You are on page 1of 10

Chapter 13 : The Arbitrage Pricing Theory

13.1

Introduction

We have made two first attempts (Chapters 10 to 12) at asset pricing from
an arbitrage perspective, that is, without specifying a complete equilibrium
structure. Here we try again from a different, more empirically based angle.
Let us first collect a few thoughts as to the differences between an arbitrage
approach and equilibrium modeling.
In the context of general equilibrium theory, we make hypotheses about
agents consumers, producers, investors; in particular, we start with some form
of rationality hypothesis leading to the specification of maximization problems
under constraints. We also make hypotheses about markets: Typically we assume that supply equals demand in all markets under consideration.
We have repeatedly used the fact that at general equilibrium with fully
informed optimizing agents, there can be no arbitrage opportunities, in other
words, no possibilities to make money risklessly at zero cost. An arbitrage
opportunity indeed implies that at least one agent can reach a higher level of
utility without violating his/her budget constraint (since there is no extra cost).
In particular, our assertion that one can price any asset (income stream) from
the knowledge of Arrow-Debreu prices relied implicitly on a no-arbitrage hypothesis: with a complete set of Arrow-Debreu securities, it is possible to replicate
any given income stream and hence the value of a given income stream, the
price paid on the market for the corresponding asset, cannot be different from
the value of the replicating portfolio of Arrow-Debreu securities. Otherwise an
arbitrageur could make arbitrarily large profits by short selling large quantities
of the more expensive of the two and buying the cheaper in equivalent amount.
Such an arbitrage would have zero cost and be riskless.
While general equilibrium implies the no-arbitrage condition, it is more restrictive in the sense of imposing a heavier structure on modeling. And the
reverse implication is not true: No arbitrage opportunities1 the fact that all
arbitrage opportunities have been exploited does not imply that a general
equilibrium in all markets has been obtained. Nevertheless, or precisely for that
reason, it is interesting to see how far one can go in exploiting the less restrictive
hypothesis that no arbitrage opportunities are left unexploited.
The underlying logic of the APT to be reviewed in this chapter is, in a sense,
very similar to the fundamental logic of the Arrow-Debreu model and it is very
much in the spirit of a complete market structure. It distinguishes itself in two
major ways: First it replaces the underlying structure based on fundamental
securities paying exclusively in a given state of nature with other fundamental
securities exclusively remunerating some form of risk taking. More precisely, the
APT abandons the analytically powerful, but empirically cumbersome, concept
of states of nature as the basis for the definition of its primitive securities. It
1
An arbitrage portfolio is a self-financing (zero net-investment) portfolio. An arbitrage
opportunity exists if an arbitrage portfolio exists that yields non-negative cash flows in all
states of nature and positive cash flows in some states (Chapter 11).

replaces it with the hypothesis that there exists a (stable) set of factors that are
essential and exhaustive determinants of all asset returns. The primitive security will then be defined as a security whose risk is exclusively determined by
its association with one specific risk factor and totally immune from association
with any other risk factor. The other difference with the Arrow-Debreu pricing
of Chapter 8 is that the prices of the fundamental securities are not derived from
primitives supply and demand, themselves resulting from agents endowments
and preferences but will be deduced empirically from observed asset returns
without attempting to explain them. Once the price of each fundamental security has been inferred from observed return distributions, the usual arbitrage
argument applied to complex securities will be made (in the spirit of Chapter
10).2

13.2

Factor Models

The main building block of the APT is a factor model, also known as a returngenerating process. As discussed previously, this is the structure that is to
replace the concept of states of nature. The motivation has been evoked before:
States of nature are analytically convincing and powerful objects. In practice,
however, they are difficult to work with and, moreover, often not verifiable,
implying that contracts cannot necessarily be written contingent on a
specific state of nature. We discussed these shortcomings of the ArrowDebreu pric- ing theory in Chapter 8. The temptation is thus irresistible to
attack the asset pricing problem from the opposite angle and build the concept
of primitive secu- rities on an empirically more operational notion, abstracting
from its potential theoretical credentials. This structure is what factor models
are for.
The simplest conceivable factor model is a one-factor market model, usually
labeled the Market Model, which asserts that ex-post returns on individual assets
can be entirely ascribed either to their own specific stochastic components or to
their common association in a single factor, which in the CAPM world would
naturally be selected as the return on the market portfolio. This simple factor
model can thus be summarized by following the equation (or process):3
rj = j + j rM + j ,

(13.1)

with Ej = 0, cov (rM ,j ) = 0, j, and cov (j ,k ) =


0, j = k.
This model states that there are three components in individual returns: (1)
an asset-specific constant j ; (2) a common influence, in this case the unique
factor, the return on the market, which affects all assets in varying degrees,
with j measuring the sensitivity of asset js return to fluctuations in the market return; and (3) an asset-specific stochastic term j summarizing all
other stochastic components of rj that are unique to asset j.
2
The arbitrage pricing theory was first developed by Ross (1976), and substantially interpreted by Huberman (1982) and Conner (1984) among others. For a presentation emphasizing
practical applications, see Burmeister et al. (1994).
3 Factors are frequently measured as deviations from their mean. When this is the case,
j
becomes an estimate of the mean return on asset j.

Equation (13.1) has no bite (such an equation can always be written) until
one adds the hypothesis cov (j ,k ) = 0, j = k, which signifies that all
return characteristics common to different assets are subsumed in their link
with the market return. If this were empirically verified, the CAPM would be
the undis- puted end point of asset pricing. At an empirical level, one may
say that it is quite unlikely that a single factor model will suffice.4 But the
strength of the APT is that it is agnostic as to the number of underlying factors
(and their iden- tity). As we increase the number of factors, hoping that this
will not require a number too large to be operational, a generalization of
Equation (13.1) becomes more and more plausible. But let us for the moment
maintain the hypothesis of one common factor for pedagogical purposes.5
13.2.1 About the Market Model
Besides serving as a potential basis for the APT, the Market Model, despite all
its weaknesses, is also of interest on two grounds. First it produces estimates for
the s that play a central role in the CAPM. Note, however, that estimating s
from past data alone is useful only to the extent that some degree of stationarity
in the relationship between asset returns and the return on the market is present.
Empirical observations suggest a fair amount of stationarity is plausible at the
level of portfolios, but not of individual assets. On the other hand, estimating
the s does not require all the assumptions of the Market Model; in particular,
a violation of the cov(i , k ) = 0, i = k hypothesis is not damaging.
The second source of interest in the Market Model, crucially dependent on
the latter hypothesis being approximately valid, is that it permits economizing
on the computation of the matrix of variances and covariances of asset returns
at the heart of the MPT. Indeed, under the Market Model hypothesis, one can
write (you are invited to prove these statements):

2 = 2 2 + 2 , j
j
j M
j
ij

2
i j M

This effectively means that the information requirements for the implementation of MPT can be substantially weakened. Suppose there are N risky assets under consideration. In that case the computation of the efficient fron2
tier requires knowledge of N expected returns, N variances, and N 2N covari2
ance terms (N is the total number of entries in the matrix of variances and
covariances, take away the N variance/diagonal terms and divide by 2 since
ij = ji , i, j).
4
Recall
5

the difficulty in constructing the empirical counterpart of M .


Fama (1973), however, demonstrates that in its form (13.1) the Market Model is inconsistent in the following sense: the fact that the market is, by definition, the collection of all individual assets implies an exact linear relationship between the disturbances j ; in other words,
when the single factor is interpreted to be the market the hypothesis cov (j ,k ) = 0, j
= k cannot be strictly valid.
While we ignore this criticism in view of our purely
pedagogical objective, it is a fact that if a single factor model had a chance to be
empirically verified (in the sense of all the assumptions in (13.1) being confirmed) the unique
factor could not be the market.

Working via the Market Model, on the other hand, requires estimating Equation (13.1) for the N risky returns producing estimations for the N j s and
2
the N j
and estimating the variance of the market return, that is, 2N + 1
information items.

13.3
13.3.1

The APT: Statement and Proof


A Quasi-Complete Market Hypothesis

To a return-generating process such as the Market Model, the APT superposes


a second major hypothesis that is akin to assuming that the markets are quasicomplete. What is needed is the existence of a rich market structure with a
large number of assets with different characteristics and a minimum number of
trading restrictions. This market structure, in particular, makes it possible to
form a portfolio P with the following three properties:
Property 1: P has zero cost; in other words, it requires no investment.
This is the first requirement of an arbitrage portfolio.
Let us denote xi as the value of the position in the ith asset in portfolio P .
Portfolio P is then fully described by the vector xT = (x1 , x2 ,..., xN ) and
the zero cost condition becomes
N
X

xi = 0 = x
1,

i=1

with 1 the (column) vector of 1s. (Positive positions in some assets must be
financed by short sales of others.)
Property 2: P has zero sensitivity (zero beta) to the common factor:6
N
X

xi i = 0 = x
.

Property 3: P is a well-diversified portfolio. The specific risk of P is (almost)


N
X
totally eliminated:
2
x2
=
i i
0.
i

The APT builds on the assumed existence of such a portfolio, which requires a
rich market structure.
6

Remember that the beta of a portfolio is the weighted sum of the betas of the assets in
the portfolio.

13.3.2 Statement and Proof of the APT


The APT relationship is the direct consequence of the factor structure hypothesis, the existence of a portfolio P satisfying these conditions, and the
no-arbitrage assumption. Given that returns have the structure of Equation
(13.1), Properties 2 and 3 imply that P is riskless. The fact that P has zero
cost (Property 1) then entails that an arbitrage opportunity will exist unless:
r P = 0 = xT r

(13.2)

The APT theorem states, as a consequence of this succession of statements, that


there must exist scalars 0 , 1 , such
that:
r = 0 1 + 1 , or
r i = 0 + 1 i for all assets i

(13.3)

This is the main equation of the APT.


Equation (13.3) and Properties 1 and 2 are statements about 4 vectors: x,
, 1, and r.
Property 1 states that x is orthogonal to 1. Property 2 asserts that x is
orthogonal to . Together these statements imply a geometric configuration
that we can easily visualize if we fix the number of risky assets at N = 2, which
implies that all vectors have dimension 2. This is illustrated in Figure 13.1.
Insert Figure 13.1
Equation (13.3) no arbitrage implies that x and r are orthogonal.
But this means that the vector r must lie in the plane formed by 1 and , or,
that r can be written as a linear combination of 1 and , as Equation (13.3)
asserts.
More generally, one can deduce from the triplet
N
X

xi =

N
X

NX

xi i =

xi ri =0

that there exist scalars 0 , 1 , such that:


ri = 0 + 1 i for all
i.
This is a consequence of the orthonormal projection of the vector ri into
the subspace spanned by the other two.
13.3.3

Meaning of 0 and 1

Suppose that there exists a risk-free asset or, alternatively, that the sufficiently
rich market structure hypothesis permits constructing a fully diversified portfolio with zero-sensitivity to the common factor (but positive investment). Then
rf = rf = 0
.

That is, 0 is the return on the risk-free asset or the risk-free portfolio.
Now let us compose a portfolio Q with unitary sensitivity to the common
factor = 1. Then applying the APT relation, one gets:
rQ = rf + 1
1
Thus, 1 = rQ rf , the excess-return on the pure-factor portfolio Q. It is
now possible to rewrite equation (13.3) as:
ri = rf + i (rQ rf ) .
(13.4) If, as we have assumed, the unique common factor is the return on the
market
portfolio, in which case Q = M and rQ rM , then Equation (13.4) is
simply the CAPM equation:
ri = rf + i (rM
rf ) .

13.4

Multifactor Models and the APT

The APT approach is generalizable to any number of factors. It does not,


however, provide any clue as to what these factors should be, or any particular
indication as to how they should be selected. This is both its strength and its
weakness. Suppose we can agree on a two-factor model:
(13.5)
+ bj2F2 +
rj = aj + bj1 F1
ej

j
with Ee = 0, cov F , = cov F , = 0, j, and cov ( , ) = 0, j =
k.

As was the case for Equation (13.1), Equation (13.5) implies that one cannot
reject, empirically, the hypothesis that the ex-post return on an asset j has two
stochastic components: one specific, (ej ), and one systematic, (bj1 1 + bj2F2 ).
F
What is new is that the systematic component is not viewed as the result of
a single common factor influencing all assets. Common or systematic issues
may now be traced to two fundamental factors affecting, in varying degrees, the
returns on individual assets (and thus on portfolios as well). Without loss of
generality we may assume that these factors are uncorrelated.
As before, an expression such as Equation (13.5) is useful only to the extent
that it describes a relationship that is relatively stable over time. The two
factors F1 and F2 must really summarize all that is common in individual asset
returns.
What could these fundamental factors be? In an important article, Chen,
Roll, and Ross (1986) propose that the systematic forces influencing returns
must be those affecting discount factors and expected cash flows. They then isolate a set of candidates such as industrial production, expected and unexpected
inflation, measures of the risk premium and the term structure, and even oil
prices. At the end, they conclude that the most significant determinants of asset returns are industrial production (affecting cash flow expectations), changes
in the risk premium measured as the spread between the yields on low- and
high-risk corporate bonds (witnessing changes in the market risk appetite), and

twists in the yield curve, as measured by the spread between short- and longterm interest rates (representing movements in the market rate of impatience).
Measures of unanticipated inflation and changes in expected inflation also play
a (less important) role.
Let us follow, in a simplified way, Chen, Roll, and Rosss lead and decide that
our two factors are industrial production (F1 ) and changes in the risk premium
(F2 ). How would we go about implementing the APT? First we have to measure
our two factors. Let I P (t) denote the rate of industrial production in month
t; then M P (t) = log I P (t) log I P (t 1) is the monthly growth rate of I
P . This is our first explanatory variable.
To measure changes in the risk premium, let us define
U P R(t) = Baa and under bond portfolio return(t) LGB(t)
where LGB(t) is the return on a portfolio of long-term government bonds. With
these definitions we can rewrite Equation (13.5) as
rjt = aj + bj1 M P (t) + bj2 U P R (t) + ej
The bjk , k = 1, 2, are often called factor loadings. They can be estimated
di- rectly by multivariate regression. Alternatively, one could construct pure
factor portfolios well-diversified portfolios mimicking the underlying factors
and compute their correlation with asset j. The pure factor portfolio P1
would be
a portfolio with bP1 = 1 and bP2 = eP1 = 0; portfolio P2 would be defined
similarly to track the stochastic behavior of U P R(t). Let us go on
hypothesizing (wrongly according to Chen, Roll, and Ross) that this two-factor model
satisfies the necessary assumptions (cov(ei , ej ) = 0, i = j) and further
assume the existence of a risk-free portfolio Pf with zero sensitivity to either of
our two factors and zero specific risk. Then the APT states that there exist
scalars 0 , 1 , 2 such that:
r j = 0 + 1 bj1 + 2 bj2
.
That is, the expected return on an arbitrary asset j is perfectly and completely
described by a linear function of asset js factor loadings bj1 , bj2 . This can
appropriately be viewed as a (two-factor) generalization of the SML.
Furthermore the coefficients of the linear function are:
0

rf

r P 1 rf

r P 2 rf

where P1 and P2 are our pure factor portfolios.


The APT agrees with the CAPM that the risk premium on an asset, r j 0 ,
is not a function of its specific or diversifiable risk. It potentially disagrees with
the CAPM in the identification of the systematic risk. The APT decomposes
the systematic risk into elements of risk associated with a particular assets
sensitivity to a few fundamental common factors.

Note the parallelism with the Arrow-Debreu pricing approach. In both contexts, every individual asset or portfolio can be viewed as a complex security,
or a combination of primitive securities: Arrow-Debreu securities in one case,
the pure factor portfolios in the other. Once the prices of the primitive securities are known, it is a simple step to compose replicating portfolios and, by a
no-arbitrage argument, price complex securities and arbitrary cash flows. The
difference, of course, resides in the identification of the primitive security. While
the Arrow-Debreu approach sticks to the conceptually clear notion of states of
nature, the APT takes the position that there exist a few common and stable
sources of risk and that they can be empirically identified. Once the corresponding risk premia are identified, by observing the market-determined premia on
the primitive securities (the portfolios with unit sensitivity to a particular factor
and zero sensitivity to all others) the pricing machinery can be put to work.
Let us illustrate. In our two-factor examples, a security j with, say, bj1 =
0.8 and bj2 = 0.4 is like a portfolio with proportions of 0.8 of the pure
portfolio P1 , 0.4 of pure portfolio P2 , and consequently proportion 0.2 in
the riskless asset. By our usual (no-arbitrage) argument, the expected rate
of return on that security must be:
rj

= 0.2rf + 0.8r P1 + 0.4r P2


= 0.2rf + 0.8rf + 0.4rf + 0.8 (r P1 rf ) + 0.4 (r P2 rf )
= rf + 0.8 (r P1 rf ) + 0.4 (r P2 rf )
=

0 + bj1 1 + bj2 2

The APT equation can thus be seen as the immediate consequence of the linkage between pure factor portfolios and complex securities in an arbitrage-free
context. The reasoning is directly analogous to our derivation of the value additivity theorem in Chapter 10 and leads to a similar result: Diversifiable risk
is not priced in a complete (or quasi-complete) market world.
While potentially more general, the APT does not necessarily contradict
the CAPM. That is, it may simply provide another, more disaggregated, way
of writing the expected return premium associated with systematic risk, and
thus a decomposition of the latter in terms of its fundamental elements. Clearly
the two theories have the same implications if (keeping with our two-factor
model, the generalization is trivial):
j (r M rf ) = bj1 (r P1 rf ) + bj2 (r P2 rf )
(13.6) Let P1 be the (market) beta of the pure portfolio P1 and similarly
for P2 .
Then if the CAPM is valid, not only is the LHS of Equation (13.6) the expected
risk premium on asset j, but we also have:
r P1 rf
r P2 rf

= P1 (r M
rf )
= P2 (r M rf )

Thus the APT expected risk premium may be written as:


bj1 [P1 (r M rf )] + bj2 [P2 (r M r f )] = (bj1 P1 + bj2 P2 ) (r M
rf )

which is the CAPM equation provided:


j = bj1 P1 + bj2 P2
In other words, CAPM and APT have identical implications if the sensitivity of
an arbitrary asset j with the market portfolio fully summarizes its relationship
with the two underlying common factors. In that case, the CAPM would be
another, more synthetic, way of writing the APT.7
In reality, of course, there are reasons to think that the APT with an arbitrary number of factors will always do at least as well in identifying the sources
of systematic risk as the CAPM. And indeed Chen, Roll, and Ross observe that
their five factors cover the market return in the sense that adding the return on
the market to their preselected five factors does not help in explaining expected
returns on individual assets.

13.5

Advantage of the APT for Stock or Portfolio Selection

The APT helps to identify the sources of systematic risk, or to split systematic
risk into its fundamental components. It can thus serve as a tool for helping the
portfolio manager modulate his risk exposure. For example, studies show that,
among U.S. stocks, the stocks of chemical companies are much more sensitive
to short-term inflation risk than stocks of electrical companies. This would be
compatible with both having the same exposure to variations in the market
return (same beta). Such information can be useful in at least two ways. When
managing the portfolio of an economic agent whose natural position is very
sensitive to short-term inflation risk, chemical stocks may be a lot less attractive
than electricals, all other things equal (even though they may both have the
same market beta). Second, conditional expectations, or accurate predictions,
on short-term inflation may be a lot easier to achieve than predictions of the
markets return. Such a refining of the information requirements needed to take
aggressive positions can, in that context, be of great use.

13.6

Conclusions

We have now completed our review of asset pricing theories. At this stage it may
be useful to draw a final distinction between the equilibrium theories covered in
Chapters 7, 8, and 9 and the theories based on arbitrage such as the Martingale
pricing theory and the APT. Equilibrium theories aim at providing a complete
theory of value on the basis of primitives : preferences, technology, and
market structure. They are inevitably heavier, but their weight is proportional
to their ambition. By contrast, arbitrage-based theories can only provide a
relative theory of value. With what may be viewed as a minimum of
assumptions, they
7
The observation in footnote 5, however, suggests this could be true as an approximation
only.

offer bounds on option values as a function of the price of the


underly- ing asset, the stochastic behavior of the latter being taken as
given (and unexplained);

permit estimating the value of arbitrary cash flows or securities using


risk- neutral measures extracted from the market prices of a set of
fundamental securities, or in the same vein, using Arrow-Debreu prices
extracted from a complete set of complex securities prices;

explain expected returns on any asset or cash flow stream once the price
of risk associated with pure factor portfolios has been estimated from
market data on the basis of a postulated return-generating process.
Arbitrage-based theories currently have the upper hand in practitioners circles where their popularity far outstrips the degree of acceptance of equilibrium
theories. This, possibly temporary, state of affairs may be interpreted as a
measure of our ignorance and the resulting need to restrain our ambitions.
References
Burmeister, E., Roll, R., Ross, S.A. (1994), A Practitioners Guide to Arbitrage Pricing Theory, in A Practitioners Guide to Factor Models, Research Foundation of the Institute of Chartered Financial Analysts, Charlottesville, VA.
Chen, N. F., Roll R., Ross, S.A. (1986), Economic Forces and the Stock
Market, Journal of Business 59(3), 383-404.
Connor, G. (1984), A Unified Beta Pricing Theory, Journal of Economic
Theory 34(1).
Fama, E.F. (1973), A Note on the Market Model and the Two-Parameter
Model, Journal of Finance 28 (5), 1181-1185
Huberman, G. (1982), A Simple Approach to Arbitrage Pricing, Journal of
Economic Theory, 28 (1982): 183191.
Ross, S. A. (1976), The Arbitrage Pricing Theory, Journal of Economic
Theory, 1, 341360.