Introduction to Uncertainty

Jesús Fernández-Villaverde
Duke University

1
 Introduction to Uncertainty

• Modern macro studies stochastic processes of observed variables.

• Two elements:

1. Dynamics.

2. Uncertainty.

• We will introduce some basic concepts by presenting a pure exchange

economy with stochastic endowments.

• In this lecture, we will present the expected discounted utility and we

will use it to assess the welfare cost of the business cycle.
2
 Time

• Discrete time t ∈ {0, 1, 2, ...} .

• Why discrete time?

1. Economic data is discrete.
2. Easier math.

• Comparison with continuous time:

1. Discretize observables.
2. More involved math (stochastic calculus) but often we have ex-
tremely powerful results.

• Calendar versus planning time.

3
 Events

• One event st happens in each period.

• st ∈ S = {1, 2, ..., N} .

• Note:

1. S is a finite set. We will later talk about measure theory.

2. S does not depend on time.

• Event history st = (s0, s1, ..., st) ∈ S × ... × S = S t+1.

4
 Probabilities

• Probability of st is π(st).

³ ´
t
• Conditional probability of st+1 is π st+1| s .

• At this moment, we are not imposing any transition probability among

states across time.

• Our notation allows the particular cases:

³ ´
π st+1 | st = π (st+1)
³ ´
π st+1 | st = π (st+1| st)

5
 Commodity Space

• One good in the economy.

• However, good indexed by event history over infinite time. Hence our
commodity space is slightly more complicated (see chapter 15 in SLP).

• Commodity space: (C, k·k) .

• We pick l∞, i.e., the space of sequences c = (c0, c1, ...) , cn ∈ R that
are bounded in the norm:
kck∞ = sup |ci|
i

6
 Household Preferences

• Preferences admit a representation:

∞ X
X
U (c) = β tπ(st)u(ct(st))
t=0 st∈S t

• This is known as the von Neumann-Morgenstern expected utility func-

tion.

• Remember:

1. Key assumptions: continuity and independence axioms.

2. Linear in probabilities.

3. Cardinal utility: unique only up to an affine transformation.

7
 Facts about Utility Function I: Time Separability

• Total utility c equals the expected discounted sum of period (or in-
stantaneous) utility u(ct(st)).

• The period utility at time t only depends on consumption in period t

and not on consumption in other periods.

• This formulation rules out, among other things, habit persistence.

• However, it is easy to relax: recursive utility functions.

8
 Facts about Utility Function II: Time Discounting

• β < 1 indicates that agents are impatient.

• β is called the (subjective) time discount factor.

1 .
• The subjective time discount rate ρ is defined by β = 1+ρ

• Assumption: constant over time→exponential discounting.

• Alternatives: hyperbolic discounting.

9
 Facts about Utility Function III: Risk Aversion

• Arrow-Pratt Absolute Risk Aversion:

u00 (c)
ARA = − 0
u (c)
Why do we divide by u0 (c)?

• Arrow-Pratt Relative Risk Aversion:

u00 (c)
RRA = − 0 c
u (c)

Interpretation.

10
 Common Utility Functions

• Constant Absolute Risk Aversion (CARA):

−e−ac

• Constant Relative Risk Aversion (CRRA):

c1−γ − 1
for γ =
6 1
1−γ
log c for γ = 1

(you need to take limits and apply L’Hôpital’s rule).

11
 CRRA Utility Functions

• γ plays a dual role controlling risk-aversion and intertemporal substi-

tution.

• Coefficient of Relative Risk-aversion:

u00(c)
− 0 c=γ
u (c)

• Elasticity of Intertemporal Substitution:

u(c2)/u(c1) d (c2/c1) 1
− =
c2/c1 d (u(c2)/u(c1)) γ

• Advantages and disadvantages.

12
 Why CRRA Utility Functions?

• Market price of risk has been roughly constant over the last two cen-
turies.

• This observation suggests that risk aversion should be relatively con-

stant over the wealth levels.

13
 Cost of Business Cycles

• Simple CRRA utility function already answers many questions.

• Lucas (1987), Models of Business Cycles: What is the welfare cost of

business cycles?

• Importance of question:

1. Limits of stabilization policy.

2. Macroeconomic priorities.

14
 A Process for Consumption

• Assume that consumption evolves over time as:

t − 12 σ 2z
ct = μ (1 + λ) e ztc
³ ´
where log zt ∼ N 0, σ 2z .

• The moment generating function of a lognormal distribution implies:

m σz 2 2
m
E (zt ) = e 2

• Then:
µ ¶
− 12 σ 2z
E e zt =1
³ ´ 1 (1−γ)2 σ 2
E zt1−σ = e2 z

15
 A Compensating Differential

• We want to find the value of λ such that:

³ ´1−γ
1−γ t
μc −1
ct −1
E =
1−γ 1−γ

• If these is true period by period and event by event, it should also be

true when we sum up.

• Moreover, the converse is also true: λ is the smallest number that

makes total utilities over time to be equal. Why? Because of the
CRRA and the i.i.d. structure of zt.

• Interpretation: λ is the welfare cost of uncertainty, i.e., by how much

we need to raise consumption in every period and state.
16
 Finding λ

• Dropping irrelevant constants, λ solves:

µ ¶1−γ
− 12 σ 2z
E (1 + λ) e zt =1⇒
− 12 σ 2z (1−γ)
(1 + λ) e Ezt1−γ = 1 ⇒
− 12 σ 2z (1−γ)+ 12 (1−γ)2σ 2z
(1 + λ) e =1⇒
1 2
(1 + λ) e− 2 γσz = 1

• Taking logs: λ ≈ 12 γσ 2z .

• Let us put some numbers here. Using quarterly U.S. data 1947-2006,
σ 2z = (0.033)2 . What is γ?
17
 Size of γ

• Most evidence suggests that γ is low, between 1 and 3. At most 10.

• Types of evidence:
1. Questionnaires.
2. Experiments.
3. Econometric estimates from observed behavior.

• Two powerful arguments from growth theory international compar-

isons. We will revisit these points when we talk about asset pricing.

• Rabin’s (2000): “Risk Aversion and Expected-Utility Theory: A Cali-

bration Theorem.”, Econometrica.
18
 An Estimate of the Cost of the Business Cycle

• Let us take γ = 1 as a benchmark number. Then, we have:

1 1
λ ≈ γσ 2z = ∗ 1 ∗ (0.033)2 = 0.0005
2 2

• Even if we take γ = 10 as an upper bound:

1 1
λ ≈ γσ 2z = ∗ 10 ∗ (0.033)2 = 0.005
2 2

• These are extremely small numbers.

• Later we will see how this finding is intimately linked with the Equity
premium puzzle.

• How could we turn around this result?

19
 Alternatives Routes

• We assumed:

1. Representative agent.

2. Exogenous lognormal consumption.

3. Expected utility.

• How important are each of these three assumptions?

20
 Representative Agent

• Representative agent: fluctuations are at the margin.

• Lucas is very explicit about the possible costs of inequality.

• We will see in the next lecture that with complete markets we will
have perfect risk sharing.

• But the interesting question is the effects of business cycles with in-
complete markets and heterogeneity.

• Krusell and Smith (2002), loss of 0.001 of average consumption, 65%

of households lose when business cycles are removed.
21
 Exogenous Lognormal Consumption

• A combined hypothesis: exogenous consumption+lognormal consump-

tion.

• Exogenous consumption⇒Cho and Cooley (2001), business cycles may

increase welfare: mean versus spread effect.

• Lognormal consumption⇒great depressions? Chaterjee and Corbae

(2005): welfare cost of 0.0187. They calibrate a great depression
every 87 years.

• A nonparametric approach by Álvarez and Jermann (2004) suggests

costs between 0.0008 and 0.0049.
22
 Problems of Expected Utility

• We have representation:
∞ X
X
U (c) = β tπ(st)u(ct(st))
t=0 st∈S t

• Three strong assumptions:

1. Intertemporal elasticity of substitution and risk aversion are deter-

mined by just one parameter.

2. Temporal separability.

3. Expected utility.

• All are problematic and they may affect our calculations.

23
 Recursive Utility

• Espstein-Zin preferences (1989):

∙ ³ ´ ρ ¸ ρ1
ρ α α
Ut = (1 − β) ct + β EtUt+1

separates elasticity of substitution:

1
γ=
1−ρ
from risk-aversion α.

• Applied to evaluate cost of business cycles by Tallirini (2000).

24
 Risk in the Long Run

• Bansal and Yaron (2004): difficult to distinguish a long run component

from a random walk.

• Implications for the equity premium.

• Croce (2006): cost of business cycle.

25
 Temporal Anomalies

• Present-bias. Frederich, Lowenstein, and O’Donoghue (2002), “Time

Discounting and Time Preference: A Critical Review”. Journal of
Economic Literature.

• Explanations:

1. Hyperbolic discounting (Phelps and Pollack, 1968, Laibson, 1996):

∞
X
δβ tu(ct)
t=0

2. Temptation: Gul and Pesendorfer (2003).

26
 Uncertainty Anomalies

1. Framing effects (Kahneman and Tversky).

2. Allais paradox. Three prizes in a lottery: {0, 1, 10}

Problem 1: L1 = (0, 1, 0) versus L2 = (0.01, 0.89, 0.1) .

Problem 2: L3 = (0.89, 0.11, 0) versus L4 = (0.9, 0, 0.1) .

3. Ellsberg paradox.

27
 Ambiguity Aversion

• Knight (1921) risk versus uncertainty.

• Gilboa and Schmidler (1989):

min EQu (c)

Q∈P

• Two possible extensions:

1. Choice over time.

2. General class of ambiguity aversion.

28
 Choice over Time

• Epstein and Schneider (2003):

∞
X
min EQ β tu(ct)
Q∈P
t=0

• Difficult technical assumption⇒rectangularity.

29
 Ambiguity and the Variational Representation of Preferences

• Maccheroni, Marinacci, and Rustichini (2006):

n o
min EQu(c) + φ (Q)
Q∈P

• The function u represents risk attitudes while the index c captures

ambiguities attitudes.

• They extend it to the intertemporal case.

• One particular example:

n o
min EQu(c) + θR (Qk P )
Q∈P

• Hansen and Sargent’s (2006) research program on robust control.

30