Utility based asset pricing

Nicola Borri

LUISS

This version: September 28, 2018

Definition
Asset pricing theory tries to understand the prices or values of
claims to uncertain payments.
I Prices and returns
I How to value an asset?
I delay
I risk

I Corrections for risk are much more important determinants of

many asset values (for example, over the last 50 years US
stocks have given a real return of about 9% on average; of
this only about 1% is due to interest rates).
I Positive vs. normative financial economics: description of the
way the world does work or the way the world should work?
I Equilibrium asset pricing
I Price equals expected discounted payoff
I Absolute versus relative pricing
I After we learn how to compute asset values, learn how to take
derivatives ∂V
∂P : risk-management, portfolio analysis, scenario
analysis, etc.
I Macroeconomics and Finance:
I Expected returns vary across time and across assets in ways
that are linked to macroeconomic variables, or variables that
also forecast macroeconomic events.
I Sizable excess returns predicted by macro variables: do we
have a good model to understand this? Campbell and
Cochrane (1999), Bansal and Yaron (2004), Barro (2006)?
I Investment and interest rates.
I Risk aversion.
I Cost of business cycles? Lucas (1987) or Croce (2007)?
I Discount factor (or GMM) approach to asset pricing:
I One framework for different classes of assets (stocks, bonds,
currencies, options, etc):

Pt = Et (Mt +1 Xt +1 ),

where Pt denotes the asset price, Mt +1 the stochastic discount

factor and Xt +1 the asset payoff.
I Mt +1 = f (data, parameters); f (.) is the model.
I Empirical procedures: Hansen (1982)’s General Method of
Moments and Fama and MacBeth (1973).
I In short, pick free parameters of the model to make it fits best,
which usually means to minimize pricing errors and evaluate
the model by examining how big those pricing errors are.
Outline

Introduction

Utility-based asset pricing

Stochastic discount factor
Prices, payoffs and returns
Risk-free rates
Stock returns
Volatility puzzle

Risk correction
Idiosyncratic risk
Expected return-beta representation and market price of risk
Mean-variance frontier
The equity premium puzzle
Random walk and time-varying expected returns
Outline

Introduction

Utility-based asset pricing

Stochastic discount factor
Prices, payoffs and returns
Risk-free rates
Stock returns
Volatility puzzle

Risk correction
Idiosyncratic risk
Expected return-beta representation and market price of risk
Mean-variance frontier
The equity premium puzzle
Random walk and time-varying expected returns
Utility-based asset pricing

I The problem:
I An investor must decide how much to save and how much to
consume, and what portfolio of assets to hold.
I In order to answer this question, we need to introduce a
simplified model of human behaviour.
I We do this through a utility function that captures basic
features of human psychology: the evidence that people prefer
money now, and money that is not very risky.
Utility-based asset pricing

I Marginal benefit and marginal cost.

I Interest rates and expectations of consumption growth.
I Risk corrections and covariances with marginal utilities.
I Consumption useful indicator of marginal utility.
I Market portfolio and the Capital Asset Pricing Model
(CAPM).
Utility-based asset pricing

I Find the value at time t of a payoff Xt +1 at time t + 1.

I Example: if you buy a stock today at price Pt , the payoff next
period is the stock price Pt +1 plus dividend Dt +1 :

Xt +1 = Pt +1 + Dt +1 .

I The payoff Xt +1 is a random variable.

I At date t, an investor does not know exactly how much he
will get from his investment at date t + 1.
I But the investor can assess the probability of various possible
outcomes.
Stochastic discount factor

I Let Q be the original consumption level (if the investor

bought none of the asset; for example think his labor income).
I Let ξ be the amount of the asset the investor chooses to buy.
I Assume investors can freely buy or sell as much of the payoff
Xt +1 at the price Pt as they wish.
I Let β be the subjective discount factor.
Stochastic discount factor

I We can write the investor’s optimization problem as follows:

max u (Ct ) + Et [ βu (Ct +1 )]

ξ

subject to:

Ct = Qt − Pt ξ,
Ct + 1 = Qt +1 + Xt +1 ξ.
Stochastic discount factor

I We assume that the investor can buy or sell the assets, i.e., no short
sales or borrowing constraints, and therefore an interior solution to
the maximization problem.
I Discounting the future by β captures investors’ impatience
(subjective discount factor, typically ≈ 0.95 on an annual basis).
I Utility function captures desire for more consumption, rather than
intermediate objectives like mean and variance of portfolio returns.
I The period utility function is increasing and concave.

I The curvature of the utility function generates aversion to risk and

to intertemporal substitution (i.e., the investor prefers a
consumption stream that is steady over time and across states of
nature).
Stochastic discount factor

I We will often use a convenient power utility form:

1
u (C ) = C 1− γ .
1−γ

I In this special case, the marginal utility of consumption is equal to

details :
u 0 (C ) = C − γ .
I Note that for γ > 1 the utility is negative: is this a problem?
Stochastic discount factor

I Substitute the constraints into the objective.

I Set the derivative with respect to ξ to zero:

Pt u 0 (Ct ) = Et [ βu 0 (Ct +1 )Xt +1 ],

where Pt u 0 (Ct ) is the loss in utility if the investor buys

another unit of the asset, and Et [ βu 0 (Ct +1 )Xt +1 ] is the
expected and discounted increase in utility he obtains from
the extra payoff Xt +1 .
I The investor continues to buy or sell the asset until the
marginal loss equals the marginal gain.
Stochastic discount factor

I The basic pricing formula is thus:

u 0 ( Ct + 1 )
P t = Et [ β Xt +1 ]. (1)
u 0 ( Ct )
I Note that this is a joint restriction on asset returns and
consumption.
Stochastic discount factor

I Let Mt +1 be the stochastic discount factor (SDF) defined as:

u 0 ( Ct + 1 )
Mt + 1 ≡ β .
u 0 (Ct )
I Then, the basic pricing formula can be expressed as:

Pt = Et [Mt +1 Xt +1 ]. (2)

I Formula (2) is the cornerstone of modern asset pricing theory

and can be used to price any assets.
Stochastic discount factor

I Note that the equilibrium condition P = E (MX ) holds after

everybody has done his/her investment choices.
I An individual will observe payoffs and prices and line-up
consumption accordingly.
I For example, if the price of an asset is too low, the individual
will buy a lot of it, reduce current consumption and thereby
increase future consumption.
I Also note that the equilibrium condition describes a marginal
investment: it is not appropriate to describe a large asset
purchase from a venture capitalist, a hedge-fund, etc.
Stochastic discount factor

I If there is no uncertainty, we can express prices via the

standard present value formula:
1
Pt = Xt +1 ,
Rtf
where Rtf is the gross risk-free rate: at t lend $1 and get a
sure payoff of Rtf at t + 1
I In this case 1/Rtf is the discount factor.
I The sub-index t denotes the fact that the payoff is known at
time t.
Prices, payoffs and returns

I The formula
Pt = Et [Mt +1 Xt +1 ]
applies to many cases:
I For stocks, the payoff Xt +1 is the price next period Pt +1 and
the dividend Dt +1 .
I We divide the payoff by the price to obtain the gross return
(Rt +1 = XPt +t 1 ).
I We can think of a return as a payoff with price one.
Prices, payoffs and returns

I The formula
Pt = Et [Mt +1 Xt +1 ]
applies to many cases:
I For a one-period bond, the payoff is 1.
I You buy it at price Pt and you get 1 dollar next period.
I Alternatively, we can think of a return on a bond in the
following way: you pay 1 dollar today, and you receive Rt +1
dollars tomorrow. The Euler equation is thus:

Et [Mt +1 Rt +1 ] = 1.
Prices, payoffs and returns

I A risk-free interest rate corresponds to the following case: you

pay 1 dollar today, and you receive Rtf dollars tomorrow,
where Rtf is known at date t (which is why it is risk-free).
Prices, payoffs and returns

I Asset prices are generally non-stationary.

I Think in terms of returns or price-dividend ratios: they are
typically stationary.
I The return on a stock is Xt +1 /Pt .
I The price-dividend ratio is Pt /Dt and it is the price of the
following payoff:
Pt +1 Dt + 1
Xt +1 = ( + 1) . (3)
Dt + 1 Dt
Prices, payoffs and returns

I Excess returns:
I if you borrow a dollar at the interest rate R f and invest it in
an asset with return R, you pay no money out-of-pocket today
and get the payoff R − R f .
I This is a payoff with a zero price.
I Zero price does not mean zero payoff: it is a bet in which the
value of the chance of losing exactly balances the value of the
chance of winning.
I We denote any such difference between returns as an excess
returns, R e .
Prices, payoffs and returns

I Real (denominated in goods) or nominal stochastic discount

factors and returns:
I Assume that prices Ptn and payoffs Xtn+1 are nominal.
I Let Πt be the price index (e.g., CPI) at date t.
I Then the Euler equation is:

Πt
Ptn = Et [Mt +1 X n ].
Π t +1 t +1
| {z }
Mtn+1
Risk-free rates

I Start from Et (Mt +1 Rt +1 ) = 1.

I The risk-free rate Rtf is:

Rtf = 1/Et [Mt +1 ]. (4)

I Assume power utility: u (C ) = C 1−γ /(1 − γ).

I Marginal utility is u 0 (C ) = C −γ .
I The risk-free rate is then:

Ct + 1 − γ
Rtf = 1/Et [ β( ) ].
Ct
Power utility

I We define the utility form u (C ) = C 1−γ /(1 − γ) as power

utility, or constant relative risk aversion (CRRA).
I Note that the limit as γ → 1 is u (C ) = ln(C ).
Certainty case

I Let us assume for now that there is no uncertainty in the

economy.
I The risk-free rate is equal to:
1 Ct + 1 γ
Rtf = [ ] . (5)
β Ct

I Three effects link R f to macro factors:

1. Real interest rates are high when people are impatient, i.e β is
low.
2. Real interest rates are high when consumption growth is high.
3. Real interest rates are more sensitive to consumption growth if
γ is large.
Uncertainty case

I If z follows a Normal distribution, then:

1 2
E (e z ) = e E (z )+ 2 σ (z )
.

I We will use this result in many derivations ( Details ).

I Note that in continuos time you can get most of the following
results without relying on the assumption of normality.
Uncertainty case

I We will use lower case letters for log variables.

I Let ∆ct +1 denote the log consumption growth:

Ct + 1
∆ct +1 = ln = ln(Ct +1 ) − ln(Ct ).
Ct
I Assume that ∆ct +1 is normally distributed (i.e., gross
consumption growth is lognormally distributed).
I For example:
∆ct +1 = µ + σet +1 .
Uncertainty case

I The risk-free rate is known in advance so we can pull it out of

the expectation:
Ct + 1 − γ − 1
Rtf = { Et [ β ( ) ]} ,
Ct
= [ βEt (e −γ∆ct +1 )]−1 ,
γ2 2
= [ βe −γEt (∆ct +1 )+ 2 σt ( ∆ct +1 ) ] −1 ,
γ2 2
= [ βe −γµ+ 2 σ ] −1 .

I Note that in this case the risk-free rate is constant.

Uncertainty case

I Let us define δ as β = e −δ and rtf = ln(Rtf ).

I Then the log of the gross risk-free rate is:

γ2 2
rtf = δ + γEt (∆ct +1 ) − σ (∆ct +1 ),
2 t
γ2 σ 2
= δ + γµ − .
2
Uncertainty case

I Compare:

γ2 2
rtf = δ + γEt (∆ct +1 )− σ (∆ct +1 ).
2 t
with:
1 Ct +1 γ
Rtf = [ ] → rtf = δ + γ∆ct +1 .
β Ct
I The new term σ2 captures precautionary savings (cf. concern
for downside risk): ⇒ When consumption is more volatile,
people want to save more, driving down interest rates.
I Concern for low consumption states details .
Uncertainty case

I Note that with low risk-aversion (γ), the precautionary

savings effect is small.

I For example, for reasonable values of σ = 2% at annual

frequency, when γ = 1 (log utility) the precautionary savings
effect is equal to 2bp (i.e., 0.02%).
Uncertainty case

I For the power utility, the curvature parameter γ

simultaneously controls IES risk aversion St Petersburg paradox :

1. intertemporal substitution - aversion to consumption stream

that varies over time,
2. risk aversion - aversion to a consumption stream that varies
across states of nature,
3. and precautionary savings.
Uncertainty case
I Note that if the risk-free rate is constant, the term structure of
interest rate is completely flat at a constant level.
I For example, consider an asset that pays 1 unit at t + 1:
Pt ,t +1 = Et [Mt ,t +1 × 1t +1 ] = (R f )−1 .

I Consider an asset that pays 1 unit at t + 2:

Pt ,t +2 = Et [Mt ,t +2 × 1t +2 ]
= Et [Mt ,t +1 Pt +1,t +2 ]
= Et [Mt ,t +1 Et +1 [Mt +1,t +2 ]]
= Et [Mt ,t +1 Mt +1,t +2 ]
= (R f ) −1 × (R f ) −1
= (R f ) −2
where we have used the law of iterated expectations ( details ) and
Mt ,t +2 is the SDF from period t to t + 2 (i.e., for CRRA utility
Mt ,t +2 = β2 ( CCt +t 2 )−γ ).
Stock returns

I A stock is an asset that pays a dividend Dt = Ct (i.e., what

can be consumed is given by the aggregate dividends paid as
in Lucas (1978) as dividends cannot be stored) in all the
states of the world:
Ct +1 −γ
Pt = E t [ β ( ) (Pt +1 + Dt +1 )].
Ct
I The price/dividend ratio is then equal to:

Pt Pt Ct + 1 1 − γ Pt + 1
= = Et [ β ( ) (1 + )], (6)
Dt Ct Ct Dt + 1

where 6 is obtained by multiplying and dividing the term in

expectation by Dt +1 = Ct +1 .
Stock returns

I Iterate forward, impose transversality condition and assume

that log consumption growth is i.i.d. normal to get Details :
2 σ2
Pt βe (1−γ)µ+(1−γ) 2
= 2 σ2
.
Ct 1 − βe (1−γ)µ+(1−γ) 2

I The price/dividend ratio is constant.

I Intuition: the expectations of future consumption growth are
independent on current consumption.
I Note how this result resembles the formula for a growing
perpetuity with no uncertainty.
Stock returns

Price/dividend ratio
When expectations of future consumption and dividend growth are
independent of the state today, and utility is CRRA, the
price/dividend ratio is constant.
Risk premium

I The expected return on equity is constant:

Pt +1
Pt +1 + Dt +1 Dt +1 + 1 Dt +1
Et ( ) = Et ( P t )
Pt D
Dt
t
1 γµ+(2γ−γ2 ) σ2
= e 2 .
β
I The geometric risk premium is then:
2
1 γµ+(2γ−γ2 ) σ2
E (R ) βe 2
= 2 2 = e γσ .
Rf 1 γµ− γ 2σ
βe

I For the US, σ = 2%. Hence, the geometric risk premium is a

small number: γ × 0.0004, or 0.4 bp for log utility.
Realized returns

I Realized returns are equal to:

Pt +1
Dt +1 + 1 Dt +1 1 (γ−1)µ+(2γ−γ2 −1) σ2 µ+σet +1
Rt +1 = Pt
= e 2 e .
Dt
Dt β

I Therefore, realized returns are i.i.d. lognormal.

Volatility puzzle

I We have seen how the price/dividend ratio implied by the

standard model is constant.
I The data counterpart is instead very volatile.
I In the model, since the P/D ratio is constant, the volatility of
returns Rt +1 is approximately equal to the volatility of
dividend growth:
Dt + 1
σ(Rt +1 ) ' σ( ).
Dt
I This is counterfactual.
I This result is equivalent to Shiller (1981).
Volatility puzzle
I Shiller (1981) starting from Pt = Et (Mt +1 Xt +1 ) states that stock
prices represent the expected present value of future dividends (i.e.,
fundamentals).
I Define with P ∗ the realization of the fundamental value of stock i
i ,t
at time t (i.e., the discounted sum of realized future dividends
starting from t + 1).
I Rational investors set Pi ,t = Et (Pi∗,t ).
I Forecast errors Pi ,t− Pi∗,t must be uncorrelated with any
information available at t (including Pi ,t ).
I Since P ∗ ≡ Pi ,t + (Pi∗,t − Pi ,t ), then:
i ,t

Var (Pi∗,t ) = Var (Pi ,t ) + Var (Pi∗,t − Pi ,t ).

I As a result, in a no-arbitrage market:

Var (Pi∗,t ) > Var (Pi ,t ).

Volatility puzzle

Figure: Figure is from Shiller (AER, 1981). For more details see
Cochrane’s blog post ”Bob Shiller’s Nobel”.
Volatility puzzle

I P ∗ is the ex-post rational price.

I The actual price varies much more than P ∗ .
I Shiller interpreted it as psychological and social dynamics,
optimism and pessimism.
I Alternative explanation (we’ll come back to this point) is that
prices are not forecasting dividend growth, but long-run
returns.
Outline

Introduction

Utility-based asset pricing

Stochastic discount factor
Prices, payoffs and returns
Risk-free rates
Stock returns
Volatility puzzle

Risk correction
Idiosyncratic risk
Expected return-beta representation and market price of risk
Mean-variance frontier
The equity premium puzzle
Random walk and time-varying expected returns
Risk correction

I By definition, the covariance between M and X is:

cov (M, X ) = E (MX ) − E (M )E (X ).

I Thus, we can rewrite our basic pricing formula as:

Pt = Et (Mt +1 Xt +1 ) = Et (Mt +1 )Et (Xt +1 ) + covt (Mt +1 , Xt +1 ).

Risk correction

I Substitute in the pricing formula the risk-free rate, we obtain:

Et (Xt +1 )
Pt = + covt (Mt +1 , Xt +1 ).
Rtf
I The first term is the standard discounted present-value
formula.
I The second term is a risk-adjustment.
Risk correction

I Let us substitute back for M in terms of consumption in

previous equation:

Et (Xt +1 ) covt [ βu 0 (Ct +1 ), Xt +1 ]

Pt = + .
Rtf u 0 (Ct )

I Marginal utility U 0 (C ) declines as C rises.

I Thus an asset’s price is lowered if its payoff covaries positively
with consumption (i.e., if it pays high (low) when
consumption is high (low) and the marginal utility low (high)).
I Insurance is an extreme example.
Risk correction
An asset’s price is lowered if its payoff covaries positively with
consumption.

I Intuition:
I Investors do not like uncertainty about consumption.
I If you buy an asset whose payoff covaries positively with
consumption, it means that it pays off well when you are
already feeling wealthy and it pays off badly when you are
already feeling poor.
I Thus, that asset will make your consumption stream more
volatile.
Risk correction: example

I Consider what happens to the volatility of consumption if the

individual investor buys a little more ξ of payoff X :

σ2 (C + ξX ) = σ2 (C ) + 2ξcov (C , X ) + ξ 2 σ2 (X ),
I For small portfolio changes: ξ 2 σ2 (X ) ' 0.
I It’s the covariance between consumption and payoff that
determines the effect of adding a bit more of each payoff on
the volatility of consumption (not the volatility of the payoff!).
Risk correction

I For returns the basic pricing equation is:

Et [Mt +1 Rti +1 ] = 1.

I The asset pricing model says that, although expected returns

can vary across time and assets, expected discounted returns
should always be the same, 1.
Risk correction

I Apply the covariance decomposition to the basic pricing equation

for returns and use the definition of risk-free rate to obtain:

Et (Rti +1 ) − Rtf = −Rtf covt [Mt +1 , Rti +1 ],

covt [u 0 (Ct +1 ), Rti +1 ]
= − .
Et [u 0 (Ct +1 )]

I All assets have an expected return equal to the risk-free rate, plus a
risk adjustment.
I Assets whose returns covary positively with consumption make
consumption more volatile, and so must promise higher expected
returns to induce investors to hold them.
I Note that assets that covary negatively with consumption, such as
insurance, can offer expected rates of return that are lower than the
risk free rate, or even negative (net) expected returns.
Idiosyncratic risk

I You might think that an asset with a volatile payoff is ”risky”,

and thus should have a large risk correction.
I However, if the payoff is uncorrelated with M, the asset
receives no risk correction to its price and pays an expected
return equal to the risk-free rate!
I This prediction holds even if X is very volatile and investors
highly risk averse.
I Why? No first order effect on the variance of consumption
stream:

σ2 (C + ξX ) = σ2 (C ) + 2ξcov (C , X ) + ξ 2 σ2 (X ).
Idiosyncratic risk
Only the component of payoff correlated with the SDF generates
an extra return.

I Idiosyncratic risk, uncorrelated with the SDF, generates no

premium.
I Only systematic risk generates risk correction.
Idiosyncratic risk

I Decompose the payoff as follows:

X = proj (X |M ) + ε.

I Projecting X on M is like regressing X on M without a

constant:
E (MX )
proj (X |M ) = M.
E (M 2 )
Idiosyncratic risk

I The residuals ε are orthogonal to the right-hand side variable

M: E (Mε) = 0, which means that the price of ε is zero.
I The price of the projection of X on M is the price of X :

E (MX )
P (proj (X |M )) = E (M M ) = E (MX ) = P (X ).
E (M 2 )
I The projection of X on M is that part of X which is perfectly
correlated with M.
I The idiosyncratic component of any payoff is that part that is
uncorrelated with M.
I Only the systematic part of payoff accounts for its price.
β pricing model

I Start from: Et (Mt +1 Rti +1 ) = 1.

I Rewrite it using the definition of covariance as:

Et (Rti +1 ) − Rtf = −Rtf covt [Mt +1 , Rti +1 ].

I Multiply and divide by vart [Mt +1 ] and rearrange terms:

covt [Mt +1 , Rti +1 ] vart [Mt +1 ]

Et (Rti +1 ) = Rtf + (− )( ),
vart [Mt +1 ] Et [ M t + 1 ]
| {z }| {z }
β i,M λM

where β i,M is the regression coefficient of the return R i on M.

β pricing model
E (Rti +1 ) = Rtf + β i,M λM is a beta-representation or beta-pricing
model. Note that λM is independent of the asset i and depends on
the volatility of the SDF. It is called the market price of risk. β i,M
is the quantity of risk. The excess return of asset i is equal to
the quantity of risk of this asset times the price of risk.
Consumption CAPM (CCAPM)

I Start from, assuming power utility:

Ct + 1 − γ i
Et [ β ( ) Rt +1 ] = 1.
Ct

Ct +1
I Assume that Rti +1 and Ct are jointly lognormal.
I Then:
i
Et [e log β−γ∆ct +1 +rt +1 ] = 1,
where lower case letters denote logs.
Consumption CAPM (CCAPM)

I Apply log-normal trick to the return on asset i and to the risk-free

rate to get the CCAPM Details :
Rti +1 1 Ri
Et [log( f
)] + Vart log Rti +1 = log Et ( tf+1 ) = γCovt (∆ct +1 , log Rti +1 ).
Rt +1 2 Rt + 1

I Intuition:

I Differences in average excess returns (corrected for Jensen’s

inequality term) depend only on covariances with consumption
growth;
I Assumption of joint log normality means that we cannot apply
same reasoning to assets with highly nonlinear payoffs (for
example, options).
β pricing model

I In a more general set up, rewrite Euler equation in terms of

the log SDF mt +1 and the log return rti +1 :

1
Et (mt +1 ) + Et (rti +1 ) + Vart (mt +1 ) +
2
1
+ Vart (rti +1 ) + Covt (mt +1 , rti +1 ) = 0.
2
I Similar equation holds for the risk-free rate rtf .
I Subtract risk-free rate from expected log return:
1
Et (rti +1 ) − rtf + Vart (rti +1 ) = −Covt (mt +1 , rti +1 ).
2
β pricing model

I Let r˜te,i
+1 be the excess return corrected for the Jensen term:
e,i
r˜t +1 = rti +1 − rtf + 12 Vart (rti +1 ). Then:

Et (r˜te,i e,i
+1 ) = −Covt (mt +1 , r˜t +1 ).
Special case: consumption CAPM

I For power utility - also known as Consumption-CAPM, the log

SDF depends only on consumption growth and is equal to:

mt +1 = log β − γg − γ(∆ct +1 − g ),

where g is the average consumption growth.

I In this case, the expected excess return is equal to:

+1 ) = γcovt ( ∆ct +1 − g , r˜t +1 ).

Et (r˜te,i e,i

I Assets whose returns covary positively with consumption must

promise positive expected returns to induce investors to hold
them.
Special case: consumption CAPM

I Assume that we have many excess returns.

I A first stage regression of each excess return r˜te,i
+1 (corrected
for the Jensen term) on a constant and consumption growth
gives a corresponding slope coefficient of:

β i,∆c = cov (∆ct +1 , rte,i 2

+1 ) /σ∆ct +1 .

I A second stage regression of all average (i.e., expected) excess

returns ET (r˜e,i ) on a constant and the corresponding β i,∆c
leads to the estimation of the market price of risk λ.
I This two-stage procedure is advocated by Fama and MacBeth
(1973).
Special case: consumption CAPM

I The expected excess return can be rewritten in terms of the

market price of risk λ∆c and the amount of risk β:

+1 ) = covt [ ∆ct +1 , rt +1 ] /σ∆ct +1

Et (r˜te,i e,i 2 2
γσ∆c .
| {z } | {zt +}1
Amount of risk β i,∆c Price of risk λ∆c

I The market price of risk λ measures the expected excess

return per unit of risk.
I Expected returns should increase linearly with their betas on
consumption growth.
Special case: consumption CAPM

I The factor risk premium λ∆c is determined by risk aversion

and the volatility of consumption.

I Intuition: The more risk averse people are, or the riskier their
environment, the larger an expected return premium one must
pay to get investors to hold risky (high beta) assets.
Present value model

I So far we used a two-period valuation: price at t and payoff

at t + 1.
I If we prefer to take into account the entire (infinite) cash flow
stream {Dt +j }:
∞ ∞
U 0 ( Ct + j )
P t = Et ∑ βj U 0 ( Ct )
D t + j = ∑ Mt,t +j Dt +j .
j =1 j =1

I If the equation above holds at time t and t + 1, then we get

our standard expression:

Pt = Et [Mt +1 (Pt +1 + Dt +1 )].

I Infinite-period model = 2-period model.

Present value model
I Consider an asset that pays a dividend flow {Dt }, and assume that
this asset has a constant expected rate of return r :
Pt +1 + Dt +1
1 + r = Et ( ).
Pt

I Rearrange and get:

E t Dt + 1 Et Pt +1
Pt = + .
1+r 1+r

I Iterate forward and assume transversality condition (TI):

∞
E t Dt + k
Pt = ∑ k
.
k =1 (1 + r )

I TI rules out ”bubbles” in which prices grow so fast that people will
buy now to resell at higher prices.
Present value model: Gordon formula

I Assume that dividends grow at a constant average rate g :

Dt +k
Et ( ) = 1 + g, k = 1, 2, . . .
Dt +k −1

I In this special case, the present value model is:

∞ ∞
Et (Dt +k ) 1+g k (1 + g )
Pt = ∑ ( 1 + r ) k
= ∑ Dt (
1 + r
) = Dt
r −g
.
k =1 k =1

I The Gordon formula is defined as:

Pt 1+g
= .
Dt r −g
Stochastic discount factor

I A stochastic discount factor (SDF) is a stochastic process

{Mt +1 } such that for any security with stochastic payoff Xt +1
at time t + 1, the price of that security at time t is:

Pt = Et [Mt,t +1 Xt +1 ]. (7)

I For example, in a representative agent economy with CRRA

preferences the SDF is:
Ct +1 −γ
Mt,t +1 = β( ) .
Ct
Stochastic discount factor

I What if there is more than one agent?

I If markets are complete, we will show that the SDF is unique.
I If markets are incomplete, but there are no arbitrage
opportunities, we will show that there always exist a positive
SDF, but there can be several SDF that satisfy equation 7.

I Markets are complete when there are enough assets to span

all the possibile states of the world. In other words, there
exists at least one asset that pays 1 in state s and 0
otherwise, for s = 1, ..., S.
Stochastic discount factor

I Start again from equation Et (Mt +1 Rt +1 ) = 1:

1 = Et (Mt +1 Rti +1 ) = Et (Mt +1 )Et (Rti +1 ) + ρt (M, R i )σt (R i )σt (M ),

Cov (X ,Y )
where ρt is the correlation coefficient ρ(X , Y ) = σX σY .

I Then:
σt (M )
Et (Rti +1 ) − Rtf = −ρt (M, R i ) σt (R i ).
Et ( M )
I This is an equilibrium relation (i.e., remember: it comes from the
optimality condition): if someone is holding these assets it must be
that their returns compensate for their risk.
Mean-variance frontier
The set of means and variances of asset returns is limited.
Mean-variance frontier

I Correlation coefficients cannot be > 1 ⇒ all assets priced by

the SDF M must obey:

σ (M )
|E (R i ) − R f | ≤ σ (R i ).
E (M )
I The Hansen and Jagannathan (JPE 1991) mean-variance
frontier is determined by |ρM,R i | = 1.
I In a graph of expected returns E (R ) as a function of their
volatility σ(R ), all assets must lie inside the wedge-shaped
mean-variance region.
Mean-variance frontier

E(R)

Slope:
sigma(M)/E(M) Mean Variance Frontier

Idiosyncratic risk
R^f

some asset returns

sigma(R)
Mean-variance frontier

I The mean-variance frontier answers a naturally interesting

question: how much mean return can you get for a given level
of variance?
I The Hansen and Jagannathan bounds are popular because
any model that does not satisfy them will be at odds with the
data.
Mean-variance frontier

I Some of you might remember from previous finance classes a

mean-variance frontier that looks like a hyperbola.
I Why in the previous figure I instead plot a cone?
I First, note that the plot refers to the mean-variance frontier of
returns. The mean-variance frontier of excess returns would
have the tip of the cone in the origin.
I Second, the slope of the mean-variance frontier in our model
is a constant.
Mean-variance frontier

I All frontier returns R mv are perfectly correlated with the SDF

and with each other (the frontier is generated for
|ρt (M, R i )| = 1).
I Returns on the upper part of the frontier are perfectly
negatively correlated with the discount factor and hence
positively correlated with consumption. They are maximally
risky and get the highest expected returns.
I Returns on the lower part of the frontier are perfectly
positively correlated with the discount factor and hence
negatively correlated with consumption. They provide the
best insurance against consumption fluctuations.
Mean-variance frontier

I Consider a payoff M/E (M 2 ).

I Its price is E (M 2 )/E (M 2 ) = 1, so it is a return.
I It is on the mean-variance frontier.
I Thus, if we know M, we can construct a mean-variance
efficient return.
I Notice that M is not on the mean-variance frontier. In fact,
P (M ) = E (M 2 ) > 0, so it is not a return.
Mean-variance frontier

I All frontier returns are perfectly correlated with each other.

I We can span any frontier return from two such returns (this is
also called the two-fund theorem).
I Example: pick R f and any single frontier return R m :

R mv = R f + a(R m − R f ),

for some number a.

Mean-variance frontier

I In addition, there exist constants a, b, c, d such that:

M = a + bR mv and R mv = d + eM.

I Thus, any mean-variance efficient return carries all pricing

information.
I Given a mean-variance efficient return and the risk-free rate,
we can find a discount factor that prices all assets and vice
versa.
Mean-variance frontier

I Given a discount factor, expected returns can be described in

a single-beta representation using any mean-variance efficient
return (except the risk-free rate):

E (R i ) = R f + β i,mv [E (R mv ) − R f ].

I This is Roll’s theorem.

Mean-variance frontier

I Proof:
I Start from: E (R i ) = R f + β i ,M λM .
I Recall that |ρ(M, R mv )| = 1.
I Thus: E (R i ) = R f + β i ,mv λM .
I Note that: E (R mv ) = R f + β i ,mv λM = R f + λM .
I Therefore: E (R i ) = R f + β i ,mv [E (R mv ) − R f ].
I Last step depends on the fact that the beta models applies to
every return including R mv itself, and R mv has a β of 1 on
itself.
Mean-variance frontier

I Even though the means and standard deviations of returns fill

out the space inside the mean-variance frontier, a graph of
mean returns versus betas should yield a straight line.
Mean-variance frontier

I In the plot of the mean variance frontier we have seen that

the return is decomposed into a priced or systematic
component and a residual or idiosyncratic component.
I The priced part is perfectly correlated with the discount
factor, and hence perfectly correlated with any frontier return.
I The residual part generates no expected return, and it is
uncorrelated with the discount factor or any frontier return.
Mean-variance frontier

I Asset inside the frontier or even on the lower portion of the

frontier are not ”worse” than assets on the frontier. The
frontier and its internal region characterize equilibrium asset
returns, with rational investors happy to hold all assets.
The equity premium puzzle

I The Sharpe ratio SR measures how much return the investor

receives per unit of volatility:

E (R i ) − R f
SR = .
σ (R i )
The equity premium puzzle

I Over the last 50 years in the US, real stock returns have
averaged 9% per year with a standard deviation of 16%.
I The real return on T-Bills has been around 1% per year.
I Thus the Sharpe ratio has been about 0.5 for an annual
investment horizon.
The equity premium puzzle

I If you borrow and put more money into a security, you can
increase the mean return of your position, but you do not
increase the SR, since the standard deviation increases at the
same rate as the mean.
The equity premium puzzle

I The slope of the mean-variance frontier is the largest available

Sharpe ratio:

E (R i ) − R f E (R mv ) − R f σ (M )
| i
| ≤ | mv
|= = σ (M )R f .
σ (R ) σ (R ) E (M )

I The T-bill rate is not very risky, so E (M ) cannot be to

different from the inverse of the mean T-bill rate:
E (M ) ≈ 0.99.
I The slope of the frontier is governed by the volatility of the
discount factor.
I As a result, σ(M ) > 0.5.
The equity premium puzzle

I Now, assume that consumption is lognormal and power utility

(M = β(Ct +1 /Ct )−γ ). Then: algebra

E (R mv ) − R f
q
σ (M ) 2 σ2
| mv
|= = eγ ∆c − 1 ≈ γσ∆c .
σ (R ) E (M )
I The slope of the mean variance frontier is higher if the
economy is riskier - if consumption is more volatile - or if
investors are more risk averse.
The equity premium puzzle

I Aggregate nondurables and services consumption growth has

a mean of 2% and a standard deviation of 1%, implying a risk
aversion coefficient of 50!
I This is the equity premium puzzle of Mehra and Prescott
(1985).
I Substitute γ = 50 in the expression for the risk free rate rtf :
the implied risk free rate is close to 100%!
I This is the risk free rate puzzle (Weil 1989).
The equity premium puzzle

I Note that in equation:

γ2 2
rtf = δ + γEt (∆ct +1 ) − σ (∆ct +1 ),
2 t
when risk aversion increases, the precautionary saving term
starts to offset the intertemporal substitution term.
Risk-free rate
f
R (γ)
60

40

20

0
Rf (%)

−20

−40

−60

−80
0 50 100 150 200 250
γ

Figure: Risk-free rate as a function of the risk aversion coefficient γ. The

risk-free rate follows (6) with δ = 0.01, Et (∆c ) = 0.01, σt2 (∆c ) = 0.01.
The equity premium puzzle

I Consequences of the equity premium:

I people are a lot more risk averse than we might have thought,
I or, the stock returns of the last 50 years were largely good luck
rather than an equilibrium compensation for risk,
I or, something is deeply wrong with the model (including utility
function and use of aggregate consumption data).
Random walk and time-varying expected returns

I Go back to Pt u 0 (Ct ) = Et [ βu 0 (Ct +1 )Xt +1 ].

I Consider a short interval of time such that the asset pays no
dividend between t and t + 1 and such that β is close to 1.
I If over the short interval consumption does not change much, then
asset prices follow a martingale (a random walk if σt2 (et +1 ) is
constant Details ):
Pt = Et [Pt +1 ].
I In this case, expected returns Et [Pt +1 /Pt ]
= 1 should be constant
and not predictable: returns should be like coin flips.
I The fact that expected returns should be constant does not mean
that realized returns should be constant as well.
Random walk and time-varying expected returns

I Since consumption and risk aversion do not change much day

to day, we might expect the random walk view to hold pretty
well on a day-to-day basis:

Pt = Et [Mt +1 (Pt +1 + Dt +1 )]
Pt = Et [Pt +1 ]
Pt + 1 ≈ Et [pt +1 ] + et +1 = Pt + et +1

I In fact, at this frequency the random walk view has been

remarkably successful.
Random walk and time-varying expected returns
S&P daily returns
0.15

E(R) =0.00

0.1 AC(1) =−0.06

0.05

−0.05

−0.1
1990 1995 2000 2005 2010

Figure: Daily total returns of the S&P 500 Composite index. Sample:
February 1, 1988 - December 9, 2011. Data are from Datastream.
Random walk and time-varying expected returns

I Note that the random walk assumption implies that only

current market prices matters in predicting future prices.
I In other words, in a regression like:

Pt +1 = α + β 1 Pt + β 2 Xt + et +1 ,

where Xt is any signal known at time t, the coefficient β 2 = 0.

I If not, investors would ”read” the signal and buy/sell till
Pt +1 = Pt + et +1 .
I This is the traditional concept of market efficiency.
I However, evidence has accumulated that long-horizon
excess returns are quite predictable.
Random walk and time-varying expected returns

I To start thinking about predictability, start from:

Et (Rt +1 ) − Rff ≈ γt σt (∆ct +1 )σt (Rt +1 )ρt (Mt +1 , Rt +1 ),

where the time-index is there to remind us that the relation

applies to conditional moments.
I In the special case that random variables are i.i.d., the
conditional and unconditional moments are the same.
I For example, conditional on tonight’s weather forecast, you
can better predict rain tomorrow than just knowing the
average rain for that date.
Random walk and time-varying expected returns

I If you look at:

Et (Rt +1 ) − Rff ≈ γt σt (∆ct +1 )σt (Rt +1 )ρt (Mt +1 , Rt +1 ),

we can see that returns could be somewhat predictable.

I If the conditional variance of returns changes over time, we
might expect the conditional mean return to vary as well.
I In this case, the return would move in and out along a line of
constant Sharpe Ratio.
I This explanation does not seem to help much in the data:
variables that forecast means do not seem to forecast
variances and vice versa.
Random walk and time-varying expected returns

I Unless we want to focus on the conditional correlation,

predictable excess returns have to be explained by changing
risk (σt (∆ct +1 )), or changing risk aversion (γ).
I It is not plausible that risk or risk aversion change at daily
frequencies.
I It is much more plausible that risk and risk aversion
change over the business cycle, and this is exactly the
horizon at which we see predictable excess returns.
I Old view of the world: returns are not predictable (like coin
flips) and expected returns are constant.
I New view of the world: there are variables that predict returns.
I We can run predictive regressions of future returns and any
signal today.
I We can test for momentum, mean-reversion, profit-taking, etc.
 Table: Return-Forecasting Regressions
σ [Et (R e )]
h b t (b ) R2 σ[Et (R e )] Et (R e )

1 month 0.3 (2.9) 0.01 0.45 0.85

1 year 3.8 (2.6) 0.09 5.46 0.76
5 year 20.6 (3.4) 0.28 29.32 0.62

Notes: The univariate regressions are: Rte→t +h = a + b × Dt /Pt + ε t +h , where

Rte→t +h denotes the h-year ahead stock market excess return. OLS regressions of
excess returns (CRSP value-weighted return less 3-month Treasury Bill) on the CRSP
value-weighted dividend-price ratio. Data are annual data, 1947-2009. Source:
Cochrane (2011).
Return-Forecasting Regressions

I Consider the 1-year regression:

I t-stat is ”significant”,
I R 2 not impressive,
I coefficient estimate is large: A one percentage point increase
in dividend yield forecasts a nearly four percentage point
higher return,
I standard deviation of expected returns is large and almost as
large as the expected value (this is interesting: according to
old view expected returns should be constant!),
I slope coefficients and R 2 increase with the horizon.
I R 2 not great measure of predictability because a + B (D/P )t
varies little compared to the variance of returns.
Return-Forecasting Regressions
4 x D/P
Return
25

20

15

10

1950 1960 1970 1980 1990 2000 2010

Figure: Dividend yield and following 7-year return. Data are from CRSP
value weighted market index.
Return-Forecasting Regressions

I High prices, relative to dividends, have reliably preceded many

years of poor returns. Low prices have preceded high returns.
I Notice that high D/P tend to occur when the economy is
doing badly.
I Possible macro-explanation that we will try to explore: in bad
times individuals require high returns to be compensated
for bearing risk.
Cash flow or return predictability?

I Consider the Campbell and Shiller (1988) decomposition.

I Start from:

1 = Rt−+11 Rt +1 = Rt−+11 (Pt +1 + Dt +1 )/Pt .

I Multiply both sides by the price-dividend ratio Pt /Dt :

Pt Pt +1 Dt +1
= Rt−+11 (1 + ) .
Dt Dt + 1 Dt
Cash flow or return predictability?

I Taking logs leads to:

pt − dt = −rt +1 + ∆dt +1 + log(1 + e pt +1 −dt +1 ).

I A first-order Taylor approximation of the last term around the

mean price-dividend ratio P/D gives:
P
pt − dt = −rt +1 + ∆dt +1 + log(1 + )+
D
P/D
+ (pt +1 − dt +1 − (p − d )),
1 + P/D
= −rt +1 + ∆dt +1 + k + ρ(pt +1 − dt +1 ).
Cash flow or return predictability?

I Iterating forward one obtains:

∞ ∞
dt − pt ≈ ∑ ρj −1 rt +j − ∑ ρj −1 ∆dt +j + ρj (dt +j − pt +j ) (8)
j =1 j =1

where ρ ≈ 0.96 is a constant of approximation (NB: this is equation

1 in Cochrane (2011)).
Cash flow or return predictability?

I Consider now OLS regressions of weighted long-run returns and

dividend growth on dividend yields dpt = dt − pt :
∞
∑ ρj −1 rt +j = ar + brj dpt + etr +j
j =1
∞
∑ρ j −1
∆dt +j j
= ad + b∆d dpt + etd+j
j =1
j
dpt +j = adp + bdp dpt + etdp+j

I Equation 8 implies that the long-run regression coefficients add up

to one:
j j
1 ≈ brj − b∆d + ρj bdp . (9)
To see this, regress left- and right-hand size of 8 on dpt .
Cash flow or return predictability?

I Equations 8 and 9 imply that if we lived in a iid world,

dividend yields would be constant since expected future
returns and dividends would never change.
I However, dividend yields are not constant.
I As a result, dividend yields must forecast:
1. long-run returns,
2. long-run dividend growth,
3. a rational bubble of ever high prices.
Cash flow or return predictability?

I Equation 8 holds ex-post, and thus also ex-ante:

∞
dpt ≈ Constant + Et ∑ ρj −1 (rt +j − ∆dt +j ) + ρj Et (dpt +j ).
j =1
Cash flow or return predictability?

I Now multiply both sides by dpt − E (dpt ):

∞ ∞
var (dpt ) ≈ cov (dpt , ∑ ρj −1 rt +j ) − cov (dpt , ∑ ρj −1 ∆dt +j ) +
j =1 j =1
j
+ρ cov (dpt , dpt +j ),

where we used cov (x, y ) = E (xy ) − E (x )E (y ).

I How big is each source of variation?

 Table: Long-Run Regression Coefficients

Coefficient
Method and Horizon brj j
b∆d j
ρj bdp
Direct k=15 1.01 -0.05 -0.11
VAR k=15 1.05 0.35 0.22
VAR k=∞ 1.35 0.45 0.00

Notes: This table reports long-run regression coefficients. For example, brj in
∑j∞=1 ρj −1 rt +j = ar + br dpt + etr +j . Sample is 1947-2009, annual data. ”Direct”
j

regression estimates are calculated using 15-year ex post returns, dividend growth, and
dividend yields as left-hand variable. The ”VAR” estimates infer long-run coefficients
from 1-year coefficients. Source: Cochrane (2006).
Cash flow or return predictability?

I The long-run return coefficients are all a bit larger than 1.

I The dividend growth forecasts are small, statistically
insignificant, and the positive point estimates go the wrong
way.
I The 15-year dividend yield forecast coefficient is also
essentially zero.
Cash flow or return predictability?

I To sum up:
I All price-dividend ratio volatility corresponds to variation in
expected returns. None corresponds to variation in expected
dividend growth, and none to rational bubbles.

I This is new: under the old idea that returns are not
predictable, high prices relative to current dividends were
considered to reflect expectations that dividends would rise in
the future.
Cash flow or return predictability?

I Implication of the new results on predictability:

I Suppose prices are low with respect to current dividends (i.e.,
D/P is high).
I Old view: Prices are low because future dividends are expected
to be low. Expected returns are expected to be constant and
equal to zero.
I New view: Prices are low because individuals want to be
compensated to bear risk. Expected returns are large.
Dividends are expected not too change much over time.
Cash flow or return predictability?

I Predictability is pervasive across markets:

I stocks: dividend yield forecast returns, not dividend growth;
I treasuries: a rising yield curve signal better 1-year return for
long-term bonds, not higher future interest rates;
I bonds: much variation in credit spreads over time and across
firms or categories signal returns, not default probabilities;
I foreign exchange: international interest rate spread signal
returns, not exchange rate depreciation;
I sovereign debt: high level of sovereign or foreign debt signal
low returns, not higher government or trade surpluses;
I houses: high price/rent ratio signal low returns, not rising
rents or prices that rise forever.
House prices and rents

20 x rent
CSW price
OFHEO
7.8

7.6 Price

7.4
log scale

7.2

7 20 x Rent

6.8

1960 1970 1980 1990 2000 2010

Date
 Table: House prices and rents

b t (b ) R2
Return 0.12 2.52 0.15
Rent change 0.03 2.22 0.07
Rent/price 0.95 16.20 0.90

Notes: Regressions of log annual housing returns, log rent growth and log rent/price
ratio on the rent/price ratio. Sample is 1947 - 2010. Source: Cochrane (2011).
Cash flow or return predictability?

I Housing regressions are almost the same as the stock market

regressions.
Cash flow or return predictability?

I You can repeat the same exercise with other variables, like
earnings or book values, and find similar results.
The frontier of research

σt (Mt +1 )
Et (Rt +1 ) − Rtf = − σt (Rt +1 )ρt (Mt +1 , Rt +1 ),
Et ( M t + 1 )
≈ γt σt (∆ct +1 )σt (Rt +1 )ρt (∆ct +1 , Rt +1 ).
The frontier of research

I Three interesting paths:

1. γt : time-varying risk aversion, a la Cambpell and Cochrane
(1999).
2. σt (∆ct +1 ): long run risk, a la Bansal and Yaron (2004).
3. σt (∆ct +1 ): disaster risk, a la Barro (2006).
Appendix
Proof: Log-normal trick (1/2)
I Recall: if r is distributed as N (µ, σ ), then its density is:

1 2 2
f (r ) = √ e −(r −µ) /2σ dr .
2πσ
I If µ = 0 and R is LN (µ, σ) then:

Z ∞
1 2 2
E (R ) = E (e r ) = er √ e −r /2σ dr
−∞ 2πσ
Z ∞
1 2 2 2
= √ e (2r σ −r )/2σ dr
−∞ 2πσ
Z ∞ −(r −σ2 )2 +σ4
1
= √ e 2σ2 dr
−∞ 2πσ
Z ∞ −(r −σ2 )2
1 2 1
= e 2σ √ e 2σ2 dr
−∞ 2πσ
| {z }
=1
1 2
= e 2σ .
Proof: Log-normal trick (2/2)

I In the general case where µ 6= 0, consider the following change of

variable: y = r − µ e dy = dr .
I Then:

Z ∞ (r − µ )2
1
E (R ) = E (e r ) = er √ e 2σ2 dr
−∞ 2πσ
Z ∞ y2
1
= e µ +y √ e 2σ2 dy
−∞ 2πσ
Z ∞ y2
1
= eµ ey √ e 2σ2 dy
−∞ 2πσ
1 2
= e µ+ 2 σ .

Back
Proof: Constant price/dividend ratio (I/II)
I Start from Euler equation for the price/dividend ratio 6 and recall
that D = C :
Pt Ct +1 −γ P C
= Et [ β ( ) (1 + t +1 ) t +1 ]
Ct Ct Ct +1 Ct
Pt C P
= Et [ β( t +1 )1−γ (1 + t +1 )]
Ct Ct Ct +1

I Solve the recursion by iterating forward for T periods and using the
law of iterated expectations:
T Ct +j 1−γ
Pt C P
= Et ∑ [ β j ( ) ] + βT Et [( t +T )1−γ t +T ] .
Ct j =1
C t Ct Ct
| {z }
→0

I Second term goes to 0 for T → ∞: transversality or no bubble

condition.

Back
Proof: Constant price/dividend ratio (II/II)
I We can write:

Pt Ct +j 1−γ
Ct
= ∑ βj Et [( Ct
) ]
j ≥1
Ct +j Ct +1 1−γ
= ∑ βj Et [( Ct +j −1 )1−γ × . . . ( Ct
) ]
j ≥1
j −1
Ct +j −k 1−γ
= ∑ β j Et [ ∏ (
Ct +j −1−k
) ]
j ≥1 k =0
2 σ2
= ∑ βj [e (1−γ)µ+(1−γ) 2 ]j
j ≥1
2 σ2
βe (1−γ)µ+(1−γ) 2
= ,
2 σ2
1 − βe (1−γ)µ+(1−γ) 2

using the fact that log consumption growth is i.i.d. normal, the
lognormal formula and the limit of a geometric series.
Back
CCAPM
I Start from Euler equation for returns under power utility:

Ct + 1 − γ i
Et [ β( ) Rt +1 ] = 1.
Ct

I Assume the log of the return and the log consumption growth are jointly
normal. Then:
i
1 = Et [e log β−γ∆ct +1 +rt +1 ],
2
log β−γEt ∆ct +1 +Et rti +1 + γ2 σσ2 + 12 Vart (rti +1 )−γCovt (∆ct +1 ,rti +1 )
1 = e ∆c ,
γ 2 1
0 = log β − γEt ∆ct +1 + Et rti +1 + σ2 + Vart (rti +1 ) − γCovt (∆ct +1 , rti +1 ).
2 σ∆c 2
I Do same for risk-free rate:

γ2 2
rtf+1 = −log β + γEt ∆ct +1 − σ .
2 ∆c
I Subtract the risk-free rate from return on asset i:

Rti +1 1 Ri
Et [log( )] + Vart log Rti +1 = log Et ( tf+1 ) = γCovt (∆ct +1 , log Rti +1 ).
Rtf+1 2 Rt + 1

Back
Martingale
A martingale is a sequence of random variables for which, at a
particular time in the realized sequence, the expectation of the
next value in the sequence is equal to the present observed value
even given knowledge of all prior observed values at current time.
Back
Law of iterated expectations

I Simple version of the law of iterated expectations (LIE) says that

(see Woolridge (2010)):

E (Y ) = EX [E (Y |X )].

I Think of X as a discrete vector taking possible values

C1 , C2 , . . . , CN with probabilities p1 , p2 , . . . , pN then:

E (Y ) = p1 E (Y |X = c1 ) + p2 E (Y |X = c2 ) + . . . + pN E (Y |X = cN ).

I E (Y ) is a weighted average of the averages.

I In words, the unconditional expectation of the conditional

expectation of Y conditional on X is equal to the unconditional
expectation of Y.
Law of iterated expectations

I With time-series, the LIE implies that your current best guess of
your best guess next period of the realization of X two periods from
now is equal to your current best guess of X two periods from now:

Et [Et +1 (Xt +2 )] = Et (Xt +2 ).

Back
Elasticity of inter-temporal substitution
I Consider a utility function over consumption bundles of this kind:

U = U ( ct , ct + 1 , . . . , ct + j , . . . ) .

I The elasticity of inter-temporal substitution (IES) between period

t + j and t + i is equal to:
∂(ct +j /ct +i )
ct +j /ct +i ∂ ln(ct +j /ct +i )
IES = − ∂(U =− ,
ct +j /Uct +i ) ∂ ln(Uct +j /Uct +i )
Uct +j /Uct +i

where Uc denote the partial derivative of utility with respect to

consumption.
I The IES measures the inverse of the percentage change in marginal
utility at j and i that one percent change in the ratio of
consumption at the same dates leads to.
Back
Elasticity of inter-temporal substitution
I Consider now the special case of additively separable CRRA utility
functions:
1− γ
ci
U ( ci ) =
1−γ
Uc = c −γ
Uct +j ct + j − γ
= ( )
Uct +i ct + i
Uct +j ct + j
ln( ) = −γ ln( )
Uct +i ct + i
ct + j 1 Uc
ln( ) = − ln( t +j )
ct +i γ Uct +i
Uc
c
∂ ln( ctt ++ji ) ∂[− γ1 ln( Uct +j )] 1
t +i
− Uc
= − Uc
=
∂ ln( Uct +j ) ∂ ln( Uct +j ) γ
t +i t +i

Back
Elasticity of inter-temporal substitution

I Note that the marginal utility of consumption is decreasing in

consumption.
I Therefore, we know that when (ct +j /ct +i ) ↑, then (Uct +j /Uct +i ) ↓
I The (inverse of the) IES says by how much.

I For CRRA utility, IES = 1 .

γ
I The larger is γ, the smaller the IES, and the less substitutable
consumption between periods j and i is.
I In an inter-temporal set-up, the larger is γ, the more an individual
will want to keep a smooth inter-temporal consumption profile.
Back
Coefficient of relative risk-aversion

I The Arrow-Pratt coefficient of relative risk-aversion (RRA) is

defined as:
−cUcc
RRA = .
Uc
I For the CRRA utility functions:

−c (−γ)c −γ−1
RRA = = γ.
c −γ
I Therefore, the RRA is the inverse of the IES.

I The parameter γ controls at the same time IES and RRA.

Back
St. Petersburg paradox (I/II)

I Consider a bet involving the toss of a fair coin.

I Start with $2.

I The game then is as follows:

1. toss a coin
2. if ”heads” double the money on the plate
3. if ”tail” game ends and win what is on the plate
I The expected payoff is

1 1 1
(2) + (4) + (8) + . . . = 1 + 1 + . . . = ∞.
2 4 8
I Therefore, the bet is a random variable with a infinite expected
value.
Back
St. Petersburg paradox (II/II)

I What is the value of this bet?

I If investor are risk-neutral (i.e., linear utility), the price should be ∞.

I Experimental investigations show investors are willing to pay little

for this bet. How come?
I Assume investors are averse to risk and have log utility and initial
wealth W . The change in utility after paying c for the bet is:
∞
1 h i
∆E (u ) = ∑ k
ln ( W + 2k
− c ) − ln ( W ) <∞
k =1 2

where k > 0 is the number a coin has been tossed.

I If W = $1m, then c = $20.88; if W = $1K , then c = $10.95.
Back
CRRA utility
I To show that the CRRA utility function1 converges to logarithmic
as γ → 1 we use the l’Hôpital’s rule.
I Differentiate both numerator and denominator with respect to γ
and then take the limit of the ratio as γ → 1.

C 1− γ − 1 e (1−γ) ln C − 1
lim = lim
γ →1 1−γ γ →1 1−γ
−e (1−γ) ln C ln C
lim = lim C 1−γ ln C
γ →1 −1 γ →1

ln C lim C 1−γ = ln C
γ →1

I Recall the l’Hôpital’s rule: if limx →c f (x ) = limx →c g (x ) = 0 and

f 0 (x ) f (x ) 0
limx →c g 0 (x ) exists, then limx →c g (x ) = limx →c gf 0 ((xx )) .

back

1 We add a −1 at numerator to get the exact result. Note that adding or

subtracting a constant does not change the utility function.
Precautionary savings (I/III)

I Consumption variance has an effect on the level of

consumption: higher variance lowers consumption today, and
therefore increases saving and decreases the interest rate.
I This effect depends on the characteristics of the utility
function: it is a positive third derivative of the utility function
that leads to precautionary savings.
back
Precautionary savings (II/III)

I Intuition (based on Jensen’s inequality):

I consider for simplicity the case in which β = 1/R f and
E [Ct +1 /Ct ] = 1.
I consider a mean-preserving expansion in the variance of Ct +1 ;
I if u 0 (C ) is strictly convex function of C , that is if u 000 (C ) > 0,
then Et [u 0 (Ct +1 )] must rise;
I if Et [u 0 (Ct +1 )] rises, so does u 0 (Ct ), which means that Ct
falls and saving rises.
back
Precautionary savings (III/III)

I You can see this also through a second-order Taylor

approximation.
I Let C̄ = E (C ).
I Take a Taylor approximation around C = C̄ of u 0 (C ):

1
u 0 (C ) ≈ u 0 (C̄ ) + u 00 (C̄ )(C − C̄ ) + u 000 (c̄ )(C − C̄ )2
2
I Take the expected value:
1
E [u 0 (C )] ≈ u 0 (C̄ ) + u 000 (c̄ )Var (C ).
2
I If and only if u 000 > 0, a rise in the variance of consumption,
holding C̄ constant, raises E [u 0 (C )].
back
Algebra for Sharpe-ratio with CRRA (I/II)

I Derive separately numerator and denominator.

γ2 2
I E (M ) = βe −γµ+ 2 σ .
2 2 2 2
I σ (M ) = [ β2 e −2γµ+γ σ (e γ σ − 1)]1/2 .
I Therefore:
σ (M ) p
= e γ2 σ2 − 1.
E (M )
back
Algebra for Sharpe-ratio with CRRA (II/II)

I For the approximation, simply apply a first order Taylor

approximation to:
2 2
e γ σ − 1,
around γ2 σ2 = 0.
I In this case:
2 σ2
eγ − 1 ≈ γ2 σ 2 ,
and p
2 σ2
eγ − 1 ≈ γσ.
back