EC 502 Lecture 11: Consumption with Infinite

Horizons and Uncertainty

Pascual Restrepo*

1 Infinite Horizon with Certainty

Consider a person choosing an infinite stream of consumption levels Ct , t = 0, 1, 2, .... This
infinite horizon assumption only makes sense if we assume that the consumer themselves is
infinitely lived. Obviously, that’s never true in reality. However, for very young people, or people
who care about leaving money after death through bequests to their children, it may be a rea-
sonable approximation to assume that their preferences are represented by the infinite stream of
discounted utility with geometric discounting and time separable utility:
∞
X
β t U (Ct ).
t=0

Let’s assume that in each period t this agent receives the income level Yt which they know ahead
of time. This is called a perfect foresight assumption, which we will relax below when we discuss
uncertainty and risk. Further assume that the consumer faces the fixed net interest rate r and can
save some amount St+1 each period. We use the time index t + 1 for savings deposited in period
t because the return on these savings is realized in period t + 1. As before, St+1 ≥ 0 implies
savings, with St+1 ≤ 0 implying borrowing. Each period t ≥ 0, the budget constraint

Ct + St+1 = Yt + (1 + r)St

must hold. So the household solves the following optimization problem

∞
X
max β t U (Ct )
{Ct }∞
t=0 t=0

Ct + St+1 = Yt + (1 + r)St .
As in the two-period example you can write this as an optimization problem involving savings
choices alone by substituting for Ct in the utility function using the budget constraint:
∞
X
max β t U (Yt + (1 + r)St − St+1 ).
{St+1 }∞
t=0 t=0

* These notes borrow heavily from notes by Stephen J. Terry.

1
Differentiating this expression with respect to St+1 , the choice of the consumer in period t, yields
the first-order condition for optimality

−β t U 0 (Yt + (1 + r)St − St+1 ) + β t+1 (1 + r)U 0 (Yt+1 + (1 + r)St+1 − St+2 ) = 0

U 0 (Ct ) = β(1 + r)U 0 (Ct+1 ).

So the Euler equation from the two-period case generalizes, and optimal consumption requires
that this Euler equation hold for every period t ≥ 0. We can generalize the present value budget
constraint from before to obtain
∞  t
X 1 P∞  1 t
Ct = t=0 1+r (Yt + (1 + r)St − St+1 ) (use budget constraints)
1+r
t=0
 P∞  1 t
∞  t


 h t=0 1+r Yt
X 1 
1 1
i
Ct = + (1 + r)S0 + S1 + 1+r S2 + (1+r) 2 S3 + ... (decompose into 3 series)
1+r 
t=0  h i
1 1
− S1 + 1+r S2 + (1+r)2 S3 + ...


∞  t
1 P  1 t
(1 + r)S0 + ∞
X
Ct = t=0 1+r Yt (cancel terms)
1+r
t=0

Just as before, the PDV of consumption must equal the PDV of income, but now also adjusting
for the initial savings S0 .1 The consumption Euler equations and the budget constraints implicitly
define optimal consumption levels Ct .

1
1.1 A Special Case β = 1+r
1
Let’s also assume β = 1+r . As in the two-period case, this implies

U 0 (Ct ) = β(1 + r)U 0 (Ct+1 ) = U 0 (Ct+1 ).

Since U 0 (Ct ) = U 0 (Ct+1 ) we must have Ct = Ct+1 .2 In other words, consumption levels are
constant with
C0 = C1 = C2 = ...
There is perfect consumption smoothing.3 The LHS of the present value budget constraint can be
written
∞  ∞ 
t t !
X 1 X 1 1 1+r
Ct = Ct = Ct 1 = Ct .
1+r 1+r 1 − 1+r r
t=0 t=0

Combining the LHS and RHS we have

∞  t
1+r X 1
Ct = (1 + r)S0 + Yt
r 1+r
t=0

1
Technically, we’re assuming that these series are convergent, which is almost always the case for utility
functions and examples which we will consider.
2
Here we’re assuming that U 0 (Ct ) is invertible, which is typically the case for the utility functions we use.
3 1
Note that perfect consumption smoothing does not always occur if β 6= 1+r !

2
 ∞  t
r X 1
Ct = rS0 + Yt
1+r 1+r
t=0

This formula is informative. In particular, consumption Ct depends upon the entire stream of
income Yt for all t ≥ 0. Also, the impact of income changes depends upon the persistence of the
income changes.

• Let’s consider an example where Y0 increases by $1, but no other income levels change,
r
with 5% interest rates and r = 0.05. Consumption increases at all dates by 1+r ≈ $0.048.
∂C0
So the MPC is very low at around ∂Y0 = 5% to this purely transitory shock. Intuitively,
the transitory shock doesn’t have much impact on the PDV of lifetime income, so it doesn’t
change optimal consumption very much at all.

• However, if income increases by $1 for all periods t, then the formulas demonstrates that
 t
r P∞ 1
consumption at all dates increases by 1+r t=0 1+r $1 = $1. The MPC is very high for a
permanent shock at ∂C ∂Y0 = 1. Intuitively, persistent shocks do change the PDV of lifetime
0

income substantially, so they do change optimal consumption a lot.

The key logic is that, absent uncertainty and borrowing constraints, and unless β(1 + r) is very
different from 1, you allocate your income equally across all time periods, regardless of when you
earn it. Hence, consumption is smoother than income. This pattern matches the lower volatility of
consumption seen in the business cycle data.

2 Uncertainty in Two Periods

In reality, households do not know exactly what their income will be in future, so to build a more
realistic model of consumption we must allow for uncertainty or risk. Let’s return to the two-period
example from the last lecture. Instead of assuming that households know their period 2 income Y2
exactly, assume that Y2 is a random variable with

Ȳ2 + ε, with probability 0.5
Y2 =
Ȳ2 − ε, with probability 0.5

where ε > 0. In this example, future income fluctuates around some mean level E(Y2 ) = Ȳ2 known
ahead of time. Although next period income Y2 is uncertain, we assume that the household does
know today’s income Y1 for certain and can choose savings S today at some fixed and known
interest rate r. Note that because Y2 is random, future consumption is also random, with
 H
C2 = Ȳ2 + ε + (1 + r)S, with probability 0.5
C2 =
C2L = Ȳ2 − ε + (1 + r)S, with probability 0.5

The expected value of future consumption is given by E(C2 ) = 0.5C2H + 0.5C2L = Ȳ2 + (1 + r)S.
Under these conditions, the household chooses savings S in period 1 to maximize discounted
expected utility, solving
max U (C1 ) + βEU (C2 )
S

C1 = Y1 − S, C2 = Y2 + (1 + r)S.

3
Rewrite this problem by using the budget constraints and the definition of expected value
 
max U (Y1 − S) + β 0.5U (Ȳ2 + ε + (1 + r)S) + 0.5U (Ȳ2 − ε + (1 + r)S)
S

with the resulting first-order optimality condition for savings S given by

U 0 (C1 ) = (1 + r)β 0.5U 0 (Ȳ2 + ε + (1 + r)S) + 0.5 ∗ U 0 (Ȳ2 − ε + (1 + r)S)

 

U 0 (C1 ) = (1 + r)βEU 0 (C2 ).

So the consumption Euler equation generalizes naturally to the case of uncertainty. Households
optimally smooth discounted expected marginal utility.

• The intuition is similar to before, with the expected marginal cost of savings – marginal
utility today – set equal to the expected marginal benefit of savings – represented by the
discounted expected marginal utility tomorrow adjusted by the return on savings.

• The expected present value lifetime budget constraint also generalizes, as you can verify

E(C2 ) E(Y2 )
C1 + = Y1 +
1+r 1+r

• As before, the consumption Euler equation and the lifetime present value budget constraint
together are enough to allow us to solve for optimal consumption levels.

• The consumption Euler equation using discounted expected marginal utilities

U 0 (Ct ) = β(1 + r)EU 0 (Ct+1 )

holds more generally than in this example as the optimality condition for consumption. In
particular, it holds for any well behaved distribution of income risk, not just the ±ε example
here, as well as for horizons longer than two periods.

2.1 Example: Quadratic Utility & Certainty Equivalence in 2 Periods

Assume that U (C) = aC − 2b C 2 in the 2-period example, where a > 0, b > 0, and β = 1
1+r . In this
very special case, marginal utility is linear with

U 0 (C) = a − bC,

and the Euler equation can be written

a − bC1 = β(1 + r)E(a − bC2 ) (Euler equation)

a − bC1 = a − bEC2 (rearrange)
C1 = EC2 (rearrange)

1
Recall that in the case with certainty from before, β = 1+r implied perfect smoothing of realized
consumption C1 = C2 . With uncertainty and quadratic utility, we instead only obtain perfect

4
smoothing of expected consumption C1 = EC2 . We can define human wealth H as the expected
PDV of income and use the lifetime present value budget constraint to get
EY2
H = Y1 +
1+r

EC2
→ C1 + = H (lifetime expected PDV budget constraint & def. of H)
 1 +r
1
C1 1 + = H (apply C1 = EC2 )
1+r
1+r
C1 = 2+r H (rearrange)

1
This is the same formula for C1 that we had in the case with certainty and β = 1+r , although
reflects expected human wealth H. So we call this very special case certainty equivalence,
since the amount of income risk ε does not change first-period consumption C1 or savings S at all.
How does this case work in practice?
• In period 1 the household chooses savings and consumption C1 to satisfy their Euler equa-
tion and perfectly smooth expected consumption C1 = EC2 .

• In period 2, the income risk ±ε is realized, and we have


C1 + ε, with probability 0.5
C2 =
C1 − ε, with probability 0.5

• Under certainty equivalence the household smooths the expected component of income
changes, i.e. the difference between Ȳ2 and Y1 , but it can not smooth the unexpected or
uncertain component of income changes.

2.2 Example: CRRA Utility & Precautionary Savings in 2 Periods

We saw above in the certainty equivalence case that the amount of income risk ±ε did not change
savings S. But this may not be entirely reasonable in practice, if we think that risk might cause
some individuals to engage in precautionary savings in response to uncertainty about their future
income. To allow for this, we consider the same 2-period example with CRRA utility rather than
1
quadratic utility. We assume β = 1+r and the same risk structure from before. Recall that the
CRRA utility function is given by
C 1−γ
U (C) = , γ>0
1−γ
Note for future reference that
• U 0 (C) = C −γ > 0

• U 00 (C) = −γC −γ−1 < 0

• U 000 (C) = γ(γ + 1)C −γ−2 > 0 → U 0 (C) is convex

• Jensen’s Inequality: If f (X) is convex, then Ef (X) > f (EX) for nondegenerate random
variables X.

5
With these facts in hand, note that the Euler equation can be written
U 0 (C1 ) = β(1 + r)EU 0 (C2 ) (general Euler equation)
 
C1−γ = E C2−γ (substitute for U 0 (C))
C1−γ > (EC2 )−γ (apply Jensen’s inequality)
C1 < EC2 (C −γ is decreasing)

With uncertainty present, and CRRA utility, the certainty equivalence equation from above C1 =
EC2 does not hold! In particular, the presence of risk with ε > 0 in period 2 leads consumers with
these preferences to consume less and save more in period 1. We call this behavior precaution-
ary savings, since it represents households savings as a precaution against the bad outcome −ε
in the next period.
• In general, precautionary savings occurs whenever U 000 (C) > 0, i.e. whenever marginal
utility is convex, due to Jensen’s inequality.
• For the quadratic utility example above, U 000 (C) = 0, so Jensen’s inequality did not apply and
C1 = E(C2 ) did hold with certainty equivalence.
Precautionary savings seems to matter in practice. Because of the possibility, for example, of
losing their job, households may put away some of their income as savings and consume slightly
less today. If risk fluctuates over the business cycle, and is higher during recessions, then precau-
tionary savings may lead people to consume less during recessions.

3 Infinite Horizon with Uncertainty

Let’s combine uncertainty with the infinite horizon framework outlined above. In this case, we
assume that an infinitely lived household maximizes the discounted expected utility
∞
X
max∞ E β t U (Ct ),
{St+1 }t=0
t=0

s.t. Ct + St+1 = Yt + (1 + r)St

where the budget constraint in each period t ≥ 0 reflects consumption Ct , income Yt , savings St+1
made in period t for the next period, and savings St brought forward from the previous period at
constant interest rate r. We allow for randomness or risk in income levels Yt , so that the resulting
behavior of consumption Ct may also be random as in the two-period example.

• As in the 2-period case with uncertainty, differentiating discounted expected utility with re-
spect to St+1 yields the following set of consumption Euler equations which must hold for all
periods t ≥ 0:
U 0 (Ct ) = β(1 + r)Et U 0 (Ct+1 ).
Note that the expectations on the right hand side of the Euler equation have a time subscript
t, reflecting the fact that expectations must be taken after incorporating information available
at the time t when the household is choosing savings St+1 and consumption Ct . In particular,
the household at time t observes the value of their current stock of savings St as well as any
past income realizations Ys , s ≤ t.

6
 • Recall the lifetime present value budget constraint from the infinite horizon with certainty
case, which must also hold here:
∞  t ∞  t
X 1 X 1
Ct = (1 + r)S0 + Yt .
1+r 1+r
t=0 t=0

This says that the entire PDV of lifetime consumption must equal the PDV of lifetime income
taking into account current savings. Using the same derivations as before we can start the
sum on the LHS at any period t, implying the formula
∞  s ∞  s
X 1 X 1
Ct+s = (1 + r)St + Yt+s .
1+r 1+r
s=0 s=0

In other words, starting at any point t, the PDV of future and present consumption must
equal the PDV of future and present income, taking into account current savings.

• Taking expectations of the final expression at time t we obtain the expected present value
budget constraint of the household starting from any time period t:
∞  s ∞  s
X 1 X 1
Et (Ct+s ) = (1 + r)St + Et (Yt+s ).
1+r 1+r
s=0 s=0

∞  s
X 1
Et (Ct+s ) = (1 + r)St + Ht
1+r
s=0

where expected human wealth Ht at time t is given by the PDV of expected income from
P ∞  1 s
time t onwards Ht = s=0 1+r Et (Yt+s ).

4 Permanent Income Hypothesis

With the infinite horizon optimal consumption framework under uncertainty in place, we can in-
troduce a version of the famous Permanent Income Hypothesis (PIH) put forward by Milton
Freidman in the 1950s. Let’s start with the optimal consumption problem under uncertainty with
an infinite horizon. However, let’s work with the special quadratic utility case U (C) = aC − 2b C 2
1
with a > 0, b > 0, and β = 1+r from a previous section. The Euler equation characterizing optimal
consumption is

U 0 (Ct ) = β(1 + r)Et U 0 (Ct+1 ) (Euler equation)

a − bCt = a − bEt Ct+1 (quadratic utility)
Ct = Et Ct+1 (rearrange)

which must hold for all periods t ≥ 0. You can show using a theorem called the Law of Iterated
Expectations that

Ct = Et Ct+1 = Et (Et+1 Ct+2 ) = Et Et+1 ...Et+s−1 Ct+s = Et Ct+s , s≥0

7
This formula implies that the best forecast of future consumption in any period is current con-
sumption. But then substitute Ct = Et Ct+s into the infinite horizon expected present value budget
constraint from time t onwards:
∞  s
X 1
Et (Ct+s ) = (1 + r)St + Ht (formula above)
1+r
s=0
∞  s !
X 1
Ct = (1 + r)St + Ht (apply Et Ct+s = Ct )
1+r
s=0
 
1+r
Ct = (1 + r)St + Ht (geometric sum)
r
r
Ct = rSt + 1+r Ht (rearrange)

So we obtain, with this very special case of quadratic utility, another form of certainty equivalence.
Only the mean level of income reflected in the term Ht matters for consumption.

4.1 Permanent and Transitory Shocks

Now assume that the risk in income takes a special form with Yt = Ȳ + εt in each period t. Ȳ is
the mean level of income called the permanent income of the household, since its value affects
income in all periods. The values εt are random variables which are mean zero and independently
and identically distributed across periods, with Eεt = 0 for all t ≥ 0. We call the random variables
εt transitory shocks to income since they only affect income in one period t. Now note that
Et Yt+s = Ȳ for all s ≥ 0. Then recall from the definition that
P∞  1 s
Ht = s=0 1+r Et (Yt+s ) (definition)
  s
Yt + ∞ 1
P
Ht = s=1 1+r Ȳ (apply Et Yt = Yt , Et Yt+s = Ȳ for t > s)
  s
Ht = Ȳ + εt + ∞ 1
P
s=1 1+r Ȳ (apply Yt = Ȳ + εt )
  s
εt + ∞ 1
P
Ht = s=0 1+r Ȳ (absorb Ȳ into sum)
1+r
Ht = εt + r Ȳ (geometric sum)
r r
rSt + Ht = rSt + 1+r εt + Ȳ (multiply by constant, then add to both sides)
1+r
r
Ct = rSt + 1+r εt + Ȳ (apply formula from above)

This final expression makes it easy to see that consumption responds very differently to transitory
shocks versus changes in permanent income. In particular,
∂Ct r
• ∂εt = 1+r , which is typically small because r is small
∂Ct
• ∂ Ȳ
= 1, which is a large response
Consumption responds very little to transitory shocks, because they don’t change the expected
PDV of lifetime income much. But consumption responds fully to permanent income changes,
since they shift the lifetime expected PDV of income one-for-one.

8
 • This set of different sensitivities to permanent and transitory shocks is exactly what economists
mean when they mention the PIH. It implies that in response to a temporary change in in-
come (e.g. a temporary tax cut meant to stimulate the economy during a recession), con-
sumption will not change very much.
• Certainty equivalence is a special case, and it is required for these results to hold exactly. If
you have a different utility function than quadratic which implies precautionary savings, e.g.
CRRA, or if borrowing constraints exist for households, then these patterns of responsive-
ness will differ from the PIH.

4.2 Time Series Tests of the PIH

Using the quadratic utility formulation with certainty equivalence laid out above, recall that

Ct = Et Ct+1

must hold given optimal consumption behavior for all periods t. But this means that the “surprise”
νt+1 in time t + 1 consumption is given by

νt+1 = Ct+1 − Et Ct+1 = Ct+1 − Ct ,

where Et νt+1 = Et Ct+1 − Et Ct+1 = 0. So then we can write

Ct+1 = Ct + νt+1

∆Ct+1 = νt+1 .
This pattern of time series behavior for Ct , in which consumption is equal to it’s lagged value
plus an unforecastable component, is called a martingale. Changes in consumption are not
predictable. If they were predictable, the household would have had an incentive to “smooth them
away.” A few notes are in order:

• The martingale consumption implication of the PIH depends upon quadratic utility and no
borrowing constraints. If you have an incentive to engage in precautionary savings with
CRRA utility, or if you face borrowing constraints, the equation Ct = Et Ct+1 may not hold
exactly.
• In a paper published in 1978 called “Stochastic Implications of the Life Cycle-Permanent
Income Hypothesis: Theory and Evidence” in the Journal of Political Economy, Bob Hall
tested the martingale theory of consumption – approximately – by running the following
regression:
log Ct+1 = γ0 + γ1 log Ct + γ2 log Yt + ηt
Hall found γˆ1 ≈ 1 and γˆ2 ≈ 0, consistent with the martingale property of consumption and
this formulation of the PIH.
• In a famous paper in 1989, John Campbell and Greg Mankiw published “Consumption, In-
come and Interest Rates: Reinterpreting the Time Series Evidence” in the NBER Macroeco-
nomics Annual. They reformulated the problem, noting that expected income growth should
have no impact on realized consumption growth. In other words, they note that the certainty-
equivalent formulation of the PIH implies that γ1 should be approximately 0 in the regression

∆ log Ct+1 = γ0 + γ1 Et ∆ log Yt+1 + ηt+1 .

9
 Campbell and Mankiw found γ1 ≈ 0.5. Their result suggests that perhaps precautionary
savings, borrowing constraints, or more broadly some departure of realized consumption
from the PIH, may be a feature of consumption behavior in reality.

4.3 Natural Experiments Testing the PIH

One implication of the PIH, noted above, is that one-time windfalls in income should have relatively
little impact on consumption, since they do not change the PDV of total lifetime income very much.
Instead of testing time series properties of aggregate consumption data - like the Hall (1978) or
Campbell and Mankiw (1989) papers - other authors have looked at empirical cases in which there
are natural experiments involving large windfalls of income in micro data.

• Alaska Permanent Fund: A famous paper called “Do Consumers React to Anticipated
Income Shocks? Evidence from the Alaska Permanent Fund” published in the American
Economic Review in 2003 by Chang-Tai Hsieh examined the case of payments to Alaskan
citizens from the Alaska Permanent Fund, which is a distribution of state oil revenues to
each citizen. The amount varies from year to year, $331/person in 1984 to $1,964/person in
2000. The payments are sent out in October, but the amount is announced far in advance (in
March) and widely publicized in the media. Through the lens of the model, you would expect
consumers to start responding to the change as soon as they find out about it, because their
expected PDV of lifetime income changes at that time. Consistent with this hypothesis, and
the basic implications of the PIH model, Chang-Tai Hsieh does NOT find a different pattern
of consumption in October in Alaska relative to the rest of the US.

• 2008 Tax Rebates: A famous paper called “Consumer Spending and the Economic Stimulus
Payments of 2008” published in the American Economic Review in 2014 by Jonathan Parker,
Nicholas Souleles, David S. Johnson, and Robert McClelland examined tax rebates in 2008
in the US. The government distributed $100 billion in tax rebates to 130 million households
as part of a package to attempt to stimulate the US economy during the Great Recession and
financial crisis. Households received between $300 and $1200 depending upon the size and
composition of their family. The IRS could not distribute all these payments simultaneously,
so they sent the checks out sequentially. Parker, et al. compared household spending for
those who received early checks versus later checks - these were randomly determined
differences based upon the social security numbers of people in the households. They
find that about 50-90% of the payments were consumed within 3 months. This is strong
evidence against the PIH, since the payments were not large enough to change lifetime
income by very much at all. So perhaps liquidity constraints, which induce a high response
of consumption to income, were at work during the Great Recession.

10