Empirical Asset Pricing: Introduction

Ralph S.J. Koijen Stijn Van Nieuwerburgh∗

September 5, 2021

∗
Koijen: University of Chicago, Booth School of Business, NBER, and CEPR. Van Nieuwer-
burgh: Columbia Business School, CEPR, and NBER. If you find typos, or have any
comments or suggestions, then please let us know via ralph.koijen@chicagobooth.edu or
svnieuwe@gsb.columbia.edu.
1. Basic structure of the notes

• High-level summary of theoretical frameworks to interpret em-

pirical facts.

• Per asset class, we will discuss:

1. Key empirical facts in terms of prices (unconditional and

conditional risk premia) and asset ownership.
2. Interpret the facts using the theoretical frameworks.
3. Facts and theories linking financial markets and the real
economy.
4. Active areas of research and some potentially interesting
directions for future research.

• The notes cover the following asset classes:

1. Equities (weeks 1-5).

– Predictability and the term structure of risk (week 1)
– Cross-section and the factor zoo (week 2)
– Intermediary-based asset pricing (week 3)
– Production-based asset pricing (week 4)
– Asset pricing via demand systems (week 5)
2. Mutual Funds and Hedge Funds (week 6).
3. Options and volatility (week 7).
4. Government bonds (week 8).
5. Corporate bonds and CDS (week 9).
6. Currencies and international finance (week 10).
7. Commodities (week 11).
8. Real estate (week 12).

2
2. High-level summary of theoretical frameworks

• To organize and interpret the empirical facts, various theoret-

ical frameworks will be used.

1. Factor models for the cross-section of expected returns.

2. Stochastic discount factor (SDF) models based on “a”
marginal investor.
3. Structural equilibrium models: Endowment and pro-
duction economies.
4. Equilibrium models: Asset pricing using demand sys-
tems.

• By moving from the first to the third framework, we impose

more structure and hence there are more testable predictions.

• Every equilibrium asset pricing model implies a demand sys-

tem. Oftentimes, the implications are not very interesting (e.g.,
the representative-agent model) or the implications have not
been tested using data on asset holdings. The fourth frame-
work tries to match holdings data and asset pricing data jointly.

• We start with a high-level overview of these frameworks. Clas-

sic asset pricing textbooks provide a more comprehensive treat-
ment of the first and third frameworks.

3
2.1. Factor Models

• Cochrane’s Asset Pricing book contains a detailed treatment of

factor models, including their estimation.

• The absence of arbitrage opportunities implies the existence

of a positive stochastic discount factor (SDF), Mt+1 , which is
unique if markets are complete.

• Given how competitive financial markets are, we typically as-

sume that there are no arbitrage opportunities (after account-
ing properly for trading costs).

• For any valid SDF Mt+1 , we have

e
Et [Mt+1 Rt+1 ] = 0,

e
for any excess return, Rt+1 = Rt+1 − Rtf , where Rt+1 denotes the
return and Rtf the one-period risk-free rate.

• Straightforward rewriting implies

e
Et [Rt+1 ] = βt λt ,

e
where βt = Covt [Rt+1 , Mt+1 ]/V art [Mt+1 ] and λt = −V art [Mt+1 ]Rtf .

• Implication: The absence of arbitrage opportunities implies a

single-factor model for expected returns, where the single fac-
tor is the stochastic discount factor.

4
• We do not observe Mt . It is common practice to use linear factor
models to approximate the SDF,

Mt+1 = a + b0 Ft+1 ,

with interpretable factors (e.g., a business cycle factor like con-

sumption growth or investment growth, an interest rate factor)
that plausibly correspond to states of high marginal utility for
households, binding constraints of institutions, . . . .

• The above derivation shows, however, that any successful K−factor

model implies a 1−factor model.

• Ideally, we find a low-factor representation of the SDF with a

clear economic interpretation of the factors.

• Two standard ways to motivate factor models:

1. APT:
– The APT starts with a statistical description of returns.
In particular, suppose that there exists a K−factor
model in returns

Ri,t = ai + βi0 Ft + ei,t ,

where Ft ∈ RK , E[ei,t | Ft ] = 0, and V ar(et ) is a diagonal

matrix (this can be relaxed to allow for some depen-
dence that vanishes in large, diversified portfolios).
– The absence of arbitrage implies:

E[Ri,t+1 ] = λ0 + βi0 λ,

where β, λ ∈ RK . λ0 is the zero-beta rate, which equals

5
 the risk-free rate when it exists.
– The key observation is that idiosyncratic risk is not
priced and only systematic risk factors matter for ex-
pected returns.
– However, the APT does not tell you what the factors are,
nor what the signs and magnitudes of the risk prices
are.
⇒ It may well be the case that the risk price of a given
factor is zero while the factor explains a significant
fraction of realized returns.
2. ICAPM / Equilibrium models:
– Equilibrium asset pricing models imply a SDF, and
therefore specify the relevant pricing factors.
– The ICAPM suggests that we identify innovations to
factors that capture the distribution of future returns,
summarizing future investment opportunities.
– The ICAPM often implies testable restrictions that the
factors should predict future returns and return vari-
ances, or restrictions on the signs (and sometimes mag-
nitudes) of risk prices.
– See Maio and Santa-Clara (2012) for a discussion and
tests of the restrictions on multi-factor models implied
by the ICAPM.
– We will see various examples of this in different asset
classes.

6
• Why are factor models useful?

1. A reduced-form, low-dimensional description of expected

returns may guide the design of equilibrium models.
– The reduced-form factor literature may be able to iden-
tify the key factors that price the cross-section of ex-
pected returns. This may provide new moment condi-
tions for theoretical models.
– For instance, the Fama and French 3-factor model (Fama
and French, 1992) had a large impact on the theo-
retical asset pricing literature where researchers tried
to explain the value premium. Three examples are
Zhang (2005), Lettau and Wachter (2007), and Gar-
leanu, Panageas, and Yu (2012).
– Analogously, in the currency literature, Lustig, Rous-
sanov, and Verdelhan, 2011 summarize the main di-
mensions of risk premia that is guiding the design of
international finance models.
2. If a factor model holds perfectly, optimal portfolio choice
reduces to choosing an optimal portfolio over the factors.
This dramatically reduces the number of parameters that
need to be estimated, which could dramatically improve
the stability of the optimal portfolio.
Consider the standard mean-variance solution

1 −1
w= Σ µ.
γ

If we plug in sample estimates of average returns and the

covariance matrix, the solution tends to be unstable and
perform poorly out of sample.

7
 If a factor model holds perfectly, we can replace

µ = βλ,

where we can estimate β accurately and we can estimate

λ potentially over longer samples. Moreover, if we have
a statistical factor model of returns, we need to estimate
fewer parameters for the covariance matrix Σ.
3. Performance measurement of mutual funds and hedge funds.
If you find non-zero alphas relative to the factor model, you
want to deviate from holding a portfolio based on the fac-
tors only.
Both points above explain the huge interest from the asset
management industry in factor models / factor investing.
4. It may help help to compute the cost of capital for firms in
case of equities and corporate bonds.
5. However, in practice, estimating factor betas and risk pre-
mia remains a difficult task (Fama and French, 1997).

8
• Limitations of factor models

1. Factor models are silent about the underlying economics

behind the factors. For instance, the fact that we see co-
movement for certain anomalies such as momentum does
not mean that there is a rational, risk-based explanation.
A factor structure in irrational beliefs (for instance, ex-
trapolation as in Barberis, Greenwood, Shleifer, and Yin,
2015) can also generate co-movement and can give rise to
a factor model. See Kozak, Nagel, and Santosh (2020) for a
detailed discussion on the interpretation of factor models.
All that is needed is that riskless arbitrage opportunities
do not exist and there exists a single-factor model. This is
a very weak condition.
The value anomaly is a prime example where researchers
(still) disagree whether the value premium is compensa-
tion for risk (Campbell, Giglio, Polk, and Turley, 2018)
or due to expectation errors (e.g., La Porta, Lakonishok,
Shleifer, and Vishny 1997).
2. Factor models cannot be used to think about policy ques-
tions or counterfactuals. This requires a fully-specified
equilibrium model.

Of course, both points simply reflect the fact that factor models
are reduced-form ways of looking at data.

9
• Some broader concerns with the current factor model litera-
ture

1. The number of factors has exploded. We now have over

400 factors that (supposedly) drive return differences in
the cross-section of expected returns. Once we add (non-
linear) combinations of these factors, we have thousands.
This provides little guidance for the design of equilibrium
models. Dimension-reducing techniques are necessary to
guide theory. Much more on this in week 2.
2. The current generation of factor models appears rather
“static.”
– The nature of shocks changes all the time (e.g., the
credit crisis in the U.S., the earthquake in Japan, the
Greek crisis/Euro crisis, geopolitical risk in Russia or
the Middle East, COVID-19 crisis, ...).
– It would be surprising if changes in the nature of shocks
would map into stable factor structures that apply uni-
formly across asset classes and across countries.
– One view is that some shocks are structural, giving
rise to the traditional, stable factors. Also, some fac-
tors (e.g., value) may generally capture cross-sectional
dispersion in expected returns, regardless of its source.
– However, the fragility of factor models may be due to
the changing nature of shocks or demand curves.
– To understand whether “changing risks” are important
for asset pricing, we may need different types of models
or data.
– In this context, it may be interesting to explore text
data. So-called topic models are designed to capture

10
 exactly the changing nature of factors (Blei, Ng, and
Jordan, 2003).
– Text data is high-dimensional, and hence sensitive to
small variations in methodology. Most of the work is
taking off-the-shelf methods from machine learning.
See Gentzkow, Kelly, Taddy (2019 JEL) for an intro-
duction.
– One concern is that any statistical approach will al-
ways have limitations. Using economic priors to dis-
cipline the dimension-reduction problem, analogous
to VAR models in macro (Del Negro and Schorfheide,
2004) may be a productive way to go.
3. Although it would have been nice to uncover a simple struc-
ture, the evidence so far suggests that the research agenda
to uncover a simple factor model has made limited progress.
There are many different factors for each asset class and
the factors across asset classes and countries are not that
highly correlated.
4. This is not because the field did not try. There has been a
large research effort in both academia and in the industry.
5. We may need to explore more (non-return) data and may
need to use more theory to understand the beliefs and ob-
jectives of the major investors in different asset classes.
This is where the work on “the” marginal investor is rele-
vant. Then, we need to impose market clearing and figure
out equilibrium asset prices. This is where equilibrium
models come in.
Traditional structural asset pricing models with heteroge-
neous agents (who differ in their preferences, beliefs, trad-
ing/production technologies, or endowments) have made

11
some strides in simultaneously accounting for prices and
quantities.
However, with current computational techniques, these
methods run into curse of dimensionality problems once
more than three (types) of agents are considered.
We will discuss recent advances in demand-based equi-
librium asset pricing models which can deal with more
agents. In both cases, taking seriously the role of insti-
tutions, such as financial intermediaries, has gained new
prominence in the aftermath of the Great Financial Crisis.

12
2.2. SDFs Based on “A” Marginal Investor

• Instead of modeling the behavior of all investors, a natural

starting point may be to model the behavior of a class of in-
vestors.

• For instance, we can model the behavior of households that

hold stocks or institutions that trade in various asset classes
(e.g., broker-dealers).

• Assume households maximize their value function over con-

sumption and their investment portfolio,
∞
X
max β s Et [u(Ct+s )] ,
s=0

for some utility function u(·).

• The first-order condition is given by

u0 (Ct+1 ) e
 
Et β 0 R = 0.
u (Ct ) t+1

• We can check whether this condition holds for subsets of house-

holds (e.g., rich, highly-educated households). This is not nec-
essarily the representative agent.

• Importantly, in this literature, we do not impose market clear-

ing.

13
• To understand equilibrium asset prices, we want to under-
stand the investment decisions of the major investors in a given
asset class. E.g., mutual funds for equities, insurance com-
panies for corporate bonds, and banks for mortgage-backed
securities.

• We will see examples below where first-order conditions appear

to hold, but the investors in question are small players in a
given asset class. What do we learn in such case?

• Their first-order condition still provides a valid stochastic dis-

count factor, but these investors are likely to be price-taking
agents. Hence, this may be a useful way to understand the
priced sources of risk (risk factors) in a particular asset class.
This represents significant progress.

• However, in any counterfactual or policy experiment, these in-

vestors may not matter much, so that a more holistic view of
the asset market is required for normative questions.

14
2.3. Equilibrium Models: Endowment and Production Economies

• The next step up from models that focus on a subgroup of in-

vestors (first-order conditions) is to impose market clearing.

• Endowment economies specify preferences, endowments, and

beliefs, starting from the classic paper by Lucas (1978).

• Leading asset pricing models in this class:

1. The habit model (Campbell and Cochrane, 1999).

2. The long-run risks model (Bansal and Yaron, 1999).
3. The variable rare disasters model (Gabaix, 2012 and Wachter,
2013 ).

• To keep things tractable, these models share three common

features:

1. Representative agent: No heterogeneity.

2. Institutions and intermediation plays no role.
3. No frictions (e.g. borrowing constraints).

• Assumptions 1+3 imply that markets are complete. In the

Lucas model, agents differ in the realizations of their endow-
ments, but they are all ex-ante identical and can trade a full
set of state-contingent securities. This enables them to share
risk perfectly. Asset prices are the same as if there were a rep-
resentative agent.

• Of course, there are many interesting extensions of these mod-

els that relax assumptions 1-3.

• Most of the models that relax assumptions 1 and 3 are heterogeneous-

agent incomplete markets models which feature imperfect risk

15
 sharing. This is a large and active area of research in both
macro-economics and finance. Since closed-form solutions are
usually not available, numerical solution techniques and pow-
erful computers are in order.

• Common justifications for assumption 2 that institutions and

intermediaries play no role:

1. Intermediaries act in the best interest of households who

own the capital that is managed by the intermediaries;
they are a veil.
2. If intermediaries did not act in the best interest of house-
holds, then households would reallocate their capital.

• The first justification requires that your fund manager, your

pension fund, . . . knows your U 0 (C). This is a strong assump-
tion.

• The second justification has testable implications. For instance,

if a mutual fund under-performs, we should see households
reallocate capital away from that fund. If regulation changes
the asset allocation of insurance companies, then households
should reallocate capital toward/away from insurers.

• In reality, it may not be easy to move your capital around (e.g.,

capital locked up in pension funds and insurance companies).
Or households may not be aware of the regulatory changes.

• Lots of interesting work can be done to understand whether

these frictions are indeed important for asset pricing. So far,
these assumptions have been largely taken for granted.

• Using detailed data on flows and holdings of different institu-

tions, we can test whether these mechanisms are active.

16
• A recent generation of models explicitly models financial insti-
tutions, see for instance Brunnermeier and Pedersen (2009)
and He and Krisnamurthy (2013). Much more in week 3.

• Note also that in models without financial institutions, many

interesting policy questions involving such institutions cannot
be answered. An example is optimal macro-prudential policy
(Ex. Elenev, Landvoigt, and Van Nieuwerburgh (2021)).

• Endowment economy models have dominated the theoretical

asset pricing literature for the last two decades. We will discuss
the empirical successes and failures of these models.

• Production-based models peel the onion one layer back and

endogenize the endowments. They start from a process for
technology and model the production and investment decisions
of firms, sometimes alongside the consumption decisions of
households. Consumption, labor income, and dividend income
are now endogenous, which makes it much more challenging
to match the data on prices and quantities. More in week 4.

• In some cases, the real side and the asset pricing side can be
“disconnected” through particular assumptions about prefer-
ences (Tallerini, 2000). It is unclear whether we want asset
markets to be this disconnected/neutral.

• A recent strand of the literature features production economy

models with financial intermediaries, see Adrien and Boyarchenko
(2015), Brunnermeier and Sannikov (2014), He and Krisna-
murthy (2019), and Elenev, Landvoigt, and Van Nieuwerburgh
(2021).

17
• There is virtually no empirical work testing these new models.
An interesting challenge is how to match these models to the
real world with heterogeneous institutions that differ in terms
of their regulatory frameworks and the liquidity of their liabil-
ities (e.g., banks versus insurance companies).

18
2.4. Asset Pricing via Demand Systems

• The models and frameworks discussed so far focus on asset

prices and macro-economic quantities (e.g., consumption).

• While some of these models are tested using micro-data such

as household-level consumption data, predictions for asset hold-
ings are rarely confronted with actual holdings data.

• There are many questions that involve a shift in the demand

curve of a group of agents:

– Do large institutions amplify volatility in bad times? Should

they be regulated as SIFIs (OFR 2013)?
– How do large-scale asset purchases (i.e., QE) affect asset
prices through institutional holdings?
– What if bond mutual funds experience large outflows?
– What is the impact on asset prices if we change the risk
weights of banks or insurance companies?
– What is the impact of China and Japan buying large quan-
tities of U.S. Treasuries?

• To answer these questions, we need a model with

1. Rich heterogeneity that matches the holdings of different

types of institutions and the household sector.
2. Correct substitution patterns across investors.

• In short, we need a realistic model of the demand system for

financial assets.

19
• Any equilibrium model implies a demand system, but the pre-
dictions tend to be rather stark. In particular, assets are often
very close substitutes and demand curves are virtually flat.

• IF true, then many policy questions have trivial answers and

the impact of changing regulation, flows, et cetera, is minimal.

• Although this implication / assumption is shared by many

models, there is little direct evidence of its empirical relevance.

• We will discuss in week 5, however, evidence of downward-

sloping demand curves in equity and fixed income markets.

• Also, we will discuss a recent literature on demand systems in

asset pricing, launched by Koijen and Yogo (2019). This model
revisits mostly forgotten work going back to Brainard and Tobin
(1968), Friedman (1980) and Friedman (1985).

• These models start from an empirical specification of demand

curves, which, combined with market clearing, leads to equi-
librium asset prices.

• The model can be “closed” by modeling how households allo-

cate capital to various institutions.

20