed by hedge funds who want to make bets on the relative prices
of financial assets which they believe to be inconsistent with economic fundamentals without
exposing themselves to other things happening in the markets.
The logic illustrated in the above example for protecting yourself against movements in com-
mon factors affecting security returns is going to be key for our next topic.

6
3 Arbitrage Pricing Theory (APT)

Stephen Ross (1976) builds on a multi-factor model to develop a theory of asset pricing. It is
based on the tenet that in a well functioning security market no arbitrage opportunities should
exist.
Specifically, the assumptions underlying the Arbitrage Pricing Theory (APT) are:

• security returns can be described by a linear factor model;

• there are sufficiently many securities available to diversify away any idiosyncratic risk;

• arbitrage opportunities do not exist.

3.1 Well diversified portfolios

The starting point of the APT is the idea that investors can construct portfolios in such a way
that they diversify away any idiosyncratic risk and are only left exposed to factor risk.
To illustrate the APT we will use a single-factor model (the arguments carry over to the
n-factor case):
r̃i = E[r̃i ] + βi F̃ + ẽi . (9)

Thus, the return on an n-security portfolio is

n
X n
X n
X
r̃P = wi E[r̃i ] + wi βi F̃ + wi ẽi
i=1 i=1 i=1
= E[r̃P ] + βP F̃ + ẽP . (10)

Computing the portfolio’s variance we obtain

σP2 = β 2 σ2 + σe2P . (11)

| P{z F} |{z}
systematic risk nonsystematic risk

Note that we used the assumption that the factor and the idiosyncratic noise terms are mu-
tually uncorrelated. If we look at an equally weighted portfolio, i.e. with wi = 1/n, the
nonsystematic risk component is given by
" n # " n #
X1 1 X
2
σeP = V ar ẽi = 2 V ar ẽi
n n
i=1 i=1
n
1 X
= V ar[ei ] (because e˜i s are uncorrelated with each other and F̃ )
n2
i=1
n
1 1 X
= V ar[ei ]
n n
i=1
| {z }
=σ̄e2

7
 2 2
ri
βwd=1 βi=1
1.5 1.5

idiosyncratic

expected return

expected return
1 1 shock ei
● common
shock F
● r
0.5 0.5
●
●
●
0 0
-0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 F -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 F

-0.5 -0.5

Realized returns: well diversified portfolio Realized returns: individual securities

Figure 1: Returns as a function of the systematic factor F

Hence, increasing the number n of different assets in the portfolio reduces more and more the
nonsystematic risk – exposure to each asset’s specific risk becomes smaller and smaller. In the
limit as n approaches infinity there is only systematic risk left since

1 2
lim σe2P = lim σ̄ → 0. (12)
n→∞ n→∞ n e

Such a portfolio is called well diversified (which we will denote by subscript wd) and has a
return that depends only on the systematic risk factor:2
X X
r̃wd = wi E[r̃i ] + wi βi F̃ . (13)
i i

Figure 1 illustrates this. The realized returns on a well diversified portfolio fall exactly on the
solid line. In contrast, realized returns for individual securities with the same factor loading
will typically not be on this line. This reflects the idiosyncratic shocks that they are exposed
to.

3.2 No arbitrage

An arbitrage opportunity exists if an investor can earn profits without exposing him or herself
to risk and making any net investment.
2
Note that theoretically such well diversified portfolios require an infinite number of securities to be available.
For all practical purposes a few hundred or thousand securities might be enough to reduce the remaining
nonsystematic risk to a negligible amount.

8
Example 2 IBM stock is sold for $ 60 on the NYSE and simultaneously trades at $ 58 on
NASDAQ. This offers a risk-free opportunity to earn money by at the same time buying IBM
stock on NASDAQ and selling IBM stock on the NYSE.

In a well functioning capital market such opportunities should not exist since market par-
ticipants would instantaneously take advantage of them and thus move prices so that they
disappear. One consequence of this tenet is the Law of one price. It states that two assets
which are equivalent in all economically relevant aspects must have the same market price.
No-arbitrage conditions such as this one underpin much of modern finance theory.
Let us turn to the implications of no-arbitrage in the context of the APT.

Example 3 Suppose that we have the following situation: well diversified portfolios A and B
have the same factor loading, i.e. βA = βB . However, E[r̃A ] = 0.10 and E[r̃B ] = 0.08.
Is this consistent with no arbitrage?

No, an arbitrage opportunity arises. A market participant can construct an arbitrage port-
folio as follows:

position return on position

long position in A E[r̃A ] + βA F̃
short position in B -E[r̃B ] − βB F̃
net proceeds E[r̃A ] − E[r̃B ]

The exposure to factor risk cancels out: βA F̃ −βB F̃ = 0. For example, buying £ 1 million of A
and selling short £ 1 million of B would yield a profit of £ 20,000 (= (0.1−0.08)×£ 1, 000, 000).
Note that this involves zero net investment and is risk-free.
The no-arbitrage condition thus implies that well-diversified portfolios with equal factor load-
ings must have equal expected returns. This means that they plot on a straight line in the
expected return - beta diagram.
The next example illustrates the repercussions of no-arbitrage for well-diversified portfolios
with different factor loadings.

Example 4 Suppose that the risk-free rate is 4 percent and that we have the following situation
for well diversified portfolios A and C:

portfolio beta expected return

A 1 0.14
C 0.5 0.08

9
 0.15
A

expected return
0.1 D
0.09
0.08
C

0.05

rf

0
-0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
be ta

Figure 2: Returns as a function of the systematic factor F

Is this consistent with no arbitrage?

Figure 2 depicts the situation. The expected return on C means that there is an arbitrage
opportunity. We can construct a portfolio that has the same beta as C and a higher return,
which we have already seen violates no arbitrage. D is such an arbitrage portfolio, it is a
portfolio invested half in the risk-free asset (with β = 0) and half in the well diversified
portfolio A. Thus βD = 0.5 · 0 + 0.5 · 1 = 0.5 and E[r̃D ] = 0.5 rf + 0.5 E[r̃A ] = 0.09.
Hence, no-arbitrage implies that well-diversified portfolios with different factor loadings must
have expected returns proportional to their betas. This means that they plot on a straight
line in the expected return - beta diagram.

3.3 Finding the factors

The CAPM is a theory that rationalizes a single-factor model. However, multi-factor models
(and the APT) do not guide us in the choice of the relevant factors. We will only briefly
mention the problems involved in this. One approach is to use economic theory to come
up with the most relevant factors. However, these factors do not always do well empirically
in explaining variation in security prices. Another approach is to use statistical methods to
extract those factors which explain most of the variation in security returns. The problem
with this approach is that the extracted factors often do not have a meaningful economic
interpretation.

10
References

Bodie, Zvi, Alex Kane, and Alan J. Marcus, 2008, Investments (Irwin McGraw-Hill: Chicago).

Grinblatt, Mark, and Sheridan Titman, 2002, Financial Markets and Corporate Strategy
(Boston, Mass.: Irwin McGraw-Hill).

Levy, Haim, and Thierry Post, 2005, Investments (FT Prentice Hall: London).

Ross, Stephen, 1976, The arbitrage theory of capital asset pricing, Journal of Economic Theory
13, 341–360.

11