Big Data Asset Pricing

Lecture 1: A Primer on Asset Pricing

Lasse Heje Pedersen

AQR, Copenhagen Business School, CEPR

https://www.lhpedersen.com/big-data-asset-pricing

The views expressed are those of the author

Electronic copyand
available at: https://ssrn.com/abstract=4068545
not necessarily those of AQR
Overview of the Course: Big Data Asset Pricing
Lectures
I Quickly getting to the research frontier
1. A primer on asset pricing
2. A primer on empirical asset pricing
3. Working with big asset pricing data (videos)
I Twenty-first-century topics
4. The factor zoo and replication
5. Machine learning in asset pricing
6. Asset pricing with frictions

Exercises
1. Beta-dollar neutral portfolios
2. Construct value factors
3. Factor replication analysis
4. High-dimensional return prediction
5. Research proposal
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 2
This Lecture: A Primer in Asset Pricing

I Fundamentals of asset pricing

I m-pricing implies β-pricing
I Projecting m on assets: mimicking portfolio
I Hansen-Jagannathan bound
I β-pricing implies m-pricing
I β-pricing in equilibrium
I Is β-pricing the same as market efficiency or rationality?
I Multi-factor β-pricing: factor models

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 3
The Idea of These Primers and Words of Warning

I The idea of primers in asset pricing and empirical asset

pricing: Brief overview so we
I see clearly how the pieces fit together
I don’t lose sight of the end goal
I can move on to the research frontier in the next lectures

I NB: The material is highly compressed

I These notes cover so much ground that it might be a full
course
I In-depth understanding of this material is important
I So read and listen carefully – and ask lots of questions
I A few places the text is not fully self-contained, so I refer to
the references in the end (e.g., certain test statistics)

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 4
 Fundamentals of Asset Pricing

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 5
Fundamentals of Asset Pricing

Question: What determines prices and expected returns?

Answer (asset pricing without frictions):

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 6
Fundamentals of Asset Pricing

Question: What determines prices and expected returns?

Answer (asset pricing without frictions):

State prices

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 7
Fundamentals of Asset Pricing

Question: What determines prices and expected returns?

Answer (asset pricing without frictions):

State prices
I Based on a state space and its natural probabilities, there are
I 3 equivalent ways of “pricing” = computing prices and risk premia:

1. state prices, ψ
2. state price deflator, m
3. risk-neutral probabilities, Q, and the risk-free rate, r f
I Equivalence:
I to each other
I to no arbitrage = law of one price (under technical conditions)

ψ exists ⇔ m exists ⇔ Q exists ⇔ no arbitrage

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 8
Fundamentals of Asset Pricing, continued
I Mickey Mouse model
I Asset with payoff dω next time period
I Each ω is a discrete state with probability Prω
I Want to find price, p, and expected return, E (r ) = E (d)/p − 1
I Three approaches – I focus on m from next slides:
1. ψω = state prices = Arrow securities = price of $1 in state ω
X
p= ψ ω dω
ω
ψω
2. mω = = state price deflator (spd) = state price density
Prω
= pricing kernel = stochastic discount factor (sdf)
X
p= Prω mω dω = E (md)
ω
ψω
3. qω = P = ψω (1 + r f ) = risk-neutral probabilities
j ψj
= equivalent martingale measure
X  
dω Q d
p= qω =E
1 + rf 1 + rf
ω https://ssrn.com/abstract=4068545
Electronic copy available at:
c Lasse Heje Pedersen 9
Fundamentals of Asset Pricing, continued

I Broader framework used from now on:

I Discrete-time economy, t = 1, 2, ...
I Riskless rate, rtf paid at time t (some textbooks call this rt−1
f
)
I Risky asset i ∈ {1, ..., N} with, at time t,
I dividend, dti
I price, pti
I excess return, rti
I Connection between prices and excess returns:
i
dt+1 i
+ pt+1
i f
rt+1 = − 1 − rt+1
pti

I Terminology:
I i
risk premium = expected excess return = Et (rt+1 )
I f i f i
discount rate: same or rt+1 + Et (rt+1 ) or 1 + rt+1 + Et (rt+1 )

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 10
Fundamentals of Asset Pricing: m-Pricing
I Price of any risky asset i ∈ {1, ..., N}:
  " ∞ #
Mt+1 i X Ms
pti = Et i
(dt+1 + pt+1 ) = Et d i
Mt Mt s
s=t+1
" ∞ #
  X
i i
= Et mt+1 (dt+1 + pt+1 ) = Et mt,s dsi
s=t+1

I Ms Mt+1
Notation: mt,s = Mt and mt+1 = mt,t+1 = Mt
I People denote both M and m as state price deflators
I In 1-period models, usually no need to distinguish
I f
Expected returns: 1 = Et [mt+1 (1 + rt+1 i )] i.e.,
+ rt+1
I f 1 i
risk-free: 1 + rt+1 = Et [mt+1 ] and risky 0 = Et [mt+1 rt+1 ]

I i ] = (1 + r f )Cov [−m
Covariances matter: Et [rt+1 i
t+1 t t+1 , rt+1 ]
I Intuition?
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 11
Deeper Answer
Question: What determines prices and expected returns?
Deeper answer from asset pricing without frictions:
m + how it is determined
I 3 frameworks for determining m (see book by Duffie (2010)):

Condition m Focus of
A. No arbitrage Must exist, can be inferred from prices Derivatives
B. Agent optimality Agent’s marginal utility Institutional
C. Equilibrium Representative agent’s marginal utility Macro-finance

I So state price deflators depend on

A. related assets (e.g., underlying stock for option) or factors
B. agent’s marginal utility of consumption
I Mt = β t u 0 (ct ) and mt+1 = βu 0 (ct+1 )/u 0 (ct )
C. aggregate consumption and all utilities aggregated with Pareto
weights, λ
I Mt = β t uλ0 (ct ) and mt+1 = βuλ0 (ct+1 )/uλ0 (ct )

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 12
Asset Pricing with Frictions

Question: What determines prices and expected returns?

Answer from asset pricing with frictions:
I m may not exist
I “No-arbitrage condition” can break for “paper profits”
I We can still look at implications of
A. no arbitrage net of frictions
B. agent optimality
C. equilibrium
I We come back to this in the last lecture

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 13
 State Prices and Betas

See Cochrane (2009) ch. 6, although my derivations are different

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 14
m-Pricing Implies β-Pricing
I Recall: m-pricing means that, for all i,
f i
1 = Et [mt+1 (1 + rt+1 + rt+1 )]

or equivalently that

i 1
0 = Et [mt+1 rt+1 ] and Et [mt+1 ] = f
1 + rt+1

I As we already saw, m-pricing implies

i f i
Et [rt+1 ] = (1 + rt+1 )Covt [−mt+1 , rt+1 ]

which is the same as

i
Et [rt+1 ] = βti λt
Cov(−m ,r i )
where βti = Vart (m
t+1 t+1
t+1 )
f
and λt = (1 + rt+1 )Vart (mt+1 ) > 0
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 15
Portfolios
To make tradable factors (considered next), investments, and many other
things in asset pricing, we need portfolios
I Portfolio weights can be measured as
I fractions of wealth invested in each asset, x i
t
I money (say, dollars) invested in each asset, x $,i
t
I shares invested in each asset, x
ˉti
I Connection: xti wt = xt$,i = xˉti pti
I Each can be useful in different circumstances
I I focus on fractions, xti , from now on
I Here, wt is the investor’s wealth, which evolves as
X X
wt+1 = xti wt (1 + rt+1
f i
+ rt+1 ) + (wt − xti wt )(1 + rt+1
f
)
i i
X
= wt xti rt+1
i f
+ wt (1 + rt+1 f
) = wt (1 + rt+1 + xt0 rt+1 )
i

I Portfolio’s excess return, wt+1 f

− 1 − rt+1 , is super simple:
wt

t t+1 x 0r
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 16
Portfolios and Excess Returns
I Portfolio’s excess return, wt+1 f
− 1 − rt+1 , is super simple:
wt

xt0 rt+1

I What is the interpretation of

P i
I
Pi xti = 1?
I
Pi xti < 1?
I
i xt > 1?

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 17
Portfolios and Excess Returns
I Portfolio’s excess return, wt+1 f
− 1 − rt+1 , is super simple:
wt

xt0 rt+1

I What is the interpretation of

P i
I
Pi xti = 1?
I
Pi xti < 1?
I
i xt > 1?

I Any linear combination of excess returns is an excess return

I An excess return can be seen as a self-financing strategy (long $1 in
the risky asset, financing by borrow $1)→so are linear combinations
I Example:
I If we run regression rti = α + β i rtMkt + εt
I then rti − β i rtMkt = α + εt is an excess return of a hedged position
I Side note: compounding excess returns:
Q Q
I
t (1Q+ rtf
+ rti ) − t (1 + rtf )
I Not i
t (1 +P
rt )
I Fine to use i
t rt for at:
Electronic copy available illustration (but it is not a cumulative return)
https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 18
m-Pricing Implies β-Pricing with Tradable Factor
I If there is m-pricing, there is no arbitrage
I When there is no arbitrage
I we can construct a mean-variance frontier, where
I the tangency portfolio, r ∗ , is risky, Vart (rt+1
∗
)>0
I ∗
(if there is arbitrage, Vart (rt+1 ) = 0, so we cannot compute
∗ i
Cov(rt+1 ,rt+1 )
βti = ∗ ) )
Vart (rt+1

Result
For any a > 0, the portfolio, xt = aVart (rt+1 )−1 Et (rt+1 ), is
mean-variance efficient. The “tangency portfolio” corresponds to
a choice of a such that portfolios weights add up to one, 10 xt = 1.
∗
For any a, the portfolio return, r t+1 =x0t rt+1 , yields beta-pricing:
i
Et [rt+1 ] = βti λt
∗ ,r i )
Cov(rt+1
for all i, where βti = t+1 ∗ )
and λt = Et (rt+1
∗ )
Vart (rt+1
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 19
m-Pricing Implies β-Pricing with Tradable Factor, cont.
I Proof:
I Look for tangency portfolio x ∈ RN with excess return
x
rt+1 = x 0 rt+1 that maximizes the following, for any γ > 0
γ
max Et (x 0 rt+1 ) − Vart (x 0 rt+1 )
x 2
I First order condition: 0 = E t (rt+1 ) − γVart (rt+1 )x
I Solution: x = γ1 Vart (rt+1 )−1 Et (rt+1 )
I Note that Et (rt+1 ) = γVart (rt+1 )x so
i
Et (rt+1 ) = βti γVart (rt+1
x
)

ei0 Vart (rt+1 )x Covt (ei0 rt+1 ,rt+1

0
x) i
Covt (rt+1 x
,rt+1 )
where βti = x )
Vart (rt+1 = Vart (rt+1x ) = x )
Vart (rt+1
I x x
Using this relation for x yields Et (rt+1 ) = γVart (rt+1 ) so we
have the desired result
i
Et (rt+1 ) = βti Et (rt+1
x
)
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 20
m-Pricing Implies β-Pricing with Tradable Factor, cont.

I The tangency portfolio in the mean-variance diagram

I How do you draw this?
I What happens when you vary the risk aversion γ from the
proof?
I Which γ corresponds to a notional exposure of 1?
I Why does the beta relation look the same for different γ’s?
I Note: in an equilibrium model, γ clearly affects equilibrium
prices and expected returns (but, again, the beta relation still
looks the same)

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 21
Projecting m on Assets: Mimicking Portfolio
Alternative characterization of the tradable factor with β-pricing
I The projection of −m on the tradable securities

I I.e., the regression of −m on r :

−mt+1 = at + bt0 rt+1 + εt+1

where Et (εt+1 ) = 0 and Et (εt+1 rt+1 ) = 0
I General formula for regression coefficients at and bt :
bt =Vart (rt+1 )−1 Covt (rt+1 , −mt+1 )
at =Et (−mt+1 ) − b 0 Et (rt+1 )
I Using properties of mt+1 :
1
bt = f
Vart (rt+1 )−1 Et [rt+1 ]
1 + rt+1
!
1 0
at = − f
+ bt Et (rt+1 )
1 + rt+1
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 22
Projecting m on Assets: Mimicking Portfolio, cont.
We have shown:

Result
b
The regression of −mt+1 on the assets, rt+1 = bt0 rt+1 , is proportional to the
tangency portfolio since b t = 1
f
1+rt+1
Vart (rt+1 )−1 Et [rt+1 ].

Result
b
The regression of m on the assets, r t+1 = bt0 rt+1 , has β-pricing (because all
mean-variance efficient portfolios have β-pricing)

Result
Any state price deflator can be written
1  
mt+1 = f
1 − Et [rt+1 ]0 Vart (rt+1 )−1 (rt+1 − Et [rt+1 ]) + εt+1
1 + rt+1

and the minimum-variance sdf has εt+1 = 0.

Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 23
Projecting m on Assets: Mimicking Portfolio, cont.

I The minimum variance sdf is the projection of the “true”

state price on the marketable space, meaning that this m is a
function of returns:
1

mt+1 = f
1 − Et [rt+1 ]0 Vart (rt+1 )−1 (rt+1 − Et [rt+1 ])
1 + rt+1

I If the market is complete, this is the unique state price

I Otherwise, there are many m’s that price the assets (i.e., many
choices of εt+1 uncorrelated with rt+1 )

I “Mimicking portfolio”: 1
f
1+rt+1
Et [rt+1 ]0 Vart (rt+1 )−1 rt+1
I The part of −mt+1 that yields beta pricing

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 24
Hansen-Jagannathan Bound
i )
Et (rt+1
I i )=
Sharpe ratio and any asset i: SRt (rt+1 i
σt (rt+1 )

Result (Hansen and Jagannathan (1991))

The Sharpe ratios of the tangency portfolio, r ∗ , and any other portfolio, r j ,
are bounded by the volatility of the sdf:

j ∗ σt (m) σt (mt+1 ) f
SRt (rt+1 ) ≤ SRt (rt+1 )= min ≤ = (1+rt+1 )σt (mt+1 )
m: m prices rt+1 Et (m) Et (mt+1 )
I Proof
I The variance of m using the projection from previous slides:

1 2 0 −1
Vart (mt+1 ) = ( ) Et [rt+1 ] Vart (rt+1 ) Et [rt+1 ] + Vart (εt+1 )
f
1 + rt+1

I Recall: Et (mt+1 ) = 1 so
1+r f
t+1

σt (mt+1 ) q
= f )2 Var (ε
Et [rt+1 ]0 Vart (rt+1 )−1 Et [rt+1 ] + (1 + rt+1 t t+1 )
Et (mt+1 )

I SR of tangency:

∗ ) q
Et (rt+1 b 0 E (r ) E [r ]0 Vart (rt+1 )−1 Et [rt+1 ]
σt (r ∗ )
= q t t t+1 = q t t+1 = Et [rt+1 ]0 Vart (rt+1 )−1 Et [rt+1 ]
0 0 −1
Electronic copy available at: https://ssrn.com/abstract=4068545
t+1 b t Var (r
t t+1 t)b Et [rt+1 ] Vart (rt+1 ) Et [rt+1 ]
c Lasse Heje Pedersen 25
β-Pricing Implies m-Pricing

Result
If there exists portfolio with excess return r ∗ s.t. for all i
i
Et [rt+1 ] = βti λt
∗ i
Cov(rt+1 ,rt+1 ) ∗
where βti = ∗ )
Vart (rt+1
and λt = Et (rt+1 ), then
 ∗

1 ∗ ∗ Et (rt+1 )
mt+1 = f
1 − (rt+1 − Et (rt+1 )) ∗
1 + rt+1 Vart (rt+1 )
i f
satisfies 0 = Et [mt+1 rt+1 ] and Et (mt+1 ) = 1/(1 + rt+1 )

I Proof:
I (Inspection) Just check
I (Constructive) Assume m = a + br ∗ and solve for a and b

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 26
 Equilibrium β-Pricing and Market Efficiency

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 27
β-Pricing in Equilibrium: CAPM
I If all investors want to maximize their Sharpe ratio
I Each chooses a combination of the tangency portfolio and r f
I Therefore the market portfolio is the tangency portfolio,
r MKT = r ∗
I So, we see

Result
If all investors maximize SR, the market portfolio, r MKT , has
beta pricing
i
Et [rt+1 ] = βti λt
i ,r MKT )
Cov(rt+1
where βti = t+1
MKT )
Vart (rt+1
MKT )
and λt = Et (rt+1
Further, the following sdf prices all assets
!
MKT )
Et (rt+1
1 MKT MKT
mt+1 = 1 − (rt+1 − Et (rt+1 ))
f
1 + rt+1 Vart (rt+1
MKT )
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 28
Is β-Pricing the Same as Market Efficiency/ Rationality? I
I Rational investors imply the existence of m- and β-pricing
I Is the reverse also true? I.e., does β-pricing imply rationality?

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 29
Is β-Pricing the Same as Market Efficiency/ Rationality? I
I Rational investors imply the existence of m- and β-pricing
I Is the reverse also true? I.e., does β-pricing imply rationality?
I No
I As long as there is no arbitrage, there is β-pricing
I Recall that the tangency portfolio always has β-pricing
I For example:
I Suppose behavioral investors do crazy things that mess up
market prices
I Rational investors (arbitrageurs) eliminate pure arbitrage and
partly correct prices, but their trades are limited by risk
aversion and other constraints
I Equilibrium prices remain partly inefficient: irrational trades
create noise in prices and information is only partially
incorporated into prices
I Nevertheless, the tangency portfolio has β-pricing
I So testing rationality requires β-pricing for an economically
meaningful factor, a structural model, or other types of tests
I Same goes for multi-factor models – see discussion of
Fama-French
Electronic factors at: https://ssrn.com/abstract=4068545
copy available
c Lasse Heje Pedersen 30
Is β-Pricing the Same as Market Efficiency/ Rationality? II

I Rational pricing also implies that SRs are bounded

(Hansen-Jagannathan bound)
I In the reverse also true? I.e., bounded SR imply rationality?
I No
I Bounded SR just means that arbitrage has removed the best
investments, i.e., the tangency portfolio is not so attractive
that arbitrageurs trade even more
I Can be true whether or not prices are affected by behavioral
effects
I If we think that near-arbitrage (=high SR) cannot exist, then
factors that are important for expected returns must also be
important for the variance-covariance matrix
I Tests of characteristics-vs-covariances are not tests of
rationality, but tests of whether near-arbitrage exists (before
transaction costs and funding constraints)
I See Kozak et al. (2018) for further results
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 31
 Multi-Factor β-Pricing: Factor Models

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 32
What is a Factor Model?

I Excess returns of N assets, rti , and K factors, ftk ,

X
rti = αi + β i ,k ftk + εit = αi + β i ft + εit
k

where E(εit ) = 0 and E(ftk εit ) = 0. In vector form:

rt = |{z}
α + β f t + εt
|{z} |{z} |{z} |{z}
N×1 N×1 N×K K ×1 N×1

I NB: This is just a statistical model – the economics comes

from testing properties of the model (the parameters, α, β, the
factors and their explanatory power for returns, the risks, etc.)

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 33
Why Multi-Factor Models?
We can always use a one-factor model for β-pricing
I As shown above, the tangency portfolio, r ∗ , gives β-pricing
I So why do people often use multi-factor models?

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 34
Why Multi-Factor Models?
We can always use a one-factor model for β-pricing
I As shown above, the tangency portfolio, r ∗ , gives β-pricing
I So why do people often use multi-factor models?
1. “Building” mimicking portfolio from observable characteristics
I Start with factors constructed based on characteristics,
ft1 , ..., ftK , such as market, size, and value (Fama-French)
I Combining these can yield a higher Sharpe ratio
P
I Say that the highest SR arises with weights k bk ftk
I If this combination has the highest SR
Pamong all portfolios,
then we have beta pricing with rt∗ = k bk ftk , i.e.,
P
i
∗
Cov(rt+1 i
, rt+1 ) k
Cov( k bk ft+1 i
, rt+1 ) X i ,f k k
Et [rt+1 ]= ∗
λ t = ∗
λt = β t λt
Vart (rt+1 ) Vart (rt+1 )
k

Vart (f k )b λt
where λkt = Vartt+1 ∗ )
(rt+1
k

P k may proxy for marginal utilities etc.

I If so,
k bk ftavailable
Electronic copy at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 35
Why Multi-Factor Models, continued

2. APT (Ross (1976))

I Notice that the variance-covariance matrix, Vart (rt+1 ), is
generated by a few factors plus (relatively) idiosyncratic risk
I Idiosyncratic risk: largely diversified away w/many assets
I So to rule out near-arbitrage, expected returns must line up
with the factors that generate the variance
3. ICAPM (Merton (1973))
I Investors are forward looking and think long term
I They like stocks that pay off when the market in general is low
(like CAPM), but also when future investment opportunities
are bad (depending on parameters), for example when
I ∗
the highest Sharpe ratio is low, SRt+1 (rt+2 )
I f
the real interest rate is low, rt+2

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 36
 Factor Models: Statistical and Economic
Implications

See Cochrane (2009) ch. 12 for more details and test statistics

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 37
Statistical Implications of a Factor Model
Factor model: rt = α + βft + εt
I Expected returns

E(rt ) = α + βE(ft )

I Variance-covariance matrix
Var(rt ) = βVar(ft )β 0 + Var(εit )

I Regression coefficients

β = Σrf Σ−1
ff and α = E(rt ) − βE(ft )

where Σff = Var(ft ) and Σrf = Cov(rt , ft )

I For asset i β i = Σr i f Σ−1
ff and αi = E(rti ) − β i E(ft )
I If r , f jointly normal, the regression is the conditional mean
E(rt |ft ) = E(rt ) + Σrf Σ−1
ff (ft − E(ft )) = α + βft
but, more
Electronic generally,
copy it at:
available is just the linear projection
https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 38
Statistical Implications: Deriving Regression Coefficients
I Factor model

rt = α + βft + εt

I Subtract the mean on both sides

rt − E(rt ) = β(ft − E(ft )) + εt

I Multiply both sides by (ft − E(ft ))0

(rt − E(rt ))(ft − E(ft ))0 = β(ft − E(ft ))(ft − E(ft ))0 + εt (ft − E(ft ))0

I Taking the expected value

Σrf = βΣff

I Conclusion
β = Σrf Σ−1
ff and α = E(rt ) − βE(ft )
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 39
Economic Implications of Factor Models
I Suppose that we have a single factor f
I Want to examine whether f “prices” all the assets, r 1 , ..., r N
(unconditonal model)
Cov(rti ,ft )
I That is, for all i , E[rti ] = β i λ where β i = Var(ft )
I This restriction can also be written as

rti = β i λ + β i (ft − E[ft ]) + εt

I We run a time series regression of rti on ft :

rti = αi + β i ft + εt
Cov(rti ,ft )
I Slope is right, β i = Var(ft )
, and intercept is αi = β i (λ − E[ft ])

When the factor is

1. tradable, λ = E(ft ), so αi = 0 for all i
2. non-tradable, where ft = mt and iid: λ = (1 + r f )Var(mt )
(see slide “m-Pricing Implies β-Pricing”), so αi need not be 0
3. non-tradable, where ft = a + bmt : even less reason for αi = 0
Here, 1 yields a time series test of model, while 2 and 3 mean that model
must be tested
Electronic using
copy a cross-sectional
available regression (see next class)
at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 40
Economic Implications of Tradable Factors
I We have observed tradable factors f
I We can estimate factor model using a time series regression
I Want to examine whether f “prices” all the assets, r 1 , ..., r N
P
I I.e., there exists b s.t. rt∗ = k bk ftk = b 0 ft has β-pricing
I If f prices the assets and we run time series regression
rt = α + βft + εt , then we must have

α=0
I This is obvious when there is 1 factor, see previous slide
I With 1 tradable factor, this is just the definition of β-pricing
I With multiple factors: see next slide
I Intuitively, this is because expected excess returns are purely
driven by factor exposures
I Said differently, if you hedge out factor exposures, then the
expected return is zero
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 41
Multiple Tradable Factors—Why Alpha is Still Zero
P
Suppose there exists b s.t. rt∗ = k bk ftk = b 0 ft has β-pricing
I Then

Cov(b 0 ft , rt+1
i
)
E(rti ) = 0
Et (b 0 ft )
Vart (b ft )
Et (b 0 ft )
= Σi b
Vart (b 0 ft ) r f
Et (b 0 ft )
= Σ i Σ−1 Σff b
Vart (b 0 ft ) r f ff
=β i bˉ

Et (b 0 ft )
where bˉ = Vart (b 0 ft ) Σff b and using β i = Σr i f Σ−1
ff
I Since f k is itself tradable with a beta equal to
e k = (0, ..., 0, 1, 0, ..., 0), we conclude that bˉk = E (ftk )
I So: αi + β i E(ft ) = E(rti ) = β i bˉ = β i E (ft )
I In other words, we see that, for all i,
αi = 0
Electronic copy available at: https://ssrn.com/abstract=4068545
c Lasse Heje Pedersen 42
References Cited in Slides I

Cochrane, J. H. (2009). Asset pricing: Revised edition. Princeton university press.

Duffie, D. (2010). Dynamic asset pricing theory. Princeton University Press.
Hansen, L. P. and R. Jagannathan (1991). Implications of security market data for models of dynamic economies.
Journal of political economy 99 (2), 225–262.
Kozak, S., S. Nagel, and S. Santosh (2018). Interpreting factor models. The Journal of Finance 73 (3), 1183–1223.
Merton, R. C. (1973). An intertemporal capital asset pricing model. Econometrica: Journal of the Econometric
Society , 867–887.
Ross, S. (1976). The arbitrage pricing theory. Journal of Economic Theory 13 (3), 341–360.

Electronic copy available at: https://ssrn.com/abstract=4068545

c Lasse Heje Pedersen 43