International Finance 12:2, 2009: pp. 123150 DOI: 10.1111/j.1468-2362.2009.01240.

Capital Asset Pricing Model and the Risk Appetite Index: Theoretical Differences, Empirical Similarities and Implementation Problems
Marcello Pericoli and Massimo Sbracia
Bank of Italy.

Abstract
We perform a thorough analysis of the Risk Appetite Index (RAI), a measure of changes in risk aversion proposed by Kumar and Persaud (2002). Building on Misinas study (2003), we rst argue that the theoretical assumptions granting that the RAI correctly distinguishes between changes in risk and changes in risk aversion are very restrictive. Then, by comparing the RAI with a measure of risk aversion obtained

We thank Lieven De Moor, Ming Liu, Miroslav Misina, Alessandro Secchi, Marco Taboga

and seminar participants at the 2004 French Finance Association Meeting, the 2005 Global Finance Association and the 2005 European Financial Management Association for helpful comments. We also beneted from the illuminating suggestions of two anonymous referees and the editor. Giovanna Poggi provided valuable research assistance. Part of the work was carried out while Sbracia was visiting the Department of Economics of the New York University and was completed while he was visiting the Einaudi Institute for Economics and Finance; kind hospitality from both institutions is gratefully acknowledged. The views expressed in this paper are those of the authors and do not necessarily reect those of the Bank of Italy.
r 2009 Blackwell Publishing Ltd. 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA

124

Marcello Pericoli and Massimo Sbracia

from the Capital Asset Pricing Model (CAPM), we nd that the estimates are surprisingly similar. We prove that if the variance of returns is sufciently smaller than the variance of asset riskiness, then RAI and CAPM provide essentially the same information about risk aversion. We also show, however, that RAI and CAPM suffer from exactly the same implementation problems the main one being the difculty in measuring ex-ante returns. At high and medium frequencies, the standard method of measuring ex-ante with ex-post returns may generate negative risk aversion and other inconsistencies. Hence, future research is needed to address this problem.

I. Introduction
In the last few years, nancial institutions and market practitioners have often explained the behaviour of international asset prices on the basis of changes in risk aversion. For instance, according to the International Monetary Fund (IMF), the decline in equity prices recorded between March 2002 and mid-March 2003 followed a sharp increase in investors risk aversion connected with the geopolitical tensions that culminated in the war in Iraq (IMF 2003a). By the same token, both the IMF and the Bank for International Settlements (BIS) explained the subsequent recovery in equity prices with a decline in risk aversion (IMF 2003b; BIS 2004). More recently, following the subprime crisis, many accounts of the decline in stock prices pointed to the role of increased risk aversion. These explanations have been supported empirically with a variety of indicators, usually created by nancial analysts.1 The favourable attitude of nancial analysts towards measuring risk aversion contrasts with the general scepticism that prevails in academic research. For instance, the classical book on the economics of uncertainty by Laffont (1993) has an entire chapter on Measuring Risk Aversion and Risk, without a single reference to empirical works.2 In his seminal contribution to the theory of risk bearing, Arrow (1970) infers the value of risk-aversion parameters on the basis of the properties of von Neumann and Morgenstern
1

IMF (2002, Box 3.1), Gai and Vause (2004) and Illing and Meyer (2005) provide valuable surveys of several atheoretic and model-based measures of risk aversion. In fact, Laffont concludes: It is of course difcult to obtain sufcient information about an agents preferences, to know whether his absolute risk aversion increases or decreases (since this requires information about the third derivative of his utility function) (Laffont 1993, p. 24, cited in Hartog et al. 2002).

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

125

utility functions, without attempting any estimate.3 This scepticism derives not only from the unobservability of individuals preferences but also from the observational equivalence between changes in risk and in risk aversion. This equivalence is due to the fact that an increase (decrease) in either of them causes asset prices to decline (rise) and risk premia to increase (decrease). Kumar and Persaud (2002) have recently made an interesting attempt to break this observational equivalence by exploiting a special feature of asset pricing models. These authors argue that standard pricing models imply that changes in risk aversion modify the rank of asset returns relative to the rank of their riskiness, while changes in asset riskiness do not affect the relative ranks. Accordingly, they build an indicator for changes in investors risk aversion, called the Risk Appetite Index (RAI), which is given by Spearmans rank correlation between expected excess returns and asset riskiness in a cross-section of assets. The RAI has been used thereafter not only in several reports of international nancial institutions but also in papers published in academic journals.4 In this work, we examine the RAI both theoretically and empirically. In the theoretical part, we rene a previous analysis of Misina (2003), who states the two main conditions granting that the RAI distinguishes between risk and risk aversion (Section II). Next, building on Kumar and Persaud (2002) and Misina (2003), we examine the RAI in the context of the Capital Asset Pricing Model (CAPM) (Section III). We focus on the CAPM because it is the workhorse of asset pricing theory. Its main prediction, that expected returns on assets are proportional to their covariance with the market risk, is shared with every other pricing model that has been taken to the data. We argue that the two main conditions granting that the RAI distinguishes between changes in asset riskiness and changes in investors risk aversion are not fullled in the general model; assumptions under which these conditions hold can be found, but are very restrictive. These assumptions require either that investors hold portfolios with equally weighted assets, or that asset returns are uncorrelated, the shocks affecting asset riskiness are idiosyncratic and the number of assets is sufciently large. Although these assumptions are restrictive in theory, we have to verify empirically the extent to which deviating from them biases the measurement of risk aversion provided by the RAI. We do so by comparing the RAI with a
3

Arrow (1970) argues that the relative risk aversion should be approximately equal to one, implying that preferences are represented by a logarithm function, as rst maintained by Bernoulli (1738). For a sample of recent applications, see Baek et al. (2005), Baek (2006), Bandopadhyaya and Jones (2006), and Misina (2008).
r 2009 Blackwell Publishing Ltd

126

Marcello Pericoli and Massimo Sbracia

measure of risk aversion derived from the estimation of the CAPM a model that does not require those assumptions. Using monthly data for the stocks included in the Dow Jones Euro Stoxx and in the Standard & Poors 500 from January 1973 to December 2008, we show that the two estimates are surprisingly similar. By focusing on the statistical properties of the RAI, we are able to explain this result, proving that if the variance of returns is sufciently smaller than the variance of asset riskiness (a condition that is always met in our case studies), then the RAI and the CAPM-based indicator provide essentially the same information about risk aversion (Section IV). In Section V, we discuss the implementation problems that affect estimates of risk aversion from several indices, including the RAI and the CAPM-based indicator. In particular, the main limitations stem from the difculties in measuring expected returns correctly a problem that is especially evident in periods of declining asset prices. In this case, at high and medium frequencies, the standard method of measuring expected with ex-post returns may generate negative risk aversion and other inconsistencies. Only by addressing this problem can the RAI and the CAPM be used to measure changes in risk aversion. Section VI condenses the main conclusions.

II. The Rank Effect and the Key Conditions

The RAI is dened as the rank correlation (measured by Spearmans coefcient) between expected excess returns and asset riskiness in a crosssection of assets. Thus, the rst item needed in order to build it is the expected excess return (excess return hereafter) on each asset i, denoted by ex Rex i , i 5 1, . . . , n. The excess return Ri is the difference between the expected return on the risky asset i, E(Ri), and the return on the risk-free asset, Rf : 5 Rex i ERi Rf . The expected return on asset i, in turn, is dened as the expected price plus the expected dividend, a sum denoted by Xi, scaled by its current price Pi, namely ERi Xi =Pi .6 In asset pricing models, a modication in a structural parameter produces a change in current prices; this, in turn, causes changes in expected and excess returns. Therefore, excess
5

Hereafter, E denotes the expectation operator, Var the variance operator, Cov the covariance operator, and Corr the correlation operator. All these operators refer to unconditional moments. We also denote by sign the operator that takes the value 1 if its arguments share the same sign, and 1 otherwise. Here, we have adopted the standard denition of expected returns (see Cochrane 2001, Chapter 1). In an early application, Kumar and Persaud (2001) had dened Xi as the long-run price of asset i.

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

127

returns adjust to a new equilibrium through changes with the opposite sign in current asset prices Pi. The second item needed to build the RAI is the riskiness of each asset i, which we denote by li. In the following sections, li is to be dened precisely within the CAPM. This denition will be critical in order to assess the properties of the pricing model and, in turn, the appropriateness of the RAI as a measure of risk aversion. As pointed out by Misina (2003), the RAI stems from an important property. We say that a change in a parameter of the pricing model that affects excess returns yields a rank effect if the following condition holds: h i ex either sign li lj DRex i DRj 1; 8i 6 j; h i 1 ex or sign li lj DRex D R 1 ; 8 i 6 j : i j In other words, denition (1) states that the rank effect is obtained when a change in a parameter of the pricing model causes changes in excess returns that are monotone (either increasing or decreasing) in asset riskiness along the cross-section of assets.7 This denition leaves indeterminate both the sign of the change in expected returns (which can be positive or negative) and the sign of the monotonicity relationship between the change in excess returns and assets riskiness (increasing or decreasing). These indeterminacies give rise to four possible cases. We argue below that in all possible cases, the occurrence of a rank effect modies the rank correlation between excess returns and risks in a cross-section of assets. The literature provides several measures of rank correlation (see Stuart and Ord 1991, Chapter 26, for a brief overview). Following Kumar and Persaud (2001, 2002), in this paper, we use Spearmans measure (denoted with rs), which takes values in the interval [ 1, 1]; specically, rs 5 1 (rs 5 1) when the rank of the values taken by sample units according to one variable is exactly the same as (inverse of ) the rank according to the other variable. To understand why the rank effect modies the rank correlation between Rex and l, suppose that, before the rank effect shows up, the correlation is lower than 1. Then, there are at least two assets, say i and j, such that li4lj ex and Rex i < Rj . Hence, consider a change in a parameter that causes a rank effect and, for now, assume that this effect raises excess returns (that is, both ex DRex i and DRj are positive). Suppose also that condition (1) holds because
7

The rank effect as dened by condition (1) generalizes the one stated by Misina (2003), who considers only changes in excess returns that are increasing in the riskiness of assets. Our denition is more relevant to the analysis of the RAI, because, as we discuss in the rest of this section, we can show that the rank correlation between excess returns and risks is also modied by changes in excess returns that are decreasing in the riskiness of assets.
r 2009 Blackwell Publishing Ltd

128

Marcello Pericoli and Massimo Sbracia

we observe monotonically increasing changes in excess returns, namely, ex DRex i > DRj > 0. In this case, if the increase in the excess return on asset i is sufciently larger than the one on asset j, it can reverse the relationship ex between excess returns into Rex i > Rj , thereby strengthening the rank correlation between excess returns and risks. Analogously, suppose that ex the rank effect reduces excess returns (that is, DRex i and DRj are both ex ex negative) and that condition (1) holds with DRj < DRi < 0; then, if the decrease in Rex j is sufciently larger (in absolute terms) than the decrease in , then the relationship between the returns on the two assets can be Rex i ex reversed into Rex i > Rj , strengthening the rank correlation between excess returns and risks. By the same token, assume that the rank correlation between excess returns and risks is higher than 1, so that there exist at least two assets, say ex assets i and j, such that li > lj and Rex i > Rj . Assume also that the change in the parameter causes a rank effect with decreasing and positive excess ex ex ex returns (that is, li > lj and DRex i < DRj with DRi ; DRj > 0). Then, the resulting rank correlation may weaken, as the change in excess returns can ex ex ex turn the relationship between Rex i and Rj into Ri < Rj . The same result is ex ex ex ex obtained if DRi < DRj and DRi and DRj are both negative. Thus, the rank effect as dened by condition (1) tends to modify (either strengthen or weaken) the rank correlation between excess returns and risks. Now suppose that in a pricing model condition (1) is fullled only by changes in the risk-aversion parameter, while changes in risk do not full it. Then, we could exploit this property in order to discriminate between changes in risk and changes in risk aversion. Specically, changes in asset prices that turn out to modify the rank correlation between excess returns and risks are due to changes in risk aversion; moreover, if changes in asset prices do not modify the rank correlation between excess returns and risks, then they can be attributed to changes in risks.8 Following Misina (2003), we can explicitly state the two conditions under which the RAI correctly discriminates between changes in risk and changes in risk aversion. One species that changes in risk aversion yield a rank effect, namely:

In order to identify correctly changes in risk aversion with changes in the rank correlation, it is also important that changes in the risk aversion of opposite signs have opposite effects on the rank correlation. If this condition did not hold, in fact, the rank correlation, as measured by Spearmans coefcient, would eventually go to one of its extrema, 1 or 1. On the contrary, when this condition holds we can identify an increase in risk aversion with, for instance, an increase in the rank correlation and vice versa. We thank Alessandro Secchi for pointing out the role of this further condition.

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

129

Condition 1. A change in investors risk aversion yields a rank effect on excess returns across different assets. Of course, this condition cannot be veried empirically, because risk aversion is an unobservable parameter. Moreover, we cannot use its implications on the rank correlation between excess returns and risks to detect changes in risk aversion because, in principle, a rank effect could show up for reasons other than changes in risk aversion. Therefore, we can introduce another condition that addresses both issues. Specically, we can assume that only changes in risk aversion yield a rank effect, or, equivalently, if we assume that risk and risk aversion are the sole parameters of the pricing model that can change over time, then the second condition states that changes in risk do not yield a rank effect: Condition 2. A change in the riskiness of assets does not yield a rank effect on excess returns across different assets. When both conditions hold, the rank effect can be used to break the observational equivalence between risk and risk aversion. Then, one can use the rank correlation between excess returns and risks to detect changes in investors risk aversion. In the following section, we examine whether Conditions 1 and 2 are fullled in the context of a standard asset pricing model such as the CAPM.

III. Theory: The RAI and the CAPM

Kumar and Persaud (2002) support the RAI as a measure of changes in risk aversion by considering the effect of changes in risk and in risk aversion in the CAPM. However, they consider only the effects on the excess returns on the market portfolio (the optimal portfolio of the representative investor) and neglect the effects on the excess returns on each risky asset. These effects, instead, are critical to verify whether Conditions 1 and 2 hold and to establish whether the RAI can effectively measure changes in risk aversion. The analysis performed by Kumar and Persaud (2002) begins with a wellknown relationship between the excess return and the variance of the market portfolio. For the sake of simplicity, suppose that there are only risky assets in the market. If investors prefer frontier portfolios dened as the portfolios with the minimum variance in the class of the portfolios with the same expected rate of return then the following relationship holds (see Huang and Litzenberger 1988, Chapter 3):
2 s2 m a ERm bERm c;

2
r 2009 Blackwell Publishing Ltd

130

Marcello Pericoli and Massimo Sbracia

where Rm is the stochastic return on the market portfolio, s2 m is its variance, and a, b and c are constants that depend on the expected returns on assets and their variancecovariance matrix. Equation (2) denes the portfolio frontier that is, the locus of all frontier portfolios which is a parabola in the s2 mERm space (the RiskReturn space). The slope of curve (2) is the risk aversion of the representative investor (see Cochrane 2001, Chapter 5, or Kumar and Persaud 2002). Hence, changes in risk aversion determine a shift of the optimal portfolio that modies both the expected return and the variance of the market portfolio, as illustrated in Figure 1. Kumar and Persaud (2002) then consider an alternative scenario in which a simultaneous change in the riskiness of all assets occurs. However, they consider only a very specic change: that which gives rise to a change in a single parameter of equation (2), which is the parameter c of the parabola. When c changes, say it goes from c to c 0 4c, it modies only the riskiness of the market portfolio, without changing its expected return. Figure 2 illustrates this effect. By comparing the different implications of these two scenarios, the authors conclude that changes in risk aversion modify the rank correlation between expected returns and risks, while a simultaneous increase in the riskiness of all assets does not affect it. However, by focusing only on the implications for the market portfolio, Kumar and Persaud neglect the implications of parameter changes for the excess returns on each risky asset, which are crucial to support the RAI as a measure of changes in risk aversion. For instance, in the case represented in Figure 1, a change in the return of the market portfolio following a change in

Return

Risk Appetite 1

Risk Appetite 2

Market Portfolio 2 Market Portfolio 1

Risk

Figure 1: Effect of a change in risk aversion

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

Return

131

Market Portfolio 1 Market Portfolio 2

Risk

Figure 2: Effect of a simultaneous change in the riskiness of all assets

risk aversion does not necessarily imply that excess returns on individual assets change monotonically in their riskiness, as Condition 1 requires. Similarly, the case represented in Figure 2 does not exclude that a change in excess returns on individual assets following a change in their riskiness yields a rank effect, as necessary in order to full Condition 2.9 Thus, even within the very specic change in riskiness considered by Kumar and Persaud (2002), the validity of the RAI remains dubious. Therefore, we now turn to examine the implications of changes in risk and risk aversion on the whole set of assets in order to verify whether Conditions 1 and 2 hold.

A. A Simple CAPM
Following Cochrane (2001, Chapter 9) and Misina (2003), and in order to set the notation and provide a simple example, in this section we consider a standard CAPM with multivariate normal returns and identical investors with Constant Absolute Risk Aversion (CARA) preferences. These assumptions are by no means necessary for our conclusions about the RAI. In fact, Appendix A shows that if we focus on an appropriate measure of risk aversion, then a more general CAPM with heterogeneous and risk-averse agents would yield the same conclusions as those obtained with the simpler CAPM analysed here.
9

By the same token, Pericoli and Sbracia (2004) show that the RAI cannot distinguish between changes in risk and changes in risk aversion in a different (non-standard) pricing model proposed in Kumar and Persaud (2001).
r 2009 Blackwell Publishing Ltd

132

Marcello Pericoli and Massimo Sbracia

Consider a representative consumer with preferences given by the CARA utility: u(C) 5 e gC, where g is the ArrowPratt coefcient of absolute risk aversion. This agent has initial wealth W, which can be split between a riskfree asset-paying Rf and a set of n risky assets paying a stochastic return R 5 (R1, . . . , Rn). Let a 5 (a1, . . . , an) denote the amount of wealth invested in each risky asset i with aiAR, 8i 5 1, . . . , n, and let af AR be the amount invested in the risk-free asset. The budget constraint then implies W af
n X i1

ai ;

P 10 while consumption is given by C af Rf n i1 ai Ri . We also assume that asset returns are multivariate normally distributed with mean ER ER1 ; . . . ; ERn and variancecovariance matrix S: R $ N ER; S. Multivariate normality implies that consumption is nor0 2 0 mally distributed: C $ N mC ; s2 C , with mC 5 af Rf 1a E(R) and sC a Sa. Hence, we can write EuC EegC egmc 2 sC :
g2
2

The consumer maximizes his expected utility (4) under the budget constraint (3). The rst-order conditions for this problem yield the following solution:  S1 a ER 1Rf ; g 5

 is the vector whose elements a i are the optimal amounts of wealth where a invested in each risky asset i, and where 1 is a vector of ones. Of course, the budget constraint (3) implies that the optimal amount of wealth invested in f W 10  the risk-free asset is a a. It is worth noticing that: (i) the strict concavity of the CARA utility function implies that the solution (5) is i is a solution of the problem for given unique; and (ii) each parameter a parameter values Rf , g and S, and for given expected returns E(R). 0 R, where the  f Rf a Thus, the total return on the investors portfolio is a latter addendum is the return on the risky portfolio, which we denote by Rm. Note that the assumption of CARA preferences implies that the amount invested in each risky asset is independent of wealth. Hence, if investors were heterogeneous in their level of wealth, they would buy the same amounts of risky assets and different amounts of the risk-free asset, the latter depending on investors level of wealth.
10

We are implicitly assuming a two-period framework where agents invest in the rst period and, in the second period, returns are distributed and consumption occurs.

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

133

In order to obtain the standard formulation of the CAPM, note that  0 R Sa . Denote by Rex the vector of the excess CovR; Rm CovR; a returns on each asset, that is Rex ER 1Rf . Then, rearranging equation (5), we obtain Rex g CovR; Rm : 6

B. Changes in Risk and Risk Aversion Yield a Rank Effect

We now verify whether Conditions 1 and 2 are fullled; that is, whether changes in risk aversion yield a rank effect (Condition 1), while changes in risk do not (Condition 2).  dened by (5), from We have noted above that the optimal coefcients a which equation (6) originates, represent the unique solution of the consumers problem for given parameters Rf , g and S, and for given expected returns. Now suppose that the representative consumer holds the optimal f ; a 0 with returns a f Rf a 0 R and let us consider what happens portfolio a when one parameter changes. In equation (5), optimal quantities are obtained for given prices. Then, in this model only one variable between prices and quantities can change. It is common to assume that, for a given f ; a , the adjustment after a change in a parameter optimal allocation a occurs through prices that is, through the excess returns Rex i . Recall, also, that for each asset i an increase (decrease) in Rex i occurs through a decrease (increase) in the asset price Pi. In other words, following a change in some ) and prices (that is, parameter, quantities remain unchanged (equal to a excess returns) adjust. A preliminary step to verify whether Conditions 1 and 2 hold concerns the denition of the riskiness of each asset i. In the CAPM (see Cochrane 2001, Chapter 1), this is dened as li CovRi ; Rm : 7

Given (7), we can rewrite equation (6) as Rex i gli . Consider now a change in risk aversion. It holds that @ Rex i =@ g li . Condition 1 is then fullled in this model. Let us turn to a change in riskiness: @ Rex i =@ li g. At rst glance, Condition 2 seems to hold, because the derivatives @ Rex i =@ li are constant for each i. Then, a simultaneous increase in the riskiness of all assets does not seem to yield a rank effect. However, as Misina (2003) points out, the result that the derivatives of the excess returns on each asset with respect to its riskiness are constant does not necessarily grant that Condition 2 holds: one has to consider explicitly which parameter caused the increase in riskiness.
r 2009 Blackwell Publishing Ltd

134

Marcello Pericoli and Massimo Sbracia

To show that Condition 2 does not hold, we just need to provide a counterexample, which we borrow from Misina (2003). Example. Assume that there are only two assets, denoted by i and j, with 2 variances s2 i and sj and covariance sij. The CAPM is  i s2 j sij ; Rex i g a i a j s2 i sij : Rex j g a j a Suppose that the riskiness of asset i changes because the covariance of asset i j . Clearly, the change in covariance with asset j increases; then @ Rex i =@ sij ga i . Now if will also affect the riskiness of asset j; therefore, @ Rex j =@ sij ga ex ex j , then @ Ri =@ sij will be different from @ Rj =@ sij , yielding a rank effect. i 6 a a This example highlights a general problem (see also Misina 2003): changes in the riskiness of one asset affect expected returns also on other assets. Therefore, all prices (returns) tend to change, and this phenomenon may yield a rank effect unless all assets have equal weights in the investors portfolio. An obvious way to preclude these patterns in the case of unequally weighted assets is to assume that returns are uncorrelated. In this case, the 2 i s2 i s2 CAPM becomes Rex i ga i , where si Var Ri . In this model, li a i.A ex change in risk aversion still yields a rank effect, as @ Ri =@ g li . Consider, ex 2 i . This change, then, may yield a however, a change in s2 i : @ Ri =@ si ga rank effect. For instance, if there are only two assets, say assets i and j, then a j . simultaneous change in their riskiness yields a rank effect as long as  ai 6 a Thus, even with uncorrelated returns, the RAI cannot discriminate between a change in risk aversion and a simultaneous change in the riskiness of all i a j 8i; j. assets, unless a Another possibility is to consider uncorrelated returns and an idiosyncratic shock, that is, a change in the riskiness of only one asset, instead of a simultaneous increase in the riskiness of all assets. With few assets, we can still provide examples in which the rank correlation also changes signicantly in this case. If the cross-section of assets is large, however, it is reasonable to presume that changes in the rank correlation following the change in the riskiness of one asset are small. In fact, consider an idiosyncratic shock to a single asset, say asset i, and suppose that assets i ex and j are such that li4lj while Rex ai > 0, an increase in s2 i < Rj . If  i causes an increase in li that cannot reverse the inequality li4lj; it also causes an ex ex ex increase in Rex i that could reverse the relationship with Rj into Ri > Rj , thereby increasing the rank correlation. But if asset i is the only asset for which we observe some change (in both its riskiness and return), given the independence of the returns on assets, we can expect that, with a sufciently
r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

135

large cross-section of assets, the change in the rank correlation would be negligible.

C. Conclusions from the Theory

The main result of the previous analysis is that, in the context of the standard CAPM, the assumptions granting that the RAI distinguishes between changes in risk and changes in risk aversion are very restrictive. In particular, two alternative sets of assumptions warrant that Conditions 1 and 2 hold. First, assets in the market portfolio are equally weighted. Second, asset returns are uncorrelated, shocks to the riskiness of single assets are idiosyncratic, and the cross-section of assets is large. While it seems implausible that assets in the market portfolio are equally weighted, the second set of assumptions deserves more attention. The assumption that asset returns are uncorrelated is implausible, but one can apply appropriate transformations to asset returns, so that the transformed returns are orthogonal (this is a strategy recently pursued by Misina 2008). Unfortunately, the assumption that shocks to the riskiness of each single asset are idiosyncratic is much more difcult to handle. Measuring risk aversion, in fact, is not very interesting when the economy is hit by a shock affecting only a single rm or a small set of rms. On the contrary, disentangling between risk and risk aversion is crucial during fully edged nancial crises. In those periods, the riskiness of all assets is potentially affected and this is the time when, in order to set an appropriate policy response, it is important to know whether asset prices are driven by increased risk or increased risk aversion. With the theory outlined above, we have identied assumptions granting that the RAI correctly works. What the theory cannot do, however, is to address the extent to which violations of those assumptions bias the measurement of risk aversion provided by the RAI an issue that we explore in the next section.

IV. Empirical Analysis

How large is the bias in the RAI when the assumptions granting that it correctly measures risk aversion are violated? How can we answer this empirical question? In principle, this issue can be addressed in two different ways. One approach would consist in estimating, through Monte Carlo simulations, the effects of departing from each assumption. For the problem at hand, however, this methodology would be extremely cumbersome. In fact, one would need to take into account not only all the assumptions (e.g. number of assets, joint distribution of returns and of the shocks affecting
r 2009 Blackwell Publishing Ltd

136

Marcello Pericoli and Massimo Sbracia

asset riskiness etc.) but also the many ways in which these may not hold. Consider, for example, the hypothesis of independence of asset returns when there is a large number of different assets. Moreover, this effort may not be rewarding, since the assumptions that we have identied above are only sufcient for the RAI to properly measure risk aversion. Hence, we cannot exclude that there are other sets of sufcient conditions that would provide the same result. A more fruitful method, instead, would consist in looking for an indicator that does not suffer from theoretical shortcomings and comparing it with the RAI. This method, albeit less precise in theory, is more relevant for empirical applications. In this section, we adopt this strategy and compare the RAI with a measure of risk aversion derived from the estimation of a CAPM. To estimate risk aversion from the CAPM, we use the classical methodology introduced by Fama and MacBeth (1973). In line with the tradition that estimates risk aversion on a cross-section of returns referred to long time periods (generally from ve to ten years), here, for each month t, we run our estimates on a cross-section of monthly returns referred to the ve years ending in t.11 Thus, we obtain a measure of risk aversion that we can compare with the RAI.

A. Methodology
To measure the risk-aversion parameter gt, we need to estimate the model 12 Rex i;t gt li;t on a cross-section of assets i (i 5 1, . . . , n) at each time t. The assets included in our analysis are the stocks of the Dow Jones Euro Stoxx and those of the Standard & Poors 500; time periods are calendar months from January 1973 to December 2008, data are end-of-month and the source is Thomson Financial Datastream.13 To estimate the model, we need excess returns (Rex i;t ) and the covariance of asset i with the market portfolio (li,t).
11

For instance, in their classical studies, Black et al. (1972) consider non-overlapping veyear periods, Fama and MacBeth (1973) use overlapping periods from ve to eight years and Sharpe (1965) uses a single ten-year period. The parameter gt is the ArrowPratt coefcient of absolute risk aversion in the context of the theoretical model (6); it can be interpreted as an aggregate relative risk aversion of the economy in the context of the more general model discussed in Appendix A.

12

13

In our benchmark estimates, we include only the stocks that are priced during the whole sample period. Hence, the number of stocks is equal to 59 for the euro area and 237 for the United States. In Pericoli and Sbracia (2006), we used all the stocks that were included in the Dow Jones Euro Stoxx and the Standard & Poors 500 at the end of the sample period, whenever they were priced; by doing so, the number of stocks was equal to 59 and 237 in January 1973, but then grew to 320 and 500 at the end of the sample period. The main results were essentially unchanged.

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

137

Computing excess returns (that are equal to ERi;t Rf ) requires nding expected returns E(Ri,t). A standard practice in the literature and a very critical step, as discussed in Section V is to use ex-post returns, hereafter denoted, for the sake of simplicity, by Ri,t. Rational expectations imply that Ri;t ERi;t ei;t , where ei,t is a white noise. Therefore, the model to be estimated is Ri;t Rf ;t gt li;t ei;t ; Rex i;t 8

is the (ex-post) excess return, and Rf,t is approxiwhere Ri;t Rf ;t mated with the one-month Libor on DM-denominated (euro-denominated from January 1999) deposits when we consider euro-area stocks and on dollar-denominated deposits when we consider US stocks.14 In order to obtain the regressors li,t we use the rst step of the two-pass procedure of Fama and MacBeth (1973). Hence, for each asset i, we estimate, using OLS, the following rolling regression:
ex Rex i;k ai bi;t Rm;k Zi;k ;

on a time series of h periods (that is, with k that goes from t h11 to t), where ai is an asset-specic constant, bi,t is the asset beta (namely, it is the covariance between the return on asset i and the return on the market portfolio divided by the variance of the market portfolio), Rex m;k is the excess return on the market portfolio and Zi,k is the residual.15 In our benchmark regression we set h 5 60 months. The outcome of this regression is a point ^ for each asset i and time t.16 The product estimate of the parameter b i;t ^ provides between the variance of Rm in the h months before period t and b i;t ^ an estimate of li,t, denoted by li;t .
14

To avoid introducing a more cumbersome notation, we use the same symbol Rex i;t to denote true (that is, ex-ante) and measured (ex-post) excess returns.

15

We identify the stock market indices with the market portfolios (for the euro area: Dow Jones Euro Stoxx in 19872008 and the euro-area index built by Thomson Financial Datastream in 197386; for the United States: Standard & Poors 500). Note that our exercise only includes stock prices. On this point, the Roll critique (Roll 1977) pointed out that the validity of the model may depend on the assets included in the portfolio: the CAPM, in fact, should include all assets, tradable and non-tradable, tangible or intangible, that add to world wealth. However, Stambaugh (1982) builds a number of market portfolios, which also included government bonds, corporate bonds, Treasury bills, real estate and consumer durables, nding that even when stocks represent only 10% of the market portfolio, inferences about the model are the same as those obtained with a stock-only index. The proper specication of the CAPM requires that asset weights in the market portfolio do not change over time, while in the market index that we have adopted weights do change. However, given the large number of assets contained in our market index, this is usually considered a good working approximation (see Ferson et al. 1987 for further discussions of this issue).
r 2009 Blackwell Publishing Ltd

16

138

Marcello Pericoli and Massimo Sbracia

We are now ready to estimate equation (8) on a cross-section of assets using GLS. Hence, for each time period t we estimate ^ Rex i;t kt gt li;t ei;t ; 9

where gt measures the risk aversion, kt is a constant representing the difference between the true and the measured risk-free rate and ei,t is the residual.17 We address the issues of serial correlation and heteroscedasticity using the NeweyWest estimator with the Bartlett window. The estimated risk aversion ^ gt has to be compared with the RAI. The latter is given, for each time t, by the rank correlation between risks and returns. Following Kumar and Persaud (2002), we measure the rank correlation with Spearmans coefcient, which we denote by rs. Therefore, we consider ex ^ ex ^ the following measure of the RAI: rs t Ri ; li , where we write Ri and li ex ^ instead of Ri;t and li;t to underscore that the correlation is measured for a cross-section of excess returns and risks at each time t. Consistent with the CAPM estimates, we consider excess returns and risks in the 60 months before t. Our application departs from the case study of Kumar and Persaud (2002) in three main respects: (i) we consider monthly stock returns instead of daily exchange-rate returns; (ii) we measure excess returns and risks over the same time period (while they consider non-overlapping time periods and obtain risks from the period that precedes the one used to compute returns); and (iii) we measure risk with a covariance and not with a variance. Points (ii) and (iii), in particular, follow directly from the CAPM and the standard nancial literature.

B. Results
ex ^ Our estimates of ^ gt and rs t Ri ; li for the euro area and the United States are illustrated in Figure 3, together with their condence intervals.18 Let us comment briey on the behaviour of the four indicators over time, and defer a more detailed discussion to Section V. For the euro area, both indicators show several ample uctuations. In particular, the indicators decrease from the late 1980s to the late 1990s, increase until 2001, decline again until 2005, and then rise almost until the end of the sample period, when the latest

17

Recall that according to the SharpeLintner version of the CAPM, kt should be zero. The Black version, instead, allows for kt 6 0; in this case, kt1Rf,t is the return on the zerocovariance portfolio.

18

Condence intervals are obtained from the GLS estimates for the CAPM risk aversion and from bootstrapping for the RAI.

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

Euro area: Dow Jones Euro Stoxx
0.8 0.4 20
RAI

139

0.0 0.4 0.8

10 0 10 20 1980 1985 1990 1995 2000 2005

United States: S&P 500

0.8 0.6 0.4 0.2 20
RAI CAPM risk aversion

0.0 0.2 0.4

15 10 5 0 5 10 1980 1985 1990 1995 2000 2005

Figure 3: CAPM risk aversion and Risk Appetite Index. The thin lines show the 95% condence intervals

observations record a deep decline. For the United States, the results are somewhat different: the two indicators increase from the mid-1980s to the late 1990s, decrease until 2005, and then rise again almost until the end of the sample period, when the latest observations show the same deep decline as the euro-area indicators. In the two areas, the average value of risk aversion from the CAPM is equal to 2.6, a result broadly in line with those of the literature (see the classic Friend and Blume 1975, or, for a survey, Cochrane 2001, Chapter 21, and the references therein). The values taken by the RAI, which is a correlation coefcient, are always included in the interval [ 1, 1]. Recall from Section II, however, that the RAI can only detect changes in risk aversion. Therefore, its level is not very informative. On the other hand, its behaviour over time is what matters most and, as Figure 3 shows,
r 2009 Blackwell Publishing Ltd

CAPM risk aversion

140

Marcello Pericoli and Massimo Sbracia

this behaviour turns out to be very similar to that of the CAPM-based indicator. The strength of the correlation between the RAI and the CAPM-based measure of risk aversion, equal to 0.85 for the Dow Jones Euro Stoxx and 0.79 for the Standard & Poors 500, is the main result of this analysis and proves to be quite robust to several changes in the estimation strategy. In particular, rst we have focused on the assets included in the regressions. In fact, plugging directly an estimate of li,t into equation (8) causes an error-in-variable problem, which is usually addressed by gathering stocks into portfolios in order to increase the precision of the betas (see Campbell et al. 1997, Chapter 5). Hence, along the lines of Fama and French (1992), we have considered (i) 20 value-weighted portfolios obtained by ranking the stocks by size and (ii) 25 portfolios obtained by ranking stock returns in 5 percentiles by size and 5 by book-to-market value and considering their intersection. Second, as an alternative method of addressing the problems of heteroscedasticity and correlation of the residuals, we have included in the regression two other explanatory variables, namely, the logarithm of market capitalization (Schwert 1983) and the systematic skewness (Kraus and Litzenberger 1976).19 Third, we have also experimented with different indices representing the market portfolio, such as indicators with equally weighted stocks and indices with a subset of value-weighted stocks. Similarly, we have performed regressions using as assets only the stocks that were always listed in the two indices during the entire sample period. Fourth, we experimented with different sizes of the moving window in the time-series regressions, using h 5 36, 24 and 12 months. This set of robustness tests is the only one that has produced signicant differences in the behaviour of risk aversion (but not in the correlations between the RAI and the CAPM-based indicator, which remained strong) with respect to the benchmark regression. Specically, by shrinking the size of the moving window, we have obtained, not surprisingly, more volatile dynamics of the risk aversion. As mentioned, however, the strong correlation between the RAI and the CAPM-based measure of risk aversion survives all our robustness tests. In fact, seldom have we obtained a correlation smaller than 0.7. Thus, despite the theoretical differences discussed in Section IV, the RAI and the CAPMbased measure of risk aversion provide essentially the same results. We devote the next section to the explanation of this puzzle.
19

The systematic skewness is introduced in order to account for the possible effect of higher-order moments of the utility function of the representative investor. Following Kraus and Litzenberger (1976), we compute it as ERi ERi Rm ERm 2 =ERm ERm 3 .

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

141

C. Explaining the Similarities

ex ^ The reason why the dynamics of ^ gt are similar to those of rs t Ri ; li lies in the statistical properties of both indicators. First, recall that Spearmans rank correlation rs is a robust statistic for the linear correlation r (see Huber 1981, Chapter 8, and Stuart and Ord 1991, 20 ex ^ ex ^ Chapter 26). Thus, not surprisingly, it is rs t Ri ; li rt Ri ; li for any t. Hence, our problem becomes that of explaining why the dynamics of ^ gt are ^ ; l . similar to those of rt Rex i i Second, the OLS estimate of gt in the cross-section of model (8) is given by ^ ^ CovRex i ; li scaled by Var li , which we can rewrite as ^ li ct ; 10 g r Rex ; ^ t t i

denotes the linear correlation between Ri and ^ li in the where cross-section of assets at time t and " #1=2 Var Rex i ct : 11 Var ^ li Equation (10) implies that if the variability of ct is sufciently lower than that of ^ ^ gt over time is dominated by that of rt Rex rt Rex i ; li , then the behaviour of ^ i ; li . ex In other words, if Var Ri is sufciently low with respect to Var ^ li , then the ^ ; l , which is the result that we are dynamics of ^ gt are similar to those of rt Rex i i trying to explain. The required condition is largely met in our as in h two samples, i ^ both of them Var ct is about 104 times smaller than Var rt Rex ; l . i i To sum up, although the analysis performed in Section III suggested that one needs very restrictive theoretical assumptions to warrant that the RAI is an indicator of risk aversion, the evidence provided in Figure 3 shows that the RAI tends to be a proxy of a standard CAPM-based risk-aversion indicator. In other words, the RAI and the CAPM-based indicator provide essentially the same information about the behaviour of risk aversion. The similarity between them stems from two main elements: (i) the rank correlation is a proxy for the linear correlation;21and (ii) if the variance ratio (11) is sufciently low, the dynamics of the regression coefcient in the equation that explains returns with risks are approximately equal to those of the linear correlation between returns and risks. In this case, violations of the theoretical assumptions discussed in Section III are not empirically relevant.
20

^ rt Rex i ; li

For both samples, the correlation between r and rs is about 0.97 in our benchmark, and remains very high in all the robustness tests performed.

21

Since the RAI is based on a robust measure of correlation, it has the advantages of other robust statistics (for instance, it is unaffected by small departures from model assumptions, such as the presence of outliers, non-normality, etc.; see Huber 1981). Its problems, however, lie elsewhere, as we discuss in Section V.
r 2009 Blackwell Publishing Ltd

142

Marcello Pericoli and Massimo Sbracia

V. Limitations
Besides the similarity between the RAI and the CAPM-based indicator, Figure 3 reveals two important problems.22 First, both indicators appear to be correlated with market returns. In fact, as shown more precisely in the rest of this section, even though there appear to be some asynchronicities with market peaks, these are mostly due to the use of a 60-month window to compute returns (and, as we discuss below, to changes in the variance of market returns). Second, the CAPM-based indicator that, we recall, measures the level of risk aversion, is often negative. These two problems are two sides of the same coin, that is, the inability to measure expected returns correctly the key limitation that impairs the estimation of risk aversion. Let us reconsider equation (6), which states that Rex i g CovRi ; Rm for any asset i. Of course, the equation must hold not only for each individual assets but also for any portfolio of assets, including the market portfolio. Then, by substituting into (6) the return on the market portfolio, we obtain the well-known result that g Rex m : Var Rm 12

In other words, equation (12) states that risk aversion can also be measured with the expected excess return on the market portfolio scaled by its variance. Equation (12) masks the relationship between the RAI and the CAPM-based indicator (because these are based on returns on individual assets), but sheds light on the main problems that affect both indicators, allowing us to simplify the discussion. For the sake of comparison, to compute g using equation (12), we keep on measuring ex-ante with ex-post returns and we maintain a 60-month rolling window. More precisely, the expected excess return on the market at time t, Rex m , is the realized excess return on the market index (Dow Jones Euro Stoxx for the euro area and Standard & Poors 500 for the United States) in the 60 months ending in t, while Var Rm is its monthly variance in the period [t 60, t].23 Figure 4 shows the behaviour of g over time, computed using
22

We are very grateful to an anonymous referee for outlining the main issues discussed in this section.

The variance at the denominator of equation (12), used to compute g as represented in Figure 4, is the variance of excess market returns. If returns on the risk-free asset were constant, then the variance of excess market returns would be equal to that of (gross) market returns. However, with returns on the risk-free asset with a low time-variability as in our sample, the difference between the two variances is negligible and does not make any difference in the behaviour of risk aversion.
r 2009 Blackwell Publishing Ltd

23

CAPM and the Risk Appetite Index

Euro area: Dow Jones Euro Stoxx
12

143

8 1980 1985 1990 1995 2000 2005

CAPM risk aversion

mean/variance ratio

United States: S&P 500

16

12

4 1980 1985 1990 1995 2000 2005

CAPM risk aversion

mean/variance ratio

Figure 4: CAPM risk aversion and mean/variance ratio of the market portfolio, is the indicator of risk aversion computed using equation (12)

equation (12), together with that of the CAPM-based indicator derived from equation (9). Not surprisingly, the two variables are strongly correlated for both the euro area and the United States (correlation is above 0.5). The behaviour of risk aversion according to these indicators (as well as according to the RAI) is disappointing and does not make much economic sense. For instance, in the last part of the sample period, with the global nancial crisis still ongoing, the indicators signal a marked decline in risk
r 2009 Blackwell Publishing Ltd

144

Marcello Pericoli and Massimo Sbracia

aversion that, clearly, simply reects the corresponding decline in stock prices. Similarly, notice also that a substantial number of observations are negative for all four indicators a problem that, in the RAI, is hidden by the use of a correlation coefcient and by the focus on changes in risk aversion, which make its level uninformative. Equation (12), then, helps us to shed light on the main problem. The expost return on the market, Rm, can be decomposed into a measurement error term, Rm E(Rm), plus the ex-ante return E(Rm), which is the variable that is needed in order to estimate g. Unfortunately, time variation in the measurement error dominates variation in the ex-ante return. In fact, over a 60-month window, the mean of monthly returns has a standard deviation of about 10% (annual returns have a standard deviation of about 20% in our sample period), which is much larger than any plausible time variation in expected market returns. For instance, predictive regressions with forecasters such as dividend yields or price/earning ratios have much lower standard deviations of about 34%.24 Thus, the dynamics of risk aversion measured with indicators based on ex-post returns (RAI and CAPM-based indicators) are mostly driven by measurement errors.25

VI. Conclusion
This paper provides a theoretical and an empirical analysis of the RAI of Kumar and Persaud (2002). The theory shows that the RAI does not necessarily identify changes in risk aversion. Theoretical assumptions granting that it correctly distinguishes between risk and risk aversion can be found, but are very restrictive. By comparing the RAI with a measure of risk aversion derived from the direct estimation of a CAPM, a model that does not require theoretically restrictive assumptions, we also nd that the behaviour of these two indicators is surprisingly similar. This puzzle can be easily explained. If the variance of returns is sufciently smaller than the variance of asset riskiness (a condition that is met in all our case studies), then the CAPMbased indicator is approximately equal to the linear correlation between risks and returns. The RAI, we recall, is dened as the rank correlation
24

For an early analysis, see Campbell and Shiller (1988), who nd that, in fact, actual returns are about four times more variable than predicted returns, even though their correlation is quite high (over 0.9 in their exercise). It is worth noting that Kumar and Persaud (2002) measure expected returns with forward exchange rates rather than with ex-post returns on currencies. Hence, their methodology is less affected by measurement errors. The problem is more relevant for later applications of the RAI to stock indices, including those performed in the studies mentioned in footnote 4, which use ex-post returns.

25

r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

145

between the same two variables. Therefore, the RAI and the CAPM provide essentially the same information about the behaviour of risk aversion. To put it differently, violations of the theoretical assumptions granting that the RAI identies changes in risk aversion are not empirically relevant. Our analysis also shows, however, that the RAI suffers from exactly the same limitations that impair the estimation of risk aversion from the CAPM. In particular, the main problem stems from the difculties in measuring expected returns correctly. If one measures them with ex-post returns, as is common in the literature, then at high and medium frequencies the measurement error dominates the estimates. As a consequence, measured risk aversion is often negative (in declining stock markets) and its behaviour tends to be incongruous. Therefore, future studies aimed at measuring changes in risk aversion with the RAI and the CAPM should rst address this problem. Once this question is solved, the RAI can broadly be regarded as a robust statistic of the CAPM-based indicator of risk aversion, as it is based on a robust measure of correlation (Spearmans rank correlation), whereas the CAPM-based indicator is approximately given by a linear correlation. Marcello Pericoli and Massimo Sbracia Bank of Italy Via Nazionale 91 00184 Rome Italy marcello.pericoli@bancaditalia.it and massimo.sbracia@bancaditalia.it

References
Arrow, K. J. (1970), Essays in the Theory of Risk Bearing. Amsterdam: NorthHolland. Baek, I. (2006), Portfolio Investment Flows to Asia and Latin America: Pull, Push or Market Sentiment?, Journal of Asian Economics, 17(2), 36373. Baek, I., A. Bandopadhyaya and C. Du (2005), Determinants of Market-Assessed Sovereign Risk: Economic Fundamentals or Market Risk Appetite?, Journal of International Money and Finance, 24(4), 53348. Bandopadhyaya, A., and A. L. Jones (2006), Measuring Investor Sentiment in Equity Markets, Journal of Asset Management, 7(3), 20815. Bernoulli, D. (1738), Specimen Theoriae Novae de Mensura Sortis, Commentarii Academiae Scentiarum Imperialis Petropolitanae, Tomus 5, 17592. Translated into English as: Exposition of a New Theory on the Measurement of Risk, Econometrica, 12, 2336, 1954.
r 2009 Blackwell Publishing Ltd

146

Marcello Pericoli and Massimo Sbracia

BIS (2004), BIS Quarterly Review International Banking and Financial Market Developments, March. Basel: Bank for International Settlements. Black, F., M. C. Jensen and M. S. Scholes (1972), The Capital Asset Pricing Model: Some Empirical Tests, in M. C. Jensen (ed.), Studies in the Theory of Capital Markets. New York: Praeger, pp. 79121. Campbell, J. Y., A. W. Lo and A. C. MacKinlay (1997), The Econometrics of Financial Markets. Princeton, NJ: Princeton University Press. Campbell, J. Y., and R. J. Shiller (1988), Stock Prices, Earnings, and Expected Dividends, Journal of Finance, 43(3), 66176. Cochrane, J. H. (2001), Asset Pricing. Princeton, NJ: Princeton University Press. Fama, E. F., and K. R. French (1992), The Cross-Section of Expected Stock Returns, Journal of Finance, 47(2), 42765. Fama, E. F., and J. D. MacBeth (1973), Risk, Return and Equilibrium: Empirical Tests, Journal of Political Economy, 81(3), 60736. Ferson, W. E., S. Kandel and R. F. Stambaugh (1987), Tests of Asset Pricing with Time-Varying Expected Risk Premiums and Market Betas, Journal of Finance, 52(2), 20120. Friend, I., and M. E. Blume (1975), The Demand for Risky Assets, American Economic Review, 65(5), 90022. Gai, P., and N. Vause (2004), Risk Appetite: Concept and Measurement, Financial Stability Review, Bank of England, London, December. Hartog, J., A. Ferrer-i-Carbonell and N. Jonker (2002), Linking Measured Risk Aversion to Individual Characteristics, Kyklos, 55(Fasc. 1), 326. Huang, C., and R. H. Litzenberger (1988), Foundations for Financial Economics. Englewood Cliff, NJ: Prentice Hall. Huber, P. J. (1981), Robust Statistics. New York: Wiley. Illing, M., and A. Meyer (2005), A Brief Survey of Risk-Appetite Indexes, Financial System Review, Bank of Canada, Ottawa, June. IMF (2002), Global Financial Stability Report Market Developments and Issues, March. Washington, DC: International Monetary Fund. IMF (2003a), Global Financial Stability Report Market Developments and Issues, March. Washington, DC: International Monetary Fund. IMF (2003b), Global Financial Stability Report Market Developments and Issues, September. Washington, DC: International Monetary Fund. Kraus, A., and R. Litzenberger (1976), Skewness Preference and the Valuation of Risky Assets, Journal of Finance, 31(4), 1085100. Kumar, M. S., and A. Persaud (2001), Pure Contagion and Investors Shifting Risk Appetite: Analytical Issues and Empirical Evidence, IMF Working Paper No. 01/ 134.
r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

147

Kumar, M. S., and A. Persaud (2002), Pure Contagion and Investors Shifting Risk Appetite: Analytical Issues and Empirical Evidence, International Finance, 5(3), 40136. Laffont, J. J. (1993), The Economics of Uncertainty and Information. Cambridge, MA: MIT Press. Misina, M. (2003), What Does the Risk-Appetite Index Measure?, Bank of Canada Working Paper No. 2003-23. Misina, M. (2008), Changing Investors Risk Appetite: Reality or Fiction?, European Journal of Finance, 14(6), 489501. Pericoli, M., and M. Sbracia (2004), The Risk Appetite Index: Theoretical and Empirical Issues, Mimeo, Bank of Italy. Pericoli, M., and M. Sbracia (2006), The CAPM and the Risk Appetite Index, Temi di discussione No. 586, Bank of Italy. Roll, R. (1977), A Critique of the Asset Pricing Theorys Test. Part I: On Past and Potential Testability of the Theory, Journal of Financial Economics, 4, 12976. Schwert, G. (1983), Size and Stock Returns and Other Empirical Regularities, Journal of Financial Economics, 12, 34160. Sharpe, W. F. (1965), Risk-Aversion in the Stock Market: Some Empirical Evidence, Journal of Finance, 20(3), 41622. Stambaugh, R. F. (1982), On the Exclusion of Assets from Tests of the Two-Parameter Model: A Sensitivity Analysis, Journal of Financial Economics, 10, 23768. Stuart, A., and K. Ord (1991), Kendalls Advanced Theory of Statistics. Volume 2: Classical Inference and Relationship, 5th edn, London: Edward Arnold.

Appendix A: Heterogeneous Agents with Risk-Averse Preferences

Section III was based on the restrictive assumptions that agents are identical and that their preferences are given by CARA utility functions. Following Huang and Litzenberger (1988), here we discuss a more general setting where agents are heterogeneous in both preferences and wealth. Our conclusions on the RAI from this more general model remain essentially unchanged once we focus on an appropriately specied global risk-aversion parameter. This is because this setting leads to a version of the CAPM that has the same functional form as equation (6).
r 2009 Blackwell Publishing Ltd

148

Marcello Pericoli and Massimo Sbracia

Assume that preferences are represented by increasing and concave utility functions. With respect to the previous section, we maintain the assumption that asset returns are multivariate normally distributed. In addition, we redene the problem in terms of shares of wealth rather than in dollar value terms. Namely, denote by ai,h the amount of wealth invested in asset i by investor h, h 5 1, . . . , H. The share of investor hs wealth, Wh, invested in that asset is wi,h 5 ai,h/Wh. P The total wealth of the H investors in the economy is Wm H h1 Wh . In equilibrium, the total wealth Wm is equal to the total value of the assets. We denote with wi,m the portfolio weight of asset i in the market portfolio, namely
H X h1

wi;m

wi;h

Wh : Wm

Using the budget constraint (3), we can rewrite the consumption Ch P af ;h Rf n i1 ai;h Ri of the investor h endowed with wealth Wh as
n X i1 n X i1

! wi;h Wh Rf

Ch

Wh "

n X i1

wi;h Wh Ri

Wh Rf

wi;h Ri Rf :

The maximization problem of this investor, whose preferences are represented by the increasing and concave function uh 2 C2 , becomes ( " max E u h W h Rf 0
a n X i1

!#) wi;h Wh Ri Rf : A:1

Let us assume that a solution to (A.1) exists. Since uh is concave, the rstorder (necessary) conditions are also sufcient and are " E u0h W h Rf
n X i1

!  i;h Wh Ri Rf w

# Ri Rf 0 ; A:2

8i 1; . . . ; n;  i;h are the optimal shares of wealth invested in asset i where the coefcients w by individual h. The optimal consumption of this investor then is
r 2009 Blackwell Publishing Ltd

CAPM and the Risk Appetite Index

 h W h Rf C
n X i1

149

 i;h Wh Ri Rf : w

Using the denition of covariance, we can rewrite condition (A.2) as  h ERi Rf Cov u0 C  E u0h C h h ; Ri 0 : In addition, using Steins lemma we obtain  h ; Ri E u00 C  h CovC  h ; Ri ; Cov u0h C h which we can substitute back into the previous expression. Recalling that ERi Rf Rex i , we nd  h Rex E u00 C  h CovC  h ; Ri : A:3 E u0h C i h We can now dene the global absolute risk aversion of the investor h as  E u00 h Ch Gh : h E u0h C  Dividing both terms of (A.3) by E u00 h Ch , summing across the H investors, and rearranging we obtain !1 H H X X  h ; Ri : G 1 CovC Rex
i h h1 h1

Note that the rst term in brackets on the right-hand side is the harmonic mean of the investors global absolute risk aversions. Moreover, we can write ! H H X X  h ; Ri  h ; Ri Cov C CovC
h1 h1

Cov

n X H X i1 h1

! wi;h Wh Ri ; Ri ! wi;m Ri ; Ri

Cov Wm

n X i1

Wm CovRm ; Ri : Substituting into the previous expression and putting ! 1 H X 1 G Wm G ; h

h1

which represent a sort of aggregate relative risk aversion of the economy, we obtain Rex i G CovRi ; Rm ;
r 2009 Blackwell Publishing Ltd

150

Marcello Pericoli and Massimo Sbracia

which has exactly the same functional form as (6). This equation shows that our conclusions on the RAI do not depend on the assumptions of identical agents with CARA preferences. For instance, as long as asset returns are multivariate normally distributed, they hold for any non-satiated and riskaverse preferences.

r 2009 Blackwell Publishing Ltd