Portfolio Choice

Factor models and Arbitrage Pricing Theory

Dr Alena Audzeyeva
[email protected]
 Factor Models: motivation
2 issues with Markowitz’ portfolio theory & CAPM:
1) In “practice” they require the knowledge of large number of
parameters:
For 50 stocks:
• 50 estimates of mean returns
• 50 estimates of variances
• 50(49)/2=1225 estimates of the covariances
For 100 stocks:
• 100 estimates of mean returns
• 100 estimates of variances
• 100(99)/2=4950 estimates of the covariances
For 3000 stocks (NYSE): over 4 million estimates!
2) The expected rate of return for an asset might depend on more
than one “factor” (determinant). The CAPM predicts dependence
on the risk premium on the market portfolio only (1 factor)… . 1
factor may not be enough.
Solution: simpler models (with simpler covariance structures)
 Single factor model
The rate of return on asset i can be decomposed into the expected component
(E[ri]) & unexpected (“surprise”) component ξi:
ri = E[ri] + ξi ,
If the asset returns can be approximated by the normal distributions (consistent
with the mean-variance framework), which are correlated → they are joint
normally distributed → they are driven by one or more common
factors/variables.
Assume, only one common variable f ( one factor) drives asset returns:
ξi = if + ei ,
the “surprise” component ξi is due to the macro (market-wide) innovations &
firm-specific innovations →
ri = E[ri] + if + ei
f : a factor which captures unanticipated macro (market-wide) surprises; f is the
same for all assets (it has an effect on all assets’ returns). It could be:
“the unexpected change in GDP growth”
“the change in expected inflation”
i: sensitivity of asset i to factor f ; called factor sensitivity (or factor loading, or
factor beta). Factor sensitivity is asset-specific; it is sometimes denoted as bi.
ei: measures the firm-specific surprise (innovation); E[ei] = 0 for all i .
 Single Factor Model: an example

Example 1 Suppose the asset A returns are driven by one

common factor: the unexpected change in the GDP growth
rate. The expected GDP growth rate was 3% for this year.
Given this information, the expected rate of return for asset A
was 10%. The factor sensitivity of A is βA=0.8. If GDP
increased by 4% this year (1% more than was anticipated),
how would it affect the return on stock A?
Assume, there were no firm-specific surprises.
 Single Factor Model: assumptions

Simplifying assumptions behind the factor models:

Cov(f, ei) = 0 for all i: The common factor f and asset-specific
innovations are not correlated;
Cov(ei,ej) = 0 for all i,j : The (non-systematic) asset-specific
innovations are not correlated across assets .
These assumptions simplify considerably the covariance
structure of the factor models:
Variance of asset i (measures total risk):  i2   i2 2f   i2 ei 
= systematic risk + firm-specific risk;
Systematic risk of an asset =  i2 2f - determined by the asset’ ß ;

The covariance between any two assets is determined by their ßs:

Covri , r j   Cov i f  ei ,  j f  e j    i  j  2f
 How does it help compared to CAPM?
The number of input variables for the market of n assets:
3*n + 2 estimates!
For 50 assets:
Number of estimates = 152 compared to 1225 (Markowitz portfolio
theory/ CAPM)
For 3,000 assets (NYSE):
9,002 estimates compared to over 4 mln using Markowitz portfolio
theory/CAPM)
However:
- factor models are descriptive models (there is no theory in the
equation)
- factor models impose a (strong) assumption of zero-covariances
between a factor & a firm-specific “surprise” terms & between all
firm-specific “surprise” terms both of which in fact might be non-
trivial - leading to errors.
Markowitz approach takes into account the complete
covariance structure (all covariances as they are) and gives a
precise answer.
 Multifactor models
The macro factor in one-factor model (captured by the market
return in CAPM) has more than one source of uncertainty:
- uncertainty about the business cycle, inflation, interest rate
risk, exchange rate risk…
A single factor model assumes that each stock has the same
relative sensitivity to each risk factor (so that they can be
aggregated in one beta) which is incorrect!
Stocks’ relative sensitivity to various risk factors differ! Hence,
a two-factor model is introduced:
ri  E ri    i1 f1   i 2 f 2  ei
A generic multifactor model with n factors:

ri  E ri    i1 f 1   i 2 f 2    in f n  ei
 Two factor model: an example
Example 2 (adapted from Body & Kane) Consider 2 companies: a
domestic utility company (Northern Electric) and an airline company
(Ryanair). Suppose that the returns of the two companies are driven
by two common factors:
– fGDP: a business cycle factor, measured by the unexpected change in
the GDP growth rate.
– fIR: an interest rate factor, measured by the unexpected change in the
local interest rate
The following 2 factor models are estimated for the two companies:
rNE  0.12  0.4 f GDP  1.5 f IR  eNE ;
(*)
rRA  0.13  1.2 f GDP  0.3 f IR  eRA

Interpret these factor equations

 A multifactor SML
Where does E[ri] come from? We need a theoretical model of
equilibrium asset returns (e.g., CAPM) to determine E[ri].
E ri   r f   i E rm   r f 
In CAPM:

This can be re-written as:

E ri   r f   i  RPm
How can this relationship be generalised when we recognise that assets
are exposed to multiple sources of systematic (market-wide) risk.
Multifactor model (a preview of APT):
E ri   r f   i ,GDP  RPGDP   i , IR  RPIR (MFM)
here
RPGDP  GDP is the risk premium associated with 1 unit of the
GDP-risk exposure;
RPIR  IR is the risk premium associated with 1 unit of the
IR-risk exposure;
GDP ,  IRsometimes called as “price of risk” or “factor prices”; they are
the same for all assets (can be positive or negative)
 Expected rate of return in the two-factor
model: an example
Example 2 (slide 8) cont. Suppose that the risk-free interest
rate is 4%; the risk premium for 1 unit of exposure to GDP risk
(GDP-factor price) is 6.61% and the risk premium for 1 unit of
exposure to interest rate risk (IR-factor price) is -3.57.
Determine the expected rates of return for Northern Electric
and Ryanair.
 Multifactor models & APT

• If the observed returns are generated by a multifactor

model (multiple sources of systematic risk) …
then the single factor models are not longer correct &
may lead to a mistakes

• The main question for us here is:

If the “observed” asset returns are generated by a
multifactor model, what do we expect about the pricing of
the asset?

Enters Arbitrage Pricing Theory (APT)

 Arbitrage Pricing Theory (APT)
• Developed by Stephen Ross in 1976.
Three main assumptions of APT:
• asset returns can be described by a factor model;
• there are sufficient assets to diversify away idiosyncratic (firm-specific) risk
• well-functioning asset markets do not allow for the persistence of arbitrage
opportunities
Arbitrage opportunity: when an investor can earn riskless profits with zero-net
investment. Two types of arbitrage:
– Type A arbitrage exists if there is an investment opportunity which produces an
immediate positive reward with no future payoff (strictly balanced portfolio)
– Type B arbitrage exists if there is an investment opportunity with zero cost but
has a positive probability of yielding a positive payoff and no probability of
yielding a negative payoff.
– If neither Type A or Type B arbitrage possibilities exist we say that there is no
arbitrage possibility (absence of arbitrage opportunities)

• A consequence of the absence of opportunities of arbitrage => law of one

price: similar assets should be priced similarly !
 Implications of APT versus CAPM
• If a riskless arbitrage opportunity arises, market participants will engage in
arbitrage activities: construct a risk-free arbitrage portfolio with mispriced
assets/portfolios (“buy low, sell high”) and trade in it until the arbitrage
opportunity is eliminated.
• In the process arbitrageurs will push the price of the underpriced asset/portfolio
up and/or the price of the overpriced asset/portfolio down until the equilibrium is
obtained.
• Because the arbitrage portfolio is risk-free, all investors will want to take
unlimited (as large as possible) positions in it – regardless of their risk
preferences.
Here comes the difference between the arbitrage (APT) & risk-return
dominance (CAPM) in regards to what happens if the market is not in
equilibrium (& hence, arbitrage opportunities exist):
• APT: each investor will want to take as large as possible position in the
arbitrage portfolio → it will take actions of only a few investors to obtain
equilibrium;
• CAPM (risk-return dominance argument): many investors will shift (re-balance)
their portfolios, depending on their risk preferences. A large number of investors
is required (who sell and buy assets in the process) to shift the prices
sufficiently & to restore the equilibrium.
• → Implications of APT are stronger compared to CAPM!
 APT: importance of diversification
Stocks cannot be “perfectly” replicated by other assets because of their
idiosyncratic characteristics (risk) → hence, the arbitrage argument holds
for well diversified portfolios! where the stocks’ idiosyncratic risks are
eliminated (rather than for individual stocks).
How? Lets look at the portfolio P of n stocks with weights wi.
Random rate of return (given one-factor economy, for simplicity):
r  Er    f  e
P P P P
(+) where

 P   wi  i ; ErP    wi Eri ; e P   wi ei
i i i

Portfolio variance can be decomposed into systematic & non-systematic

components:
 
 P2   P2 2f   2 eP , where  2 eP   Var  wi ei    wi2 2 ei 
 i  i
Assume (for simplicity) that the portfolio is equally weighted: wi=1/n → the
non-systematic portfolio variance becomes:
1  2 ei  1 2
2
1 2
 e P    w  ei       ei   
2 2
i
2
  ei 
i i n n i n n
When n is large,  2 e   0. . E e   0, Hence any realised value of eP~0.
P P
 APT: importance of diversification
• Substituting the last result into (+) gives

rP  ErP    P f

• Non-systematic (asset-specific) component disappears

for well-diversified portfolios!
• Only a factor (systematic) risk commands a risk premium
as the non-factor (stock-specific) risk can be diversified
away → Investors are not rewarded for bearing stock-
specific risk!
• This result holds for well-diversified portfolios in a
multifactor economy. If n=2:

rP  ErP   1P f1   2 P f 2
 Single factor economy
Rate of Return, %

- random rate of return on a single stock S

- random rate of return on a well diversified portfolio P

rS  8.0  1.0 f  es
rP  8.0  1.0 f

8%

(0,0) Factor, f
 Arbitrage opportunity: could we have well-diversified
portfolios P & N with the pattern of returns shown?
Rate of Return, %

rN  10.0  1.0 f
rP  8.0  1.0 f

Arbitrage Opportunity!

(0,0) Factor, f
 Arbitrage portfolio
when portfolio betas are the same
What is the arbitrage opportunity here?
• P & N have the same exposure to systematic risk captured by factor f:
βP= βN =1.0. → we would expect the expected returns on these two
portfolios to be the same. However, E[rN] = 10% > E[rP] = 8% →
P is overpriced relative to N!

Risk-free arbitrage strategy:

Sell short £1mln (say) of P
Buy £1mln of N

The net earnings from the arbitrage strategy:

Long position in N: (Investment amount) * rN = £1mln x (0.10 + 1.0f )
Short position in P: (Investment amount) * rP = - £1mln x (0.08 + 1.0f )
Risk-free net proceeds: = £1mln x 0.02 = £20,000
E[R] Arbitrage opportunity: portfolios with different betas
Consider a new portfolio C with βC= 0.5 & E[rC] = 5%.
N has βN= 1.0; E[rN] = 10% (as before); the risk-free rate rf=3%.
Construct D: a synthetic portfolio (0.5 F, 0.5 N) with βD= βC= 0.5

0.10
N

>0
0.05
C
rf =
0.03
F

(0,0) β=0.5 β=1 β

D has the same beta as C and a higher rate of return → Arbitrage opportunity!
The expected rate of return for all assets must lie on the line through F and N!
 Factor-price equation: 2 factor case
Take eqn. (MFM) (slide 9), move rf to the left hand side & use generic
factors f1 & f2 (also, substitute factor risk premiums for lambdas:
RPi=i):

rP*   1P 1   2 P  2 (APT)

where rP*  ErP   r f

Eqn (APT) is the fundamental pricing result of the APT. It relates the
risk premium rP* of a large, well-diversified portfolio P to the factor
loadings of this portfolio, using the coefficients λ as weights.
Factor price λ1 can be interpreted as a risk premium of a portfolio that’s
exposed only to factor 1 but not to factor 2: β1=1 and β2=0. Similarly
for λ2 (β1=0 and β2=1). Such portfolios are called factor portfolios.
The risk premium of a well-diversified portfolio P can be determined
from its factor loadings, β1P and β2P and from the factor prices λ1 & λ2
(APT)
 Finding the factor prices
It follows from eqn. (APT) that we can infer the factor prices
λ1 & λ2 from the factor loadings and risk premia of any
two observed well-diversified portfolios. Suppose,
portfolios L and M are both well diversified. Then , by
(APT):
rL*  1L 1   2 L 2
rM*  1M 1   2 M 2
If betas and risk premia of L & M are known (say, estimated
from empirical data), we can solve the system above for
the two unknown lambdas.
Once we know the lambdas, we can price any other well-
diversified portfolio & check the market for arbitrage
opportunities.
 Worked example: 2-factor economy
Suppose, asset returns are governed by a two-factor APT model; factor
f1represents domestic risk and factor f2 represents export risk. Both
factors have zero mean. Returns on three well-diversified portfolios
L, M and N satisfy the three factor-price equations:
rL  8.0  1.75 f1  0.25 f 2 , rM  10.0  0.75 f1  1.0 f 2 , rN  9.0  1.0 f1  0.5 f 2 ,

where fi is the (uncertain) value of factor i (i = 1; 2), and rj is the

(uncertain) rate of return of portfolio j (j = L; M;N); fi and rj are given
in per cent. The risk-free rate of return on government bonds G is
rf = 6%.

The Task:
(i) From the information we have for bond G and portfolios L and M,
find the two factor prices λ1 and λ2.
(ii) Identify a risk-free arbitrage portfolio involving L, M and N.
(iii) Determine the appropriate correction in the rate of return of N that
would eliminate the arbitrage opportunity from (ii).
(solution on white board)
 Reading
➢ Study Worked Example (Appendix to these lecture notes)
➢ Read one of the following:
Standard reading:
• Essentials by Bodie, Kane & Marcus, Ch.7, Sections 7.4 -7.5
• Brown & Reilly, Ch.9
More advanced reading:
• Investments by Bodie, Kane & Marcus, Chs.8 and 10
• Luenberger, Ch. 8.
➢ Additional reading (available on KLE):
• Burmeister, Roll & Ross, “A Practitioner’s Guide to Arbitrage
Pricing Theory”, CFA publications (1994)
• Fama & French, “Multifactor Explanations of Asset Pricing
Anomalies”, Journal of Finance 51 (1996), pp.55-84.
• Roll & Ross, “An Empirical Investigation of the Arbitrage
Pricing Theory”, Journal of Finance 35 (1980), pp.1073-1103.