As noted in Meyer and Meyer (2005), there is a longstanding and continuing tradition of utilizing “introspection” to draw inference on the degree of risk aversion. For example, economists have long argued that there must be an upper bound to utility, in order to avoid a St. Petersburgstyle paradox whereby agents will choose to take on gambles with a large probability of considerable losses, but a tiny probability of enormous gains.
In the context of a CRRA utility function, an upper bound requires the degree of relative risk aversion to be greater than one. It will prove useful to review the mathematics of this result, in a variation of the St. Petersburg paradox game.
Consider the “game” where your wealth (after paying an “entrance fee”,
F) will multiply by a factor
X, for every successive time you toss a head on a fair coin. The first time a tail appears the game stops. If a tail comes up on the first toss, your terminal wealth,
W_{T}, is
W_{0 }– F, where
W_{0} is the initial wealth.
If a tail comes up on the second toss,
your terminal wealth is X(
W_{0 } F); if a tail comes up on the third toss, your terminal wealth is
X^{2}(
W_{0 } F), and so on. The expected utility for playing the game when preferences are CRRA is
(I.1)
Suppose we have a prospective player whose initial wealth is $100,000 and X = 4.1 with a = 0.5. According to expected utility maximization, this individual would play this game even for an entrance fee of $99,999. Notice, however, that our player would have to toss at least 9 consecutive heads in order for terminal wealth to exceed $100,000, an event with a probability of less than twotenths of one percent. The idea that vanishingly small probabilities of enormous gains could have such a large effect on the willingness to take on gambles leads economists to argue that utility must be bounded from above.
Arrow (1971) also recognized that similar problems arise if utility was not also bounded from below.^{1} In this case, vanishingly small probabilities of enormous losses may preclude agents from taking on a gamble. This can be easily illustrated with the “inverse game” to the one above.
In this game, you are paid an entrance fee, F, to play, but every time you toss a head on a fair coin, you have to return a part of your remaining wealth (including the entrance fee). The game stops when the first tail appears. In particular, if a tail comes up on the second toss, you return and your terminal wealth is also (W_{0 }+ F)/X. If a tail comes up on the third toss, you return, and your remaining wealth is (W_{0 }+ F)/X^{2}, and so on.
The expected value for playing this game with CRRA preferences is given by
(I.2)
Now, again suppose that we have a prospective player whose initial wealth is also, $100,000, but whose relative risk aversion is 1.5 and we set X =4.1. Unfortunately, our potential player will refuse the gamble for all finite values of the entrance fee, F. So for example, our player would find the game to be “too risky,” even if after the entrance fee, her wealth was $1 billion (=$100,000^{2}). Again, in this case she has less than a 0.2% chance of being worse off after playing the game, but those very small probabilities of a disaster are sufficient for her to forego the $1 billion.
B. Indifference to “Compound Gambles”
A related problem with CRRA utility is that an individual’s preferences for a gamble do not depend on the number of times the gamble is performed.^{ 2} Another simple example is useful to illustrate this point. Consider the gamble where wealth will increase by the factor X >1, with probability p > ½ and will decrease by 1/X, with probability (1p). Assume that the values of p and X have been selected so that our CRRA subject is indifferent to the gamble. In this case, the values of X and p are related as follows:
(I.3)
Now consider an alternative gamble where the outcomes are based on two binomial trials with the same values for X and p. In this case, the possible outcomes are:
The agents’ expected value from this “compounded” gamble is given by
(I.4)
We are not imagining the second gamble as being sequential. Instead, we are asking our constant relative risk aversion subject to evaluate two onetime gambles with different outcomes and different probabilities. Our CRRA subject, however, will be indifferent to the outcomes based on 1 or 2 binomial trials, as long as the payoffs are compounded. Of course, an exactly similar argument can be made to a gamble based on 3 or more binomial trials where the outcomes are adjusted accordingly.
But herein lies the paradox: with p > 0.5, the probability of a loss gets smaller and smaller as the number of trials increases. As p increases without limit, the chance of a loss approaches zero, but our CRRA subject will value this gamble as equal to the first gamble. In the case of , this is because even though the probability of losses are going to zero, the maximum size of those losses is increasing without limit and the utility consequences of those losses have finite consequences.^{2}
It is worth showing this last result more explicitly. Consider the onetrial gamble with the following properties:
Now consider an alternative gamble with a larger value for p (call this p_{Y}) and smaller value for Y (call this Y_{P}) such that the individual is still indifferent to the gamble:
which implies
Since
, there always exists a
Y_{P}.> 0 that will meet the indifference condition as long as
p_{Y} < 1.
Now consider a (p_{Y} , Y_{P} ) couple that satisfies the indifference condition and p_{Y} is arbitrarily close to (but less than) 1. Also consider a Y that is arbitrarily close to (but less than) Y_{P} . A person with constant relative risk aversion will not accept this gamble even if winning is a nearly “sure thing.” The size of the disaster in the bad state of nature is large enough that the agent would rather not take the gamble even if the probability of a disaster is tiny. This strikes us as implausible as people do seem to take chances such as those in many everyday situations (e.g. driving, flying on an airplane).
The example above illustrates why the Law of Large Numbers doesn’t work for individuals with CRRA preferences (with > 1). Think of the payoff X in our simple gamble as the expected return from a gamble. The Law of Large Numbers implies that as the sample size increases, the sample mean will become arbitrarily close to the population mean. In our simple gamble, this is captured by p_{Y} approaching 1. Thus with CRRA, even though the sample return is approaching the expected return, expected utility is not changing. This is a consequence of unbounded utility, where events with vanishingly small probabilities can have finite effects on expected utility.
C. Diversification over Time.
The last of our related problem with CRRA was stated most clearly by Samuelson (1969, pp.8834):
Third, being still in the prime of life, the business man can “recoup” any present losses in the future. The widow or the retired man nearing life’s end has no such “second or n^{th} chance.”
Fourth, (and apparently related to the last point), since the businessman will be investing for so many periods, “the law of averages will even out for him,” and he can afford to act almost as if he were not subject to diminishing marginal utility.
However, before writing this paper, I had thought that points three and four could be reformulated to give a valid demonstration of businessman’s risk, my thought being that investing for each period is akin to agreeing to take a 1/n^{th }interest in insuring n independent ships.
The present model does not itself (emphasis in the original) introduce extra tolerance for riskiness at early, or any, stage of life.
The “present model” was the discrete time version of the lifetime portfolio selection model where a closedform solution was derived for the CRRA case. When theory (based on CRRA preferences) conflicted with introspection, Samuelson chose theory.
^{3} Samuelson had modeled the instantaneous utility from consumption,
U(
C ), as exhibiting CRRA and then had shown that the indirect utility function for wealth,
V(
W) was also CRRA with the same degree of relative risk aversion.
To illustrate this point, consider the following lifetime portfolio selection problem:
(I.5)
where
is the discount factor,
is the share of assets allocated to the riskyasset, and
and
R are the gross rate of return on the risky and riskfree assets respectively.
only enters the problem through the second term of (I.5). If
V(
W) is CRRA, then the first order condition for
can be written as:
(I.6)
so that the optimal
does not depend on the level of wealth but only on the individual’s relative risk aversion and the distribution of asset returns.
Note that as long as the riskfree rate and the distribution of the risk premium are time invariant, optimal portfolio choice will not depend on time. Therefore, there are no lifecycle or timehorizon effects on optimal portfolio choice. This very important and powerful result is based on CRRA preferences, and it strikes many (including Samuelson pre1969) as implausible.
II. Bounded Utility Where the Degree of Relative Risk Aversion Is Constant Across Individuals.
As was
pointed out in the beginning, the assumption of CRRA preferences is often justified by the empirical fact that people with different levels of wealth on average exhibit similar relative risk aversion. But there is an important distinction between the assumption of CRRA for individual preferences and the observation that average relative risk aversion among groups of individuals appears to be constant across groups with different average levels of wealth. The strong theoretical results discussed in the last section depend on preferences where at a moment in time the individual’s elasticity of marginal utility of wealth is constant for all possible levels of wealth. In this section, we show that it is possible to develop a utility function for wealth where the elasticity of marginal utility is an increasing function of wealth, so that utility is bounded, but the average degree of relative risk aversion across groups of individuals with different average levels of wealth is constant.
Our utility function is based on a variation of “habit formation” with constant absolute risk aversion, where the specification of the habit level of consumption has been chosen for analytical convenience.^{4}
(II.1)
(II.2)
(II.3)
Hence in this CARA utility function, the RRA depends on the absolute risk aversion times the relative wealth, W/Z.^{5}
We have not been able to find a closedform solution for the direct utility function of consumption that gives rise to this indirect utility function for wealth, but we do know that this unknown function is homogeneous of degree zero in C/Z. While this function exhibits CARA for any given level of habit consumption, it can also be consistent with the Friend and Blume (1975) evidence of the relative constancy of relative risk aversion across groups of individuals. The key point is that when we observe individuals with different wealth levels, they will also have different habit levels of consumption. If wealth and habit consumption are roughly proportional across individuals, so too will be the average value of RRA for different wealth categories. In some ways, our solution to the crosssection evidence is similar to Friedman’s solution to the consumption puzzle, where the average value of the Average Propensity to Consume (APC) is roughly constant across groups of the poor and the affluent, but for an individual the marginal propensity to consume is less than the APC. Similarly, the degree of RRA at the initial level of wealth can be constant across groups of individuals while decreasing in wealth for any individual.
With CARA, the utility from wealth is bounded both from above and below and thereby St. Petersburg Paradoxes are eliminated. For example, in our “inverse” gamble from section I.A., where RRA =1.5 and X is equal to 4, a reverse entrance fee of about $450,000 would induce our CARA subject to accept the gamble. This individual would take the gamble, in spite of the fact that there is a 25% probability that terminal wealth will be below initial wealth.
Similarly, the CARA utility function implies that the individual will not be indifferent to the number of trials. We were somewhat surprised that there is not a general statement for whether the 2trial gamble will be preferred to the onetrial gamble when the expected utility of the onetrial gamble is equal to the utility of initial wealth. We have been able to show through numerical simulations that for gambles of more than 2% of wealth and where RRA is greater than 3, the twotrial gamble is preferred to the onetrial gamble.
More importantly, however, the CARA function implies that events with vanishingly small probabilities of occurrence can have no finite effect on expected utility. To see this,
Now consider the pairs of
p_{Y} and
Y_{P} that will provide indifferent gambles:
Note that
Therefore, with CARA, there exist a
p_{Y} sufficiently close to 1, where the individual will prefer the gamble, even when the bad state of nature involves the total loss of wealth. Thus
the Law of Large Numbers will, in fact, work to the advantage of CARA individuals.
Time diversification is a more complicated issue, but CARA preferences do not lead to the strong, but implausible results of CRRA preferences. The lifetime portfolio selection problem and the firstorder condition for (the analogs of (I.5) and (I.6)) can be written as: (II.4)
(II.5)
and in this case the optimal value of will vary with the relative level of wealth.
By taking a secondorder Taylorseries expansion around the expected value of relative wealth in period t+1, it can be shown that (see Appendix) the optimal portfolio allocation, , would also maximize:
(II.6)
Note that
and thus the optimal portfolio allocation depends on the moments of the distribution of the returns on the risky asset and on preferences as measured by relative risk aversion.^{6}
The difference is that relative risk aversion can vary over time with a CARA utility function. In our case, it will vary positively with relative wealth, and it is certainly possible that lifecycle effects will lead to increasing relative wealth and relative risk aversion as the planning horizon shortens with age. Timevarying portfolio allocation in this case is more about the tolerance for risk varying with the horizon, than it is about taking advantage of diversification over time. Time diversification is taking advantage of the Law of Large Numbers and our CARA individual does not avail herself of this opportunity, even though it could work to her advantage.
We can state this result in a slightly different way. If you force an individual to make one portfolio allocation and then stick to that allocation for the rest of his life, the CRRA individual will not find this constraint to be a problem and different individuals with different planning horizons will pick the same allocation if they have the same degree of relative risk aversion. This is not the case for an individual whose preferences are given by the CARA utility function in relative wealth. Individuals with the same relative wealth and the same degree of relative risk aversion but with different planning horizons will not select the same portfolio allocation. In general, the longer the horizon the riskier the allocation selected. This is just a generalization of the “Law of Large Numbers” working for our CARA person.
The reason that individuals with CARA preferences do not timediversify is that they can do even better. Consider the CARA individual who is indifferent to the onetrial gamble but prefers the 2trial gamble. Time diversification means that the “trials” are sequential and because of liquid financial markets one is not forced to accept the second gamble. For example, consider the onetrial gamble for a CARA individual (where we have replaced X with 1+x):^{7}
(II.7)
Let p and x be such that the individual is indifferent between taking the gamble or not:
Our CARA subject will prefer the twotrial gamble when the following condition holds:
As already noted this condition does not always hold for small gambles and low risk aversion. But consider the case where H < 1 does hold. Because these trials occur over time (i.e., the first trial in period 1 and second in period 2) a better gamble is the following: Accept the first period gamble and then if one loses (in which case relative risk aversion will be lower) take the second gamble, which is now preferred to not gambling. If on the first trial one wins, call off the second gamble because the increase in relative risk aversion implies that you prefer certain wealth of W_{0}(1+x) to the expected utility of the second gamble.^{8}
Individuals with CARA preferences choose not to “take advantage of the Law of Large Numbers” because a better strategy is to adapt one’s portfolio allocation to evolving relative risk aversion due to past successes or failures or to lifecycle effects. This timevarying portfolio allocation is due to changing risk tolerance and not to the desire to have winners offset losers over time. Perhaps the folk wisdom that people in the prime of life should take on a riskier portfolio than widows and orphans is due to the combination of lifecycle effects of risk tolerance along with the recognition that most people will solve the portfolio allocation very infrequently. Given the fact that people voluntarily tie their hands, thereby forcing themselves into multitrial gambles, the longer the planning horizon the riskier the portfolio should be.^{9}
III. CARA Utility for Relative Wealth and Asset Pricing Puzzles
The original models of habit formation were developed to explain the equitypremium puzzle [Mehra and Prescott (1985)] and the related riskfree rate puzzle [Weil (1989)]. While our specification of “relative” habit formation is different from that of the original models, it is still capable of explaining the puzzles. The key idea in habit formation models is that trend growth in consumption leads to less of a decrease in marginal utility than does a onetime change in consumption. We can illustrate this idea with a firstorder Taylor series expansion of marginal utility:
(III.1)
The Euler equation for the riskfree asset relates the change in utility from saving more today and investing in the riskfree asset.
(III.2)
Using (III.1), the riskfree rate puzzle (assuming there were, in fact, a completely riskfree asset) can be written as:^{ }
(III.3)
Therefore, if R is less than one, then any growth in consumption is inconsistent with the Euler equation. If R is greater than one, then growth in consumption can be consistent with the Euler equation so long as the rate of growth in consumption and/or the degree of relative risk aversion are not too large. Thus, in general, the Euler equation for the riskfree rate requires a small value for relative risk aversion in order to accommodate trend growth in consumption.
The two related asset pricing puzzles arise because while the degree of relative risk aversion has to be low to accommodate trend growth in consumption, it has to be high to accommodate the volatility in the risk premium. The Euler equation relating to the risk premium, is derived from the condition for the optimal share of the portfolio in the risky asset and can be written in either one of two equivalent ways:
(III.4)
(III.4’)
where (III.4) shows the expected consequences of the excess return on the risky asset for consumption in period t+1, and (III.4’) shows the consequences for wealth. Using (III.1) with (III.4) yields:
(III.5)
Notice that the last term in parentheses will be (relatively) small because it is the product of a growth rate and interest rate. Therefore, trend growth in consumption is not terribly important in this Euler equation. Volatility in consumption growth is more important in helping to determine the covariance term. But if consumption is not very volatile and the average risk premium is high, relative risk aversion will have to be high to accommodate the Euler equation.
In our model of habit formation we cannot directly test the Euler equation related to the riskfree rate, (III.2) because we do not know the functional form of the direct utility function . While we don’t know the exact functional form, we can write the analogous equations to (III.1) and (III.2) and derive similar qualitative results: 1) If R is less than one, no growth in the argument of the utility function is consistent with the Euler equation; 2) If R is greater than one, then the Euler equation can be fulfilled if either growth in the argument of the utility function and/or relative risk aversion is small. The important difference with the habit utility function is that not consumption, but consumption relative to the habit level is the argument of the utility function; and in a steady state there is no growth in this relative consumption. With little or no growth in the argument of utility, the Euler equation for the riskfree rate can be fulfilled for higher degrees of relative risk aversion.
Similarly, we cannot use equation (III.4) to test the equity premium, Euler equation, but in this case we can use the firstorder condition from the indirect utility function. As we showed in (II.5) the firstorder condition for leads to a condition very similar to (III.3), where we can write:
(III.6)
While (III.6) looks similar to (III.3), it is important to note that the degree of relative risk aversion in the indirect utility function is, in general, not the same as in the direct utility function [c.f. Meyer and Meyer (2005)], and the distribution of the growth in wealth is not the same as the distribution for the growth in consumption. But we can, and in the next section do, look at historical data to see whether the data are consistent with the firstorder condition on optimal portfolio choice.
It is important to note that our version of habit formation “solves” the two asset pricing puzzles the same way that the traditional models do. Habit formation allows relative risk aversion in the indirect utility function to be large enough to solve the equity premium puzzle, while lessening (and in our case eliminate for the steady state) the effect of trend growth in consumption on the riskfree puzzle.
IV. Can CARA Preferences over Relative Wealth Explain the Asset Pricing Puzzle?
To investigate whether CARA preferences over relative wealth can explain the asset pricing puzzles, we will follow the methodology laid out in Kocherlakota (1996) and Meyer and Meyer (2005). We see if the Euler conditions implied by the lifetime portfolio allocation problem are satisfied given data and reasonable parameter values.
Before turning to CARA preferences, we first show how the equity premium and riskfree rate puzzles emerge under CRRA utility function. Consider the following 3 statistics (where is the constant degree of relative risk aversion):^{ 10}
(IV.1)
(IV.2)
(IV.3)
For a particular value of , is the statistics that corresponds to the Euler condition related to the riskfree rate, equation (III.3). If that Euler condition holds this statistic has an expected value of zero. If the sample mean of the statistics is statistically significantly different from zero, however, we can reject the hypothesis that this Euler condition is fulfilled for that particular value for the degree of relative risk aversion.
Similarly, is the statistic that corresponds to the Euler equation relating the risk premium and the direct utility function U(C), i.e. (III.4), and is the statistic that corresponds to the Euler equation relating the risk premium and the indirect utility function V(W), i.e. (III.4’). ^{11} The related asset pricing puzzles are the values of A that are needed to avoid rejecting the riskfree Euler equation (viz., low values) lead to the rejection of the riskpremium puzzles (viz., high values) and vice versa.
The variable C is matched to Real, per Capita Consumption Expenditures data from the NIPA tables of the Bureau of Economic Analysis (BEA). W, per capita wealth, is matched to the (beginningofperiod) net financial assets of households (i.e. Net worth – Tangible assets) deflated by the Personal Consumption Expenditure (PCE) deflator. The household wealth data were obtained from the Sectoral Balance Sheets provided in the Flow of Funds (FOF) Accounts of the Federal Reserve Board and the deflator is from NIPA of the BEA. The real return on risky asset, , was matched to the real rate of return of the S&P 500 and the real return on the riskfree asset, R, was matched to the real rate of return of 3month U.S. Treasurybills (the inflation numbers used were derived from the PCE deflator, however the resultant series are very similar to those derived from the GDP deflator). Both series were obtained from Ibbotson (2006). The data used was for the years 19552005. Data are available for earlier postwar years, but the Fed’s practice of pegging the Tbill rate in the preaccord period led us to discard these data [c.f. McGrattan and Prescott (2003)]. Table 1 reports the summary statistics of the data used.
Table 1: Summary Statistics of the Data






Mean

0.0122

0.0684

0.0240

0.0213






Variance

0.0005

0.0291

0.0003

0.0046






Covariance

0.0005





0.0007

0.0291




0.0001

0.0001

0.0003



0.0005

0.0006

0.0006

0.0046

The statistics in Table 1 suggest that CRRA preferences will have a hard time fulfilling the two Euler equations: 1. There is significant average growth in consumption and wealth per capita. 2. The average risk premium is high, but the covariance of the risk premium with growth in consumption per capita or wealth per capita is relatively low. The first fact implies that a high degree of risk aversion will make it more difficult to fulfill the riskfree Euler equation and the second that a low degree of risk aversion will have trouble fulfilling the riskpremium Euler equations.
We test these results statistically, using the data described above. Table 2 presents the sample mean and the tstatistics of the 3 test statistics for different values of the relative risk aversion parameter.
^{12}
Table 2: Evaluating Euler Equations (CRRA preferences)





mean

tstat

mean

tstat

mean

tstat

1

0.02119

6.25683

0.06670

2.77791

0.06653

2.67405

1.5

0.03252

7.94274

0.06588

2.76397

0.06569

2.59931

2

0.04364

8.80311

0.06508

2.74973

0.06491

2.51906

2.5

0.05456

9.25390

0.06430

2.73518

0.06417

2.43397

3

0.06530

9.50445

0.06353

2.72034

0.06348

2.34482

3.5

0.07584

9.65225

0.06277

2.70521

0.06284

2.25243

4

0.08620

9.74405

0.06203

2.68981

0.06224

2.15766

4.5

0.09638

9.80355

0.06130

2.67414

0.06168

2.06139

5

0.10638

9.84346

0.06059

2.65821

0.06116

1.96447

5.5

0.11620

9.87093

0.05989

2.64203

0.06067

1.86768

6

0.12585

9.89019

0.05921

2.62561

0.06023

1.77177

6.5

0.13534

9.90384

0.05853

2.60896

0.05981

1.67738

7

0.14465

9.91351

0.05787

2.59208

0.05944

1.58510

7.5

0.15380

9.92025

0.05723

2.57499

0.05909

1.49538

8

0.16279

9.92477

0.05659

2.55770

0.05878

1.40863

8.5

0.17163

9.92757

0.05597

2.54022

0.05850

1.32514

9

0.18031

9.92896

0.05535

2.52255

0.05825

1.24514

9.5

0.18883

9.92917

0.05475

2.50472

0.05803

1.16878

10

0.19721

9.92838

0.05416

2.48671

0.05785

1.09616

The results above confirm the intuition reached in Section III. The first column shows that there is no value of risk aversion that makes the data consistent with the riskfree Euler equation. It is worth noting that not only is the average growth in consumption during this period high, but it is uniformly positive (in 46 out of the 51 years), thus the first term in (IV.1) is uniformly below 1 (in 45 out of the 51 years) even when the degree of relative risk aversion is equal to 1.
The equitypremium Euler equation, however, can be reconciled with the data using higher values of relative risk aversion (in particular greater than 3).^{13} Note the results are qualitatively invariant to the choice of the Euler equation specified in terms of the direct utility or indirect utility functions. In each case, the tstatistic falls with higher degrees of relative risk aversion. The tstatistics are uniformly lower for the indirect utility function, which probably reflects the greater volatility of wealth (see Table 1).
We now repeat the above exercise using our version of habit formation with CARA preferences over relative wealth. The Euler condition for the risk premium is given by equation (II.5) and can be tested by testing whether the following statistic has a sample mean different from 0:
(IV.4)
We first calculate our summary statistics for the data using the growth of consumption and wealth relative to the habit level of consumption.^{14}
Table 3: Summary Statistics of Data using Relative Variables






Mean

0.01219

0.06838

0.00066

0.00190






Variance

0.00046

0.02912

0.00046

0.00472






Covariance

0.00046





0.00074

0.02912




0.00010

0.00091

0.00046



0.00050

0.00139

0.00081

0.00472

The most important thing to note in Table 3 is that there is no positive trend growth in relative wealth or relative consumption. So the major problem of fulfilling the riskfree puzzle is gone.
We again calculate the mean and tstatistic of the statistic given in (IV.4) for different values of
.
^{15} We pick the value of
A, so that the degree of relative risk version at the sample mean of
W/Z varies from 1 to 10. Note from (IV.1) that only the value of
A/ matters. The results are summarized below in Table 4:

Table 4




mean

tstat

1

0.06696

2.64036

1.5

0.06626

2.54786

2

0.06555

2.44935

2.5

0.06482

2.34555

3

0.06406

2.23729

3.5

0.06328

2.12544

4

0.06245

2.01091

4.5

0.06157

1.89459

5

0.06063

1.77738

5.5

0.05962

1.66011

6

0.05851

1.54357

6.5

0.05731

1.42846

7

0.05599

1.31540

7.5

0.05453

1.20494

8

0.05293

1.09751

8.5

0.05116

0.99348

9

0.04919

0.89313

9.5

0.04701

0.79665

10

0.04459

0.70419

As shown above, the equity premium puzzle again can be solved for relative risk aversion greater or equal to 3 using CARA preferences over relative wealth. Because we cannot find the functional form for the direct utility function, we cannot test formally the riskfree puzzle. But we believe that the data on relative consumption growth in Table 3 are suggestive that our preferences are consistent with the riskfree puzzle. In a steady state, relative consumption would have no trend growth. In our sample, average growth in relative consumption is negative (see Table 3). With this low relative growth there are many periods where relative consumption is falling as well as rising. For example, in 25 of the 51 years in our sample, relative consumption growth is negative. Therefore, while we don’t know the functional form of marginal utility from relative consumption, it less likely that there is a trend in marginal utility that is causing the riskfree premium.
V. Conclusions and Further Thoughts.
We have shown that the assumption of constant relative risk aversion and unbounded utility leads to problems, paradoxes and strong but counterintuitive implications about portfolio choice. All of these problems can be resolved by assuming CARA preferences over relative wealth. In addition, because the degree of relative risk aversion depends on relative wealth, our model is not inconsistent with the evidence of Friend and Blume (1975).
But do these preferences meet the “introspection test?” For further insights we can appeal to Daniel Bernoulli himself, as quoted by the psychologist Daniel Gilbert (2006):
Although a poor man generally obtains more utility than does a rich man from an equal gain, it is nevertheless conceivable, for example, that a rich prisoner who possesses two thousand ducats but needs two thousand ducats more to repurchase his freedom, will place a higher value on a gain of two thousand ducats than another man who has less money than he. Though innumerable examples of this kind may be constructed, they represent exceedingly rare exceptions.
Our model doesn’t directly involve imprisonment, but we can easily construct an example where the rich man gains more utility from an extra 2000 ducats than does a poorer man. Consider a rich man who has suffered some recent losses (perhaps he was just imprisoned), so that his relative wealth is low. The gain of 2000 ducats will raise his relative wealth by less than that of a poor man who has a lower level of habit consumption. But if the poor man started with relative wealth sufficiently greater than the rich man, the poor man will gain less of an increase in utility.^{16} We therefore share Gilbert’s conclusion (although for somewhat different reasons) that,”…the ‘innumerable exceptions’ that Bernoulli swept under the rug are not exceedingly rare.
Appendix
A. The General Form of Bounded Utility in Relative Wealth
In the text we have used the simplest possible specification of bounded utility in relative wealth. A more general specification of utility is the following
Since relative wealth must be nonnegative, the negative exponential guarantees boundedness. In case (A):
In case (B):
In
both cases, relative risk aversion depends on relative wealth, and therefore on average can be constant among groups of people with different average levels of wealth. The simplest example of case (A) is the preferences assumed in the text where
. We use these preferences in the text because it keeps the algebra simple.
Some, however, might be concerned that for a given value of Z, case (A) implies increasing relative risk aversion in W. This implication is not necessary in the more general specification of preferences. For example, consider the case (B) example where . In this case, for a constant Z relative risk aversion is , which for a given value of Z is decreasing in W
B. Compound Gambles with CARA preferences.
Consider the onetrial gamble for a CARA individual (where we have replaced X with (1+x)):
Let
p and
x be such that the individual is indifferent between taking the gamble or not:
The expected utility of the two trial gamble is given by:
Our CARA subject will prefer the twotrial gamble when the following condition holds:
This condition does not hold for in general (e.g., it is violated for p=0.50745, x =.01, AW/Z=4), but it generally holds for larger gambles (e.g., it holds for p=0.51485, x =.02, AW/Z=4). Now consider the case where H < 1, but the gambles are sequential and the individual turns down the second gamble if the good state of nature occurs in the first period.
The alternative gamble is preferred to the twotrial gamble if the expected utility of the second gamble after a good state of nature on the first trial is less than the utility of the wealth after the first trial (i.e., V(W_{0}(1+x))):
C. The Homogeneity of
U(C, Z) when
Consider two individuals with different levels of absolute wealth but the same levels of expected relative wealth at the start of the terminal period.
Now consider the firstorder condition at time T1:
Because this last relationship must hold true for all
D. Optimal Portfolio Choice with CAR preferences Over Relative Wealth
The secondorder Taylor series approximation of aroundcan be written as: