Jessica Wachter Class notes for Finance 911

IV Arbitrage Pricing Theory

1. Linear Factor Model

2. An Economy with 1 Factor and No Residual Risk

3. An Economy with K Factors and No Residual Risk

4. An Economy with K Factors and Residual Risk

Recall CAPM: An equilibrium theory that implied that expected excess returns on
assets depend on the covariance with the market. The reason that some company has a
higher return on average than another company is that it co-moves more with the
market.
Stated differently, excess expected return was a reward for bearing a certain kind of risk
- the risk associated with the market.
But to derive this result, we had to make some very specific assumption, namely
quadratic utility, normal returns or two fund separation.
The idea here is that we are going to make less restrictive assumptions, but hope that
we still get something interesting.

1 Linear Factor Model

Note: this is a statistical model for asset returns, not an economic model.
Assume

• n risky securities with returns r̃i

• 1 riskless security rf

• ∃ai bik and portfolios with returns δ̃k such that

r̃i = ai + ΣK
k=1 δ̃k bik + 
˜i K<n

where
h i
E[˜i ] = E ˜i δ̃k = 0 (i)
E[˜i ˜j ] = 0 (ii)

1
Jessica Wachter Class notes for Finance 911

Note: we can always find coefficients such that (i) is true by linear projection. (ii),
however, is restrictive.
The content here is that a small number of portfolios captures all the common variation
in the assets. We can relax the assumption that the δ̃k s are portfolios at the cost of
making the proof more complicated.

Differences between this statement and the CAPM

The CAPM says:

E r̃i = rf + βi (E r̃m − rf )
where
Cov(r̃i , r̃m )
βi = .
Var(r̃m )
This implies that r̃i can be written as

r̃i = (1 − βi )rf + βi r̃m + ˜i ,

for
E˜i = 0.
Note that the definition of βi implies that

E[r̃m ˜i ] = 0.

How is this different from a LFM?

• Most obviously: The equation for the CAPM has only one factor, and the CAPM
tells us what it is – the market portfolio (namely δ̃ = r̃m ).

• Also, the CAPM tells us what the intercept of this regression is: ai = (1 − βi )rf .

• Does this mean that the CAPM is a special case of the LFM? No. The CAPM
does not require (or imply) that E[˜i ˜j ] = 0.

In what follows, we will assume different forms of the linear factor model.
We will also introduce some economics through the assumption of no arbitrage.
Definition. An arbitrage opportunity is a trading strategy with

zero initial investment

2
Jessica Wachter Class notes for Finance 911

non-negative payoff in all future states of the world and positive payoff in at least one
future state.

An arbitrage opportunity is sometimes called a “free lunch”.

Note that any investor with u0 > 0 would take an arbitrage opportunity. In other
words, free lunch ⇒ no equilibrium. So ruling out arbitrage is a very weak assumption.

Aside: Short-selling versus Buying Because an arbitrage portfolio involves zero

initial investment and is riskfree, it usually requires that one takes a “short” position in
an asset. In that case, one earns the negative rate of return.
Let’s dig into this statement a bit, just to make sure we understand what a zero-cost
portfolio means.
Suppose you were to invest $1 in asset i. Let’s say asset i has price Pi0 when you buy
it, and asset P̃i at the end of the period (we continue to assume a one-period problem
and, for the moment, either ignore dividends or assume they are part of P̃i ). Your net
return is the rate of change in your investment:

P̃i − Pi0
r̃i =
Pi0
Note that your gross return is the amount you have in asset i at the end of the period,
per $1 invested:
P̃i
1 + r̃i =
Pi0

Suppose you wanted to finance this purchase by investing in the riskfree asset. In that
case, you would borrow $1 (say), and pay it back with interest rf . You could say that
your net return is rf . The total amount you owe is −(1 + rf ). If you do this for every
dollar you invest in asset i, you have a zero-cost portfolio, and your payoff, per $1
invested is:
Payoff = (1 + r̃i ) − (1 + rf ) = r̃i − rf
namely, what you receive minus what you owe. The return on this strategy is undefined
because your initial investment is zero.
Now suppose that instead of financing the purchase of asset i by borrowing at the
riskfree rate, you finance the purchase with a short-sale of asset j. When you sell an
asset short, you borrow the asset, sell it, and then buy it back (returning it to
whomever you borrowed it from) at the end of the period. If you do this for one share

3
Jessica Wachter Class notes for Finance 911

of asset j, then your cash inflow is P0,j (note that it is positive; short-selling is a form of
borrowing), and your outflow is P̃j at the end of the period. If you execute this strategy
for $1 rather than P0,j , your inflow is $1 at the start of the period, and
−P̃j /P0,j = −(1 + r̃j ) at the end of the period. Because prices can be arbitrarily high,
short-selling is risky. It is not a limited liability activity.
Your payoff per dollar, if you buy asset i and short-sell an equal dollar amount of asset
j is:
Payoff = (1 + r̃i ) − (1 + r̃j ) = r̃i − r̃j .

2 An Economy with 1 Factor and No Residual Risk

Assumption:
r̃i = ai + bi δ̃ ∀i
(Note that we are setting the ˜ equal to zero.)
Claim: no arbitrage ⇒ ai = (1 − bi )rf .

Proof. • Assume that ai > (1 − bi )rf .

Because δ̃ is a portfolio, we can trade it.
Consider the following portfolios:

1. Asset i
2. weight bi in δ̃ and weight (1 − bi ) in riskfree asset.

Strategy: buy $1 of (1) and short-sell $1 of (2). Note that the initial investment is
zero.
Payoff:

= $1(1 + ai + bi δ̃) − $1(bi (1 + δ̃) + (1 − bi )(1 + rf ))

= $1(ai + bi δ̃ − bi δ̃ − (1 − bi )rf )
= $1(ai − (1 − bi )rf ) > 0

Portfolio costs nothing and always makes money. Hence it is an arbitrage.

• What if ai < (1 − bi )rf ?

Strategy: buy $1 of (2) and short-sell $1 of (1). Note that the initial investment is
zero.

4
Jessica Wachter Class notes for Finance 911

Return: $1((1 − bi )rf − ai ) > 0 ⇒ arbitrage.

Note the form of this proof:

To find the arbitrage, you need two portfolios that have exactly the same risk. Relative
to one another, one of them is a “good” investment and one is a “bad” investment.
Then buy the good investment and short the bad investment.
Also note that we could have put in $1000 or $100,000. The cost would still be nothing
and the payoff would still be certain.
This idea is what makes arbitrage so powerful. Not only do all investors with u0 > 0 like
arbitrage, they will also want to put an infinite amount of money into the arbitrage
portfolio.

3 An Economy with K Factors and No Residual Risk

Assumption:
K
X
r̃i = ai + bik δ̃k ∀i
k=1
PK
Claim: no arbitrage ⇒ ai = (1 − k=1 bik )rf .

Proof. Consider the following portfolios:

1. Asset i
PK
2. weight bik in δ̃k and weight (1 − k=1 bik ) in riskfree asset

PK
• If ai > (1 − k=1 bik )rf ,
buy $1 of (1) and short-sell $1 of (2). Payoff:
X X X X
ai + bik δ̃k − (1 − bik )rf − bik δ̃k = ai − (1 − bik )rf > 0
k k k k

Note that the risky parts cancel ⇒ Arbitrage.

• If ai < (1 − K
P
k=1 bik )rf , buy $1 of (2) and short-sell $1 of (1). ⇒ Arbitrage.

5
Jessica Wachter Class notes for Finance 911

Implication: !
X
E r̃i = 1− bik rf + bi1 E δ̃1 + · · · + bik E δ̃k
k

E[r̃i − rf ] = bi1 E[δ̃1 − rf ] + · · · + bik E[δ̃k − rf ]

The expected excess return of an asset is equal to the sum of its loading on each source
of risk multiplied by the reward from bearing that risk.

4 An Economy with K Factors and Residual Risk

Sections 2 and 3 presented very special cases. Now we will get to what is clearly the
most interesting case, namely ˜i not equal to zero.
Here we can not use the same arbitrage arguments, because we have ˜i , which are
impossible to get rid of.
Idea: ai ≈ (1 − bik )rf for most assets, as long as there are no trading strategy close to
arbitrage.
Of course, we have to define what we mean by the approximation, what we mean by
most and what we mean by close to arbitrage.
This was done in 1976 by Stephen Ross.
Consider a sequence of economies indexed by n = number of risky assets in the
economy.
Definition. An asymptotic arbitrage opportunity is a sequence of trading strategies
with payoff ỹ n such that
1. ỹ n requires zero initial investment ∀n.
2. There exists a constant J > 0 such that E ỹ n > J for all n and

lim Var(ỹ n ) = 0
n→∞

We will assume a linear factor model for each of the n economies. We also have to
assume the following:

Assumption ∃σ̄ such that Var(˜ni ) ≤ σ̄ 2 ∀n.

6
Jessica Wachter Class notes for Finance 911

Theorem. No asymptotic arbitrage ⇒ ∀ζ, ∃N̄ such that

X
|ani − (1 − bnik )rfn | ≥ ζ (*)

for at most N̄ assets.

Note:

• The theorem says that at most N̄ assets are mispriced by the amount ζ.
• N̄ is regardless of n. Thus the fraction of assets that are mispriced falls as n → ∞.
• We have complete freedom to choose ζ. If we choose ζ small then the bound will
be tighter and it will be violated by more assets (N̄ will be bigger).

Proof. Let M (n) be the number of assets s.t. ? holds in the nth economy. WLOG,
assume they are the first M (n) assets:
X
|ani − (1 − bnik )rfn | ≥ ζ for i = 1, ..., M (n)

Suppose by contradiction /∃ N̄ s.t. M (n) < N̄ ∀n. Then there exists a subsequence of
{n}, denote {nl } such that M (nl ) → ∞ as nl → ∞.
Consider the nl th economy. An asset satisfying ? must have either
K
!
X
ani l > 1 − bnikl rf
k=1

or !
K
X
ani l < 1− bnikl rf .
k=1

Recall the proof in the case of no residual risk. We will use the same argument as
before: we adopt appropriate long/short positions. Except in this case, we spread out
the $1 among all the M (nl ) assets.

• If
K
X
ani l > (1 − bnikl )rf ⇒
k=1
1 1
invest $ M (nl)
in asset i (“good asset”); short $ M (nl)
of the portfolio with weights
P
bik in δk , (1 − bik ) in rf (“bad asset”).

7
Jessica Wachter Class notes for Finance 911

• If
K
X
ani l < (1 − bnikl )rf ⇒
k=1
1 1
short $ M (n in asset i; invest $ M (n of the portfolio with weights bik in δ̃k ,
P l) l)
(1 − bik ) in rf .

Let
+1 if ani l > (1 − K nl
 P
Sinl = k=1 bik )rf
−1 if ani l < (1 − K nl
P
k=1 bik )rf

Payoff:

K K
1 X n X n X n
[(ani l + bikl δ̃k + ˜ni l ) − ((1 − bikl )rf + bikl δ̃k )]Sinl
M (nl ) k=1 k=1
1 X n
= [|ani l − (1 − bikl )rf | + ˜ni l Sinl ]
M (nl )

Each of these portfolios have zero cost. So you can take all of them. The sum has zero
initial investment and payoff
M (nl )
1 X n X n
ỹ nl
= (|ai l − (1 − bikl )rf | + ˜ni l Sinl )
M (nl ) i=1

Taking the expectation

M (nl )
nl 1 X n X n M (nl )
E[ỹ ] = (|ai l − (1 − bikl )rf |) ≥ ζ
M (nl ) i=1 M (nl )

Because E[˜ni l Sinl ] = E[˜ni l ]Sinl = 0.

8
Jessica Wachter Class notes for Finance 911

Taking the variance,

 
M (nl )
1 X n n
Var(ỹ nl ) = Var  ˜i l Si l 
M (nl ) i=1
 
M (nl )
1 X
= Var  ˜ni l Sinl 
M (nl )2 i=1
 
M (nl )
1 X
=  Var(˜ni l Sinl )
M (nl )2 i=1
 
M (nl )
1 X
= 2
 (Sinl )2 Var(˜ni l )
M (nl ) i=1
 
M (nl )
1 X
= 2
 Var(˜ni l )
M (nl ) i=1
 
M (nl )
1 X
≤  σ¯2 
M (nl )2 i=1

M (nl ) 2 σ̄ 2
= σ̄ = .
M (nl )2 M (nl )

Note that we have used the fact that the ˜i s are uncorrelated.
Because M (nl ) → ∞ as nl → ∞, Var(ỹ nl ) → 0 as nl → ∞. Therefore {ỹ nl } constitutes
an asymptotic arbitrage, which contradicts our assumption.