A Model-Free CAPM with High Order Risks *

Transcription

1 A Model-Free CAPM with High Order Risks * Te-Feng Chen Department of Finance Hong Kong University of Science and Technology Clear Water Bay, Kowloon, Hong Kong tfchen731@gmail.com Tel: (852) Fax: (852) San-Lin Chung Department of Finance National Taiwan University Taipei, Taiwan chungsl@ntu.edu.tw Tel: (886) Fax: (886) K.C. John Wei Department of Finance Hong Kong University of Science and Technology Clear Water Bay, Kowloon, Hong Kong johnwei@ust.hk Tel: (852) Fax: (852) This version: August 27, 2014 * We appreciate the helpful comments and suggestions from Yakov Amihud, Chuang-Chang Chang, Huimin Chung, Jin-Chuan Duan, Shing-Yang Hu, Chuan Yang Hwang, and Yaw-Huei Jeffrey Wang. John Wei acknowledges financial support in the form of a Research Infrastructure Grant from Hong Kong s Research Grants Council (RI/93/94.BM02). All remaining errors are ours.

2 A Model-Free CAPM with High Order Risks Abstract Allowing for non-normality in market returns, we present an approximate capital asset-pricing model in which besides first-order co-moment risks, higher-order co-moment risks are also important for pricing individual stocks. We find that the second-order risk is significantly and negatively priced and contributes to the inverse U-shaped relation between expected returns and betas. We then construct three risk factors with each to mimic second-order risk premiums. We show that each of our newly constructed second-order risk factors is related to the market variance risk premium. More importantly, a two-factor model consisted of a market factor and one of our second-order risk factors can fully explain the total (idiosyncratic) volatility puzzle and the MAX puzzle and partially explain the betting against beta premium. Keywords: Model-free CAPM; Cumulants; High order risks; Nonlinear risk-return trade-off.

3 1. Introduction The capital asset-pricing model (CAPM) developed by Sharpe (1964), Lintner (1965), and Black (1972) predicts that expected stock returns are positively and linearly related to their systematic risk measured by market beta. Earlier studies find support for the CAPM that there is a simple positive relation between beta and expected return during the early period (Black, Jensen, and Scholes, 1972; Fama and MacBeth, 1973). However, recent studies show that the relation between beta and expected return is flat during the recent period (Fama and French, 1992). In addition, stock returns are significantly related to some firm characteristics such as firm size, book-to-market equity, profitability, and capital investment (Fama and French, 1992, 1995, 2006, 2008; Novy-Marx, 2013; Titman, Wei, and Xie, 2004). All these motivate Fama and French (1993, 2014) and others (e.g., Hou, Xue, and Zhang, 2012) to propose the use of empirically motivated linear asset-pricing models with multiple factors, in the spirit of Merton (1973), to explain the cross section of stock returns. At the meantime, another strand of asset-pricing models assumes that investors consider higher order moments in stock returns. For example, Kraus and Litzenberger (1976), Sears and Wei (1985), Harvey and Siddique (2000), Dittmar (2002), Chung, Johnson, and Schill (2006), Mitton and Vorkink (2007), and Barberis and Huang (2008), among others, show that skewness or co-skewness is important in pricing the cross section of stock returns. Chang, Christoffersen, and Jacobs (2013) further find that risk-neutral skewness significantly predicts future individual stock returns. Due to its inherited nature, the linear asset-pricing models have difficulty in explaining the pricing effect embedded in higher orders of asset returns. For example, Ang, Hodrick, Xing, and Zhang (2006; 2009) find that stocks with higher total or idiosyncratic volatility earn lower future 1

4 returns, a phenomenon which they refer to as the volatility puzzle. Bali, Cakici, and Whitelaw (2011) report that stocks with the maximum daily returns over the past month earn substantially lower returns in the following month. We refer this as the MAX puzzle. Moreover, Baker, Bradley, and Wurgler (2011) document that high beta stocks earn lower future returns, which contradicts to the predictions of any linear asset pricing models. A recent study by Frazzini and Pedersen (2014) further documents that a betting against beta (BAB) strategy of longing leveraged low beta assets and at the same time shorting high beta assets produces significantly positive risk-adjusted returns. More importantly, the volatility puzzle, the MAX puzzle, and the BAB puzzle cannot be explained by the existing linear asset-pricing models, including the Fama and French three-factor model augmented by the Carhart (1997) momentum factor. All these motive us to explore whether an asset-pricing model that incorporates the nonlinear pricing of the systematic risks can explain these puzzles. We assume that the market return is non-normal and that an individual stock s log returns follow a simple linear model with the market s log return and that the linear structure is preserved in the risk-neutral measure. We then characterize the dynamic of the market return through the cumulant generating function, which provides analytical solution to our model and shows that the market return has high-order risk premiums. 1 Based on these assumptions, we are able to present an approximate model-free capital asset-pricing model in which in addition to first-order co-moment risks documented in the existing literature, higher-order co-moment risks 1 To be discussed in more detail later, the cumulant generating function of a random variable is defined as the logarithm of the moment generating function. The j-th cumulant, which is defined as the j-th derivative of the cumulant generating function evaluated at zero, is related to the j-th moment. Since the pioneer work of Jarrow and Rudd (1982) on option valuation using the approximation method, a growing literature shows that the cumulant generating function can be used to quantify the impact of higher moments on the pricing structure of implied volatility (Backus, Foresi, Li, and Wu, 1997; Bakshi, Kapadia, and Madan, 2003), to identify the risk-neutral measure for heteroskedasticity volatility models (Christoffersen, Elkamhi, Feunou, and Jacobs, 2010; Corsi, Fusari, and La Vecchia, 2013), and to study the equity premium in representative agent models with non-normal distributions (Backus, Chernov, and Martin, 2011; Martin, 2013; Duan and Zhang, 2014; Backus, Chernov, and Zin, 2014). 2

5 and their corresponding higher-order risk premiums are also important for pricing individual stocks. The intuition for this nonlinear risk-return trade-off in our model is simple. Since the market return has higher-order risk premiums due to non-normality, expected individual stock returns should comprise compensation for investors bearing the corresponding high-order systematic risks. To gauge the impact of high order risk premiums, we estimate the market risk premium in each moment (more specifically risk-neutral moments) implied by the S&P 500 index (SPX) and the SPX index options, following the model-free methodology proposed by Bakshi, Kapadia, and Madan (2003). 2 The estimation result shows that the market mean risk premium and the market skewness risk premium are positive. In contrast, the market variance risk premium and the market kurtosis risk premium are negative. We then perform cross-sectional regressions using returns on Fama and French 25 size-b/m portfolios as well as returns on portfolios formed on the basis of historical CAPM betas. Interestingly, we find that the first-order risk is significantly and positively priced and the second-order risk is significantly and negatively priced, while higher order risks are not significantly priced. The finding suggests that an asset-pricing model consisted of the first two order risks are sufficient to price individual stocks in the cross section and that the cross-sectional relation between expected return and market beta is inversely U-shaped. Our finding of a negative second-order risk premium is consistent with the well-documented evidence for the negative market variance risk premium (e.g., Bollerslev, Tauchen, and Zhou, 2009; Carr and Wu, 2009; Bollerslev, Marrone, Xu, and Zhou, 2013). 3 2 It has been shown that the risk-neutral moments can be inferred in a model-free fashion from a collection of option prices without the use of a specific pricing model (see, for example, Carr and Madan, 1998; Britten-Jones and Neuberger, 2000; Bakshi, Kapadia, and Madan, 2003; Jiang and Tian, 2005). 3 While Driessen, Maenhout, and Vilkov (2009) suggest that individual options covered by the S&P 100 index do not embed a negative variance risk premium, Han and Zhou (2012) find that, with the realized variance measured from high frequency stock prices, the individual variance risk premium is significantly negative in their larger 3

6 We next examine the economic value of the second-order risk premium with three different approaches. First, we examine the second derivative (i.e., curvature) of expected returns with respect to the first-order co-moment risk (i.e., market beta). We separately form portfolios based on the stock return sensitivity to the market return (i.e., market beta), the stock return sensitivity to the squared market return (i.e., first-order coskewness), and the stock return sensitivity to the cubic market return (i.e., first-order cokurtosis). Consistent with the inverse U-shaped return-beta pattern implied by the model, all of the first-order co-moment risk sorted portfolios exhibit the same return pattern that stocks in the bottom and top portfolios have lower returns than stocks in the middle portfolios. To exploit the economic value of the inverse U-shaped return-risk pattern, we construct a curvature portfolio (CUR1) by the sum of the twice difference in the portfolios formed on each of the first-order co-moment risks. 4 We find that the average annual return on the curvature portfolio is significantly negative at 12.00% for the market beta sorted portfolios, at 11.76% for the first-order coskewness sorted portfolios, and at 11.52% for the first-order cokurtosis sorted portfolios. Second, we examine the first derivative of expected returns with respect to the second-order co-moment risk (i.e., market beta squared), which includes the squared stock return sensitivity to the market return (i.e., second-order coskewness) and the squared stock return sensitivity to the squared market return (i.e., second-order cokurtosis). We then construct another curvature portfolio (CUR2) that longs the top portfolio and shorts the bottom portfolio formed on each of the second-order co-moment risks. We find that the average annual return on this curvature portfolio is significantly negative at 15.24% for the second-order coskewness sorted portfolios and at 12.72% for the second-order cokurtosis sorted portfolios. Third, we estimate the sample of equity options covered by OptionMetrics. Bali and Hovakimian (2009) and Han and Zhou (2012) also show that individual variance risk premiums can predict the cross-sectional stock returns. 4 We will provide the detailed definition later. 4

7 risk-neutral variance beta (i.e.,,, ), which essentially corresponds to the market beta squared, using a linear market model based on the second moment. While the pricing implication of the co-moment risks depends on how the utility preference (i.e., the pricing kernel) is specified, the empirical tests based on the risk-neutral variance beta are genuinely mode-free since these estimates are inferred in a preference-free fashion. We find a similar result for the systematic risk premium estimated from the risk-neutral moments using individual equity options and index options. The result shows that the average annual return on the curvature portfolio (CUR2RN), which longs the top portfolio and shorts the bottom portfolio formed on,,, is significantly negative at 16.08% for the risk-neutral variance beta sorted portfolios. We finally investigate whether the second-order risk can help explain the cross section of stock returns. We construct risk factors that mimic second-order risk premiums using the curvature portfolios discussed above. Stocks with high curvature factor loadings, by construction, are less risky because they are more sensitive to the market variance risk, thereby providing hedging against market volatility risk. We find that mimicking curvature factors are priced risk factors. We then investigate whether our newly constructed mimicking factors are able to explain the anomalies that are essentially exposed to the second-order systematic risk. We show that our mimicking factors are able to explain the total/idiosyncratic volatility puzzle documented by Ang, Hodrick, Xing, and Zhang (2006) as well as the MAX puzzle by Bali, Cakici, and Whitelaw (2011). Moreover, we find evidence that our mimicking factors help explain the betting against beta (BAB) anomaly documented by Frazzini and Pedersen (2014). Although Frazzini and Pedersen attribute the anomaly to margin constraints, our model, in contrast, suggests that the BAB anomaly is exposed to high order systematic risks. Our paper is closely related to, but different from, Hong and Sraer (2012), in which they 5

8 demonstrate that the disagreement about the market return leads to speculative overpricing for high beta stocks. Their model implies that the shape of the security market line (SML) is kinked and the slope of the SML decreases with the aggregate disagreement. Our model, in contrast, suggests that the curvature of the SML is determined by the market variance risk premium (i.e., the second-order risk premium). More importantly, we find evidence that the market variance risk premium explains our curvature factors better than the aggregate disagreement. Our paper also complements Conrad, Dittmar, and Ghysels (2013), in which their focus on systematic risks is limited to first-order risk-neutral co-moments. In contrast, we show that the second-order systematic risk measured from the risk-neutral variance beta is strongly related to future stock returns. Furthermore, we find evidence that our mimicking factors are priced in the firm-level cross-sectional regression and that the results are robust to the inclusion of individual risk-neutral moments. The remainder of the paper is organized as follows. The next section presents our approximate capital asset pricing model. Section 3 discusses the empirical implications of our model. Section 4 describes the data and presents the estimation of co-moment risks. In Section 5, we show empirical evidence on the second-order risk premiums in the cross section of stock returns. Section 6 focuses on the construction and verification of the mimicking curvature factors. In Section 7, we test the performance of mimicking curvature factors in explaining the cross section of stock returns. Finally, Section 8 contains our concluding remarks. 2. The model 2.1. The market risk premium and the cumulant generating function We derive asset prices based on the pricing kernel,. Denoted any asset return as 6

9 , where ( ) is the stock price at time t (t+1) and is dividend paid between t and t+1. Then the standard asset pricing suggests that 1, (1) where is the expectation operator at time t. Exploiting the pricing condition for the risk-free return,, it follows that the expected value of the pricing kernel is a discount factor with the risk-free rate, i.e.,,. The risk-neutral measure Q, which corresponds to the physical measure P, is defined by the Radon-Nikodym derivative, Q. The P standard asset-pricing condition implies that Q,. (2) Eq. (2) suggests that the expected value of any gross return in the risk-neutral measure is the gross risk-free rate. The physical dynamic of the market return,,, where W is the value of the market portfolio, is defined by its cumulant generating function, (i.e., the logarithmic value of the moment generating function),, log,. The power-series expansion yields,,, (3)! where, is the n-th derivative of, at 0, corresponding to n-th moment of,. In particular,, is the mean (, ),, is the variance (, ),, is 7

10 the unstandardized skewness, and, is the unstandardized excess kurtosis. 5 The risk-neutral dynamic of, is similarly defined by its cumulant generating function, Q, Q,!, where Q, corresponds to the n-th risk-neutral moment of,. We define the expected market risk premium or the excess market return as,,, log exp,,. 6 The risk-neutral pricing Eq. (2) implies that Q, 1,. Therefore, the market risk premium can be conveniently expressed in terms of the cumulant generating functions as follows (ignoring the terms higher than the fourth moment):,, 1 Q, 1, Q, 1 2, Q, 1 6, Q, (4) 1 24, Q,,,,,. The result implies that the market premium, in general, comprises all of the risk premiums stemmed from the differences between the physical moments and the risk-neutral moments. Therefore, the market risk premium can be approximately decomposed into the market mean risk premium (, ), the market variance risk premium (, ), the market skewness risk premium (, ), and the market kurtosis risk premium (, ) An approximate capital asset-pricing model 5 More specifically,,,, and,,3,, where,,,. 6 The gross market return from t to t+1 is defined as 1, exp,, where, is the simple market return ( ), and the gross risk-free rate from t to t+1 is known at time and is similarly defined as 1, exp,. Hence, the expected excess simple market return,,, explog exp, exp, can be approximated by the log expected excess gross market return. That is,,, log exp,,. 8

11 For any asset return,,,, we assume that its log return follows a simple, linear model with the log gross market return as follows:,,,,,, (5) where, is the idiosyncratic component, which is independent of the market return, i.e.,,,. It follows that the cumulant generating function of, is,,,,,,, where,, is the cumulant generating function of,. We further assume that the idiosyncratic component is independent of the pricing kernel (i.e.,, ) and that the linear market model structure is preserved in the risk-neutral measure. Therefore, the risk-neutral dynamic of, is represented by, Q, Q,,,,. Now, we are ready to calculate the expected excess stock return defined as,,,,,,, 1 Q,, 1. It follows that,,, Q,,. (6) The power-series expansion (ignoring the terms higher than the fourth moment) yields the following proposition for,. Proposition 1. (Nonlinear beta representation of expected excess stock return),, Q,, 1 2, Q,, 1 6, Q,, 1 24, Q,,,,,,,,,,. (7) The result implies the feature of a nonlinear risk-return relation. The first term,,,, measures the first-order risk premium of the classical CAPM, whereas the remaining terms 9

12 capture the pricing effect for the market s higher moment risk premiums. In particular, the market variance risk premium (, Q, ) determines the risk price (, for the second-order systematic risk (, ). Similarly, the scaled market skewness risk premium (, ) and the scaled market kurtosis risk premium (, ) are relevant for the third-order systematic risk (, ) and the fourth-order systematic risk (, ), respectively. It is worth noting that our model does not rely on specific assumptions on the economic preference (i.e., model free). Instead, our model is an approximate identity for any linear market model under arbitrary identification of the economic preference. In the Appendix, we examine some well-known economic preferences studied in prior literature as examples to illustrate the role of high-order systematic risks in the cross section of stock returns A calibration: The risk premium implied by the S&P 500 index and its options To gauge the impact of the high order risk premiums, we estimate the market risk premium in each moment implied by the S&P 500 index (SPX) and SPX index options, following the model-free methodology proposed by Bakshi, Kapadia, and Madan (2003). Using the index option prices from the Option Price file, we follow the procedure of Chang, Christoffersen, and Jacobs (2013) to estimate the 30-day risk-neutral market moments for each day during the period from 1996 to We then average daily estimates from the index option prices to obtain the full sample risk-neutral market moments. The physical market moments are computed using the full sample logarithmic monthly SPX returns. Table 1 presents the estimates of physical market moments, risk-neutral market moments, and their differences. The estimation result shows a positive market mean risk premium ( Q ) of 0.359% per month and a positive market skewness risk premium ( Q ) of 0.049% per 10

13 month. In contrast, we find a negative market variance risk premium ( Q ) of 0.260% per month and a negative market kurtosis risk premium ( Q ) of 0.021% per month. 7 We then use the estimates of market moments to generate random samples and then perform the kernel smoothing density to estimate the physical density and the risk-neutral density. We estimate the pricing kernel by the ratio of the risk-neutral density to the physical density. As can be seen from Panel A of Figure 1, the risk-neutral density of the market return (β=1) is more volatile, more negatively skewed, and more fat-tailed than its corresponding physical density. Furthermore, we find that, as shown in Panel B of Figure 1, the implied pricing kernel for the market return (β=1) is U-shaped, consistent with the recent findings in the literature. 8 In particular, Christoffersen, Heston, and Jacobs (2013) show that their variance dependent pricing kernel which implies the quadratic preference of the market return can generate a U-shaped pricing kernel. Among the three pricing kernels in our examples discussed in the Appendix, the quadratic pricing kernel and the stochastic volatility pricing kernel are more likely to reconcile the shape of the implied pricing kernel demonstrated in Figure 1. To investigate how the systematic risk interacting with the pricing kernel affects the expected stock returns, we generate random samples from individual stocks based on,,, where, = 0.5, 1, 2, and 4 and, is restricted by the risk-neutral pricing relation (i.e., Q ). 9 As shown in Panel A of Figure 1, as, increases, both the physical density and the risk-neutral density become more dispersed. Moreover, as can be seen from Panel B of Figure 1, the pricing kernel puts more weight on the positive region as, increases. In other words, high beta assets have higher volatile future payoffs and therefore are 7 These estimated results suggest an estimated annual market risk premium of 2.9% during our sample period. 8 See, for example, Bakshi, Madan, and Panayotov (2010), Christoffersen, Heston, and Jacobs (2013), Chabi-Yo, Garcia, and Renault (2008), and Brown and Jackwerth (2012), among others. 9 We do not consider the idiosyncratic randomness here since it does not affect the expected returns in our model. 11

14 capable of earning the upside variance premium provided by the increasing region of the pricing kernel. Hence, high beta assets are more likely to have high prices and low expected returns, supporting the recent work of Bakshi, Madan, and Panayotov (2010) on the property of contingent claims on the upside. We calibrate our approximate capital asset-pricing model with the risk prices implied by SPX and SPX options and plot the cross-sectional expected stock returns in Panel A of Figure 2. As can be seen from the graph, the security market line (SML) is inversely U-shaped, suggesting that the first two risk prices are economically important. The effect of the interaction between the market variance risk premium and the market beta can be confirmed in Panel B of Figure 2. The figure suggests that the SML has a positive slope in the absence of the market variance risk premium, whereas the negative market variance risk premium leads to a concave SML. The link between the market variance risk premium and the curvature of the SML is a unique feature of our model Cross-sectional regressions: Pricing high order systematic risks We next examine our approximate capital asset pricing model in the cross section. In each month, we sort stocks into 25 portfolios based on the historical CAPM beta (, ) estimated from the past one month daily returns. 10 We then compute the equal-weighted portfolio returns. To achieve higher testing power, we also adopt the Fama-French 25 value-weighted portfolio returns formed on size and B/M. 11 We first estimate the following time-series regression for each portfolio on the Fama-French (1993) and Carhart (1997) four-factor model: 10 Although our model assumed in Eq. (5) is based on log returns, to be consistent with the literature we use raw returns to estimate the historical CAPM betas and the Fama and French factor loadings. Nevertheless, our results remain unchanged if we use log returns to estimate the market betas or factor loadings. 11 The detailed description of our data will be detailed discussed in Section 4. 12

15 ,,,,,,,. (8) In the second stage, we use the Fama-MacBeth (1973) cross-sectional regression to estimate the prices of high order risks while controlling for common factor loadings as follows:,,,,,,,, (9) where,,,, and, are the orthogonalized higher order market risks with respect to their lower order risks. We use the robust Newey and West (1987) t-statistics with eight lags that account for autocorrelations to test whether the estimated risk premiums are significant different from zero. Panel A of Table 2 reports the estimates for the prices of high order risks using equal-weighted portfolio returns formed on the historical CAPM beta. Consistent with the first-order market risk of classical CAPM, as reported in Model 1, we find that the first-order market risk premium (, ) is positive (0.711% per month) with a significant t-statistic of Furthermore, as reported in Model 2,, is negative ( 2.372) with a significant t-statistic of 5.69, consistent with the well-documented evidence for the negative market variance risk premium. In Model 4, we find that, has a significant positive value of 1.567% (t-stat = 5.40) and, yields a significant negative value of 2.116% (t-stat = 4.69). In contrast,, and, are insignificant. Overall, our model-free results suggest that only the first two order risks are important in pricing the cross section of stock returns. Panel B of Table 2 reports the estimates using Fama-French 25 value-weighted portfolio returns formed on size and B/M. The results in Panel B are similar to those reported in Panel A. For example, Model 4 shows that, is positive (0.925% per month) with a significant 13

16 t-statistic of 4.87 and, is negative ( 0.826%) with a significant t-statistic of 2.18, whereas, and, are insignificant. Overall, Table 2 provides supporting evidence for the pricing of the first two orders of market risks. The first-order risk is significantly and positively priced, while the second-order risk is significantly and negatively priced. More importantly, our findings imply that the cross-sectional relation between expected return and market beta should be inversely U-shaped, which constitutes the main idea of our empirical tests in the remaining part of the paper. 3. Empirical implications of the second-order risk The result in Table 2 shows that only the first two orders of market risks are significantly priced. As a result, in the remaining part of the paper, we only consider a simplified model of Eq. (7) with the first two order systematic risks as follows:,,,,,,, (10) where, 0,, 0, and, 0. We then discuss the empirical implications of the second-order risk from this simplified model, including the inverse U-shaped return-beta relation, the volatility-return relation, and the betting against beta anomaly as follows The shape of the security market line First of all, the slope of the security market line (SML) is,,,, 2,,. (11) In this case, there exists,,, such that,,, 0 when,, and 14

17 ,,, 0 when,,. The result implies that the expected excess stock return in the cross section is first increasing with, and then decreasing with,. Furthermore, the curvature of the security market line (which is the second derivative of the stock excess return with respect to the first-order risk,, ) is positively related the second-order market risk premium as follows:,,, 2,. (12) The first derivative of the cross-sectional risk-adjusted returns with respect to the second-order risk (i.e.,, ) is also positively associated with the second-order market risk premium:,,,,,. (13), In sum, the curvature of the SML or the first derivative of the risk-adjusted SML with respect to the second-order risk corresponds to the second-order risk price (, ). Since, < 0, the results imply that the SML is inversely U-shaped. The following corollary summarizes our results. Corollary 1. (Market variance risk premium and security market line) When the market variance risk premium is negative (i.e.,, 0), the slope of the security market line is decreasing in,. Furthermore, the security market line is inversely U-shaped and the curvature and the first derivative of the risk-adjusted security market line with respect to the second-order risk (, ) are both positively related to the market variance risk premium. 15

18 3.2. The cross-sectional volatility-return relation We now illustrate how the second-order risk price affects the cross-sectional volatility-return relation. First, define, as the stock return variance as follows:,,,,,,,,,. (14) In our case,, 0 and,,,, stock return with respect to, is, /. Then the first derivative of the cross-sectional,,,,, 2,,,,,. (15) Therefore, there exists, and,,,,,,, such that,,,, 0 when,,. The result yields the following corollary. 0 when,, Corollary 2. (The second-order risk price and the cross-sectional volatility-return relation) The expected excess stock return in the cross section is first increasing with, and then decreasing with,. In particular,, contributes to the negative cross-sectional volatility-return relation. We now show that the second-order risk price in our model can also help explain the cross-sectional return differentials with respect to idiosyncratic volatility documented by Ang, Hodrick, Xing, and Zhang (2006). Define the idiosyncratic volatility as the residual variance of the stock return adjusted for the first-order risk premium, 16

19 , ar,,,, ar,. (16) In our case,, 0 and,,, /. The first derivative of the cross-sectional stock return with respect to, is,,,, 2, ar,, 4, ar,. (17) There exists,,, such that,, 0 when,,, and,,,, 0 when,,. The result yields the following corollary. Corollary 3. (The second-order risk price and idiosyncratic volatility) The expected excess stock return in the cross section is first increasing with, and then decreasing with,. In particular,, contributes to the negative cross-sectional idiosyncratic volatility-return relation The betting against beta anomaly Recently, Frazzini and Pedersen (2014) suggest that because constrained investors bid up high beta assets, high beta assets are associated with low risk-adjusted returns (alphas). They show that a betting against beta (BAB) strategy that longs leveraged low beta assets and shorts high beta assets generate significant positive risk-adjusted returns. Our model, in contrast, suggests that the betting against beta strategy is exposed to high order systematic risks. To illustrate our claim, define the return on the BAB strategy as 17

20 ,,,,,, (18) where,, is the return for the low beta assets, and, is the return for the high beta assets. In our model, it follows directly that,,,, 0.,, (19) Thus, our model implies that the second-order risk price contributes to the premium of the BAB strategy. The result yields the following corollary. Corollary 4. (The second-order risk price and the betting against beta premium) The BAB factor is negatively related to,. 4. Data and summary statistics 4.1. Data Our sample comprises all NYSE/AMEX/NASDAQ ordinary common stocks over the period from January 1963 to December Daily and monthly stock return data (with share code =10 and 11) are from the Center for Research in Security Prices (CRSP). Stocks with share prices less than $1 at the end of the previous month are excluded. Financial statement data are from COMPUSTAT. Fama and French (1993) factors, their momentum UMD factor, and their Size-B/M portfolios are obtained from the online data library of Ken French. 12 We obtain daily data from OptionMetrics for equity options and S&P 500 index options over the period from

21 January 1996 to December The expected market variance risk premium (ERV IV) is obtained from Hao Zhou s personal website. 13 The risk-neutral expectation of variance (IV) is measured as the end-of-month VIX-squared de-annualized (VIX 2 /12), whereas the realized variance (RV) is the sum of squared 5-minute log returns of the S&P 500 index over the month. As described in Drechsler and Yaron (2011) and Zhou, the expected realized variance (ERV) is a statistical forecast of realized variance with one lag of implied variance and one lag of realized variance. As a result, the expected market variance risk premium is defined as EVRP= ERV IV. Data on government bond yields, corporate bond yields, and the TED spread are from the FRED database of the Federal Reserve Bank of St. Louis. Following the literature (e.g., Petkova and Zhang, 2005; Petkova, 2006), we construct a set of variables for macro-economy. Specifically, we use the CRSP value-weighted portfolio to measure the dividend yield (DIV) by the sum of dividends over the last 12 months, divided by the level of the index. The term spread (TERM) is measured by the difference between the yields of the 10-year and 1-year government bonds. The default spread (DEF) is computed as the difference between the yields of long-term corporate Baa bonds and long-term government bonds. We use stock analyst forecasts of the long-term growth rate (LTG) for earnings per share (EPS) obtained from the unadjusted I/B/E/S summary database. Following Yu (2011) and Hong and Sraer (2012), the standard deviation of LTG forecasts is used to proxy for the firm-level disagreement. The aggregate disagreement (DIS) is measured as the cross-sectional value-weighted average of the individual stock disagreements

22 4.2. Estimation of co-moment risks Our model-free CAPM in Eq. (7) is very general. In the remaining part of the paper, we focus on the model in Eq. (10) with only the first-order and second-order risk premiums. If this special case of the model is supported, it suggests that our more general case of the model-free CAPM in Eq. (7) is also supported for the nonlinear risk-return trade-off. In contrast, if the special case of the model is rejected, we cannot make any conclusion whether our more general case of the model-free CAPM is accepted or rejected. To test our model, we use the first-order co-moment risks to identify the first-order risk and, is measured by,,,,,,,,,, or,,,, ; similarly, we use the second-order co-moment risks to identify the second-order risk and, is measured by,,,, or,,,,. 14 In many cases, through the specification of the pricing kernel as exemplified in the Appendix, the risk prices for these co-moment risks are determined by high-order market moments with the following properties:,,,,,,,,,,,,,,,,,,,,,,,,,. While the risk prices for the co-moment risks are preference-dependent, they are nested in our model-free CAPM in Eq. (7), i.e.,,,,,,,,,,, and,,,,,,,. As a result, these co-moment risks provide the channel through which the systematic risks can be measured and we will focus on testing the pricing effects for, and,, which are in general not sensitive to any specification of utility preferences. 14 We need the condition that,0. 20

23 To do so, in each month we estimate co-moment risks using the daily stock returns over the past month. Following Ang, Hodrick, Xing, and Zhang (2006), only stocks with more than 17 daily observations are included. First of all, the historical CAPM beta (, ) is estimated by the following linear regression:,,,,. (20) Define, as stock i residual return, i.e.,,,,, ) and define, as the demeaned market return, i.e.,,. Then the first-order coskewness (, ) and first-order cokurtosis (, ) are computed respectively by,,,. (21) Our approach is similar to but different from that of Harvey and Siddique (2000). 15 In particular, our estimation for these first-order systematic risks is performed separately as follows. We estimate the co-moments divided by market variance using the daily data from the regression whereas the market skewness ( ) and market kurtosis ( ) are computed using the full sample monthly market returns. While preserving the cross-sectional ranks of the security betas, this procedure ensures that the denominators estimated from the short regression window would be well-behaved. The second-order coskewness (, ) and second-order cokurtosis (, ) are similarly estimated by,,,, (22) where the absolute value is required in both estimates since the second-order systematic risk 15 While the numerator of, and that of, in Harvey and Siddique (2000) are the same, these two co-moment risks are different in which the residual volatility of stock returns is used as the denominator of their,. 21

24 should be nonnegative. Moreover, to identify the second-order risk in a model-free fashion, we use the second moment condition in the risk-neutral world through the model-free implied variances estimated from individual options and index options, i.e., Q,, Q,,. We first apply the model-free approach of Bakshi, Kapadia, and Madan (2003) to estimate the 30-day risk-neutral moments for each day. 16 For each month, we then estimate the risk-neutral variance beta (,, ) with the daily estimates of risk-neutral individual variance and risk-neutral market variance, exploiting the linear market model relation for the second moment as follows:,,,,,,,,,,, (23) where,, is the risk-neutral variance of stock i and,, is the risk-neutral variance of the market. The parameters are estimated using nonnegative least squares method, since both,, and,, should be nonnegative. Unlike the co-moment risks that are preference-dependent, the estimate of risk-neutral variance beta (,, ), which essentially corresponds to the second-order risk, has a model-free asset pricing implication Summary statistics Table 3 reports summary statistics for variables used in this study. In each month, we first compute the cross-sectional statistics for each security and then report the time-series average. The historical CAPM beta (, ) ranges from (at the 5 th percentile) to (at the 95 th percentile); the first-order coskewness (, ) ranges from to 4.280; the first-order kurtosis (, ) ranges from to The mean and median values of these first-order 16 In particular, we follow the procedure outlined in Chang, Christoffersen, and Jacobs (2013) except for one thing. We use linearly interpolated implied volatilities since the cubic spline interpolation requires more available observations across moneyness and sometime produces inconsistent negative estimates for implied volatilities. 22

25 co-moment risks tend to decrease as the orders of the market returns increase. The second-order coskewness (, ), ranging from to , is more dispersed than the second-order kurtosis (, ) which ranges from to The mean and median values of, are larger than those of,. The risk-neutral variance betas (,, ), estimated from stocks with available equity options, range from to 8.675, which is less dispersed than those of the second-order co-moment risks estimated from the entire common stock sample. We also report the statistics for average daily estimates for risk-neutral variance (, ), risk-neutral skewness (, ), and risk-neutral kurtosis (, ) as well as statistics for firm characteristics, including the book-to-market equity ratio (B/M), market capitalization (Size, in billion), average past 11-month returns prior to last month (RET_2_12), and Amihud s illiquidity measure (ILLIQ, in million). 5. Evidence on the second-order risk premium As shown in Corollary 1, the second-order risk premium (, ) is positively associated the curvature of the SML (i.e., Eq. (12), which is the second derivative of the SML with respect to the first-order risk (, ). Since,,,,,,,,,,,,,,,,,, we use, in Eq. (20) and, and, in Eq. (21) to proxy for, in our empirical tests. We discuss the results based on these first-order systematic risks in Section 5.1. Corollary 1 further shows that the second-order risk premium (, ) is also positively associated the first derivative of the risk-adjusted SML with respect to the second-order risk (, ). Furthermore, since,,,,,,,,,,,,, we therefore use 23

26 , and, in Eq. (22) to proxy for, in our empirical tests and report the results in Section Performance of portfolios formed on the first-order systematic risks We examine the performance of portfolios formed on the first-order co-moment risks, including the market beta (, ), the first-order coskewness (, ) and the first-order cokurtosis (, ). In each month, all stocks are sorted into 25 portfolios from the lowest (1) to the highest (25). After portfolio formation, we calculate the equal-weighted monthly stock returns for each portfolio. 17 For each portfolio, we compute the risk-adjusted return with respect to Fama-French (1993) and Carhart (1997) four-factor model (intercept) estimated from a time-series regression. Table 4 and Figure 3 present the results for,,,, and, in Panels A, B, and C, respectively. Consistent with an inverse U-shaped pattern implied by the model, the results in Table 4 and Figure 3 find that all first-order co-moment risk sorted portfolios exhibit a pattern that stocks in the bottom and top portfolios have lower stock returns than stocks in the middle portfolios. To exploit the economic value of the inverse U-shaped pattern, we construct a curvature portfolio (CUR1) by the sum of the twice difference for the portfolios formed on each of the first-order co-moment risks, which is equivalent to Eq. (12). That is, for N risk-sorted portfolios, the curvature portfolio is defined as Δ, where Δ and Δ Δ Δ. That is, for the 25 portfolios, CUR1 is the curvature portfolio that longs the difference in the top two portfolios (25-24) and shorts the difference in the bottom two portfolios (2-1). 17 The results in Tables 4 and 5 (to be discussed in the next subsection) are re-evaluated based on value-weighted returns and the findings remain unchanged. 24

27 We find that CUR1 is significantly negative at 1.00% per month (t-stat = 5.79) for, sorted portfolios, at 0.98% (t-stat = 5.62) for, sorted portfolios, and at 0.96% (t-stat = 6.18) for, sorted portfolios. Controlling for the Fama-French (1993) and Carhart (1997) four-factor model, CUR1 still generates a significant alpha of 1.11% with a t-statistic of 7.15 for, sorted portfolios, 1.01% with a t-statistic of 6.83 for, sorted portfolios, and 1.03% with a t-statistic of 7.78 for, sorted portfolios. In summary, consistent with Corollary 1, we find an inverse U-shaped pattern for portfolios formed on the first-order risks. We also find that the curvature portfolios based on the trading strategy exploiting the inverse U-shaped pattern or the second derivative of the SML generates significant abnormal returns, of which the Fama-French (1993) and Carhart (1997) four-factor model cannot explain. The findings suggest that the second-order risk premium is negative and is statistically and economically significant Performance of portfolios formed on the second-order co-moment risks We examine the performance of portfolios formed on the second-order co-moment risks, including the second-order coskewness (, ) and second-order cokurtosis (, ). In each month, all stocks are sorted into 25 portfolios based on, or, from the lowest (1) to the highest (25) and the portfolio returns are equal-weighted. Table 5 and Figure 4 present the results for, and, in Panels A and B, respectively. Consistent with the negative risk premium for the second-order risk (Corollary 1), Table 5 and Figure 4 show that portfolio returns exhibit a decreasing pattern in, or,, albeit slightly increasing initially. To exploit the economic value of the second-order risk premium, we 25

28 construct another curvature portfolio (CUR2) that longs the top portfolio and shorts the bottom portfolio formed on each of the second order co-moments risks, which is equivalent to Eq. (13). We find that CUR2 is significantly negative at 1.27% per month (t-stat = 4.32) for, sorted portfolios and at 1.06% (t-stat = 3.05) for, sorted portfolios. Controlling for the Fama-French (1993) and Carhart (1997) four-factor model, CUR2 delivers an even more negative and significant alpha of 1.47% with a t-statistic of 8.37 for, sorted portfolios, and 1.37% with a t-statistic of 6.74 for, sorted portfolios. In Panel C, we find similar results for the 15 portfolios formed on the risk-neutral variance beta (,, ). The curvature portfolio (CUR2RN = 15-1) is significantly negative at 1.34% with a t-statistic of 2.21 and has a significant alpha of 1.38% with a t-statistic of 3.71 adjusted for the Fama-French (1993) and Carhart (1997) four-factor model. In summary, consistent with Corollary 1, we find a negative cross-sectional relation between stock returns and second-order risks. We also find that the curvature portfolios generate significant abnormal returns, of which the Fama-French (1993) and Carhart (1997) four-factor model cannot explain. The findings confirm that the second-order risk premium is negative and is statistically and economically significant. 6. Mimicking curvature factors 6.1. Properties of the mimicking curvature factors We construct three mimicking factors for the second-order risk premium using the curvature portfolios studied in the previous section. We construct our first mimicking factor, FCUR1, based on the average of the three CUR1s formed on,,,, and,. Similarly, we construct our second mimicking curvature factor, FCUR2, based on the average of the two CUR2s formed on 26

29 , and,. Our third mimicking curvature factor, FCUR2RN, is the curvature portfolio CUR2RN formed on,,. 18 Table 6 reports the performance of our mimicking curvature factors. In Panel B, FCUR1 is significantly negative at 0.98% per month (t-stat = 6.90) and FCUR2 is also significantly negative at 1.17% (t-stat = 3.66) during the sample period from January 1963 to December Moreover, FCUR1 and FCUR2 remain significantly negative during the sub-sample period from January 1996 to December For each curvature factor, we compute the risk-adjusted return with respect to Fama-French (1993) and Carhart (1997) four-factor model. Our three mimicking curvature factors, FCUR1, FCUR2, and FCUR2RN, have significantly negative abnormal returns of 1.05% (t-stat = 6.90), 1.42% (t-stat = 7.65), and 1.38% (t-stat = 3.71), respectively. Panel C of Table 6 presents the Spearman correlations. Our curvature factors are related to each other. For example, the correlation between FCUR1 and FCUR2 is high at and the correlation between FCUR2 and FCUR2RN is also high at The market factor (MKT) has positive correlations of with FCUR1, with FCUR2, and with FCUR2RN. Furthermore, the size factor (SMB) also shows positive correlations of with FCUR1, with FCUR2 and with FCUR2RN. The value factor (HML) and the momentum factor (UMD) show smaller negative correlations with our mimicking factors. More importantly, consistent with our model, we find that the expected market variance risk premium (ERV IV) has positive correlations of with FCUR1, with FCUR2, and with FCUR2RN. 18 In untabulated results (which are available upon request), we find that the performance of our curvature mimicking portfolios are robust to alternative total numbers of portfolios formed (10 or 15 for FCUR1 and FCUR2 and 10 for FCUR2 RN). 27

30 6.2. The market variance risk premium and the mimicking curvature factors To examine how our mimicking curvature factors are related to the macroeconomic variables, we regress these factors on a set of state variables. Following the literature, we use the aggregate dividend yield (DIV), the default spread (DEF), the term spread (TERM), and one-month Treasury bill yield (TB) as the explanatory variables. Moreover, we also include the expected market variance risk premium (ERV IV), aggregate disagreement (DIS), and the TED spread (TED). Our model implies that the second-order risk premium is related to the market variance premium and therefore ERV IV should explain our mimicking curvature factors. Table 7 presents the results for the regression results of the mimicking factors. Consistent with the prediction of our model, Model 1 in Panel A shows that ERV IV has a significant and positive slope of with a t-statistic of 2.19 in explaining the variation of FCUR1, whereas all other explanatory variables are insignificant. Controlling for DIS and TED, Model 2 in Panel A reveals that ERV IV remains significant with a t-statistic of Moreover, Model 2 of Panel B and Model 2 of Panel C report that ERV IV has significantly positive slopes with the t-statistics of 2.47 and 2.69 in the regressions of FCUR2 and FCUR2RN, respectively. The time-series co-movement between FCUR1 (FCUR2) and the expected market variance risk premium (ERV IV) is also confirmed in Panel A (Panel B) of Figure 5. Hong and Sraer (2012) demonstrate that the disagreement about the market return leads to speculative overpricing for high beta stocks. Their model implies that the shape of the SML is kinked and the slope of the SML decreases with the aggregate disagreement. Their theory implies that the aggregate disagreement should explain our mimicking factors. However, Table 7 shows that the market variance premium explains the curvature factors better than does the aggregate disagreement. Overall, the second-order risk premium in our study cannot be fully 28