Investor Preferences and Portfolio Selection: Is Diversification an Appropriate Strategy? C. James Hueng 1 and Ruey Yau 2 1. Department of Economics, Western Michigan University, Kalamazoo, MI 49008, U.S.A. 2. Department of Economics, National Central University, Taoyuan, Taiwan 32001, R.O.C. Email: James.Hueng@wmich.edu and ryau@mgt.ncu.edu.tw Abstract: This paper analyzes the relationship between diversification and several distributional characteristics that have risk implications for stock returns. We develop a flexible three-parameter distribution to model the stock returns. Using data on the current 30 DJIA stocks, we show that an investor’s strategy on diversification depends on the measures of risk for particular concerns. For example, investors who desire to increase positive skewness would hold a less diversified portfolio, while those who care more about extreme losses would hold a more diversified portfolio. Experimenting with a more general pool of stocks yields the same conclusions. JEL Classification: C51, G11, G12. Keywords: diversification; asymmetric generalized t distribution; skewness. 1. Introduction Research conducted in the mean-variance framework has long shown that diversification offers the benefit of reducing unsystematic risk (e.g., Sharpe, 1964, Lintner, 1965, Evans and Archer, 1968, and Fielitz, 1974). However, empirical evidence has revealed that investors do not tend to hold fully diversified portfolios. For example, Goetzmann and Kumar (2002) examine more than 40,000 equity investment accounts over a six-year period from 1991 to 1996 and find that the average investor has four stocks in her/his portfolio, while less than five percent of investors hold more than ten stocks in their portfolios. Furthermore, more than 25 percent of portfolios contain only one stock, and more than 50 percent contain fewer than three. Given the benefits typically associated with holding a diversified portfolio, the apparent lack of diversification among most investors is somewhat puzzling. The perception that investors do not weigh downside risk equally with upside potential provides an explanation to this phenomenon. In this case, the variance is no longer an appropriate measure for risk. Kraus and Litzenberger (1976) build a three-factor CAPM framework by explicitly incorporating skewness in investors’ preference. Malevergne and Sornette (2005) derive a modified efficient frontier where higher moments (up to the eighth order) replace variance as the measure of risk. They demonstrate how the traditional portfolio optimization framework can be improved. Cooley (1977) conducts an experiment to test the perception of risk on the part of institutional investors and finds that, among 56 institutional investors who were asked to rate distributions according to perceived risk, at least 29 associated the asymmetry of return distributions with risk. In particular, the investors associated increases in risk with increases in negative skewness, indicating a preference for positive skewness. 1 Harvey and Siddique (2000) and 1 Arrow (1974) theoretically shows that risk averse investors with non-increasing risk aversion prefer positively-skewed investment positions. 1 Premaratne and Tay (2002) explore the role of skewness in asset returns. If an asset contributes positive skewness to a diversified portfolio, then that asset will be valuable and has a higher price (or lower expected return). Investors who prefer positive skewness would seek to construct portfolios that have this characteristic. However, this complicates the portfolio selection decision, because a desire on the part of investors to obtain positive skewness may not be compatible with the familiar method of constructing a diversified portfolio in order to reduce risk. As shown in Simkowitz and Beedles (1978), even though portfolio variance decreases as diversification occurs, skewness may either increase or decrease with diversification. Based on computations of the raw skewness, they find that skewness is rapidly reduced by diversification for a pool of 549 common stocks. Aggarwal and Aggarwal (1993) and Cromwell, Taylor, and Yoder (2000) also record similar results. Aside from variance and skewness, there exist other risk measurements that are of particular concern to specific investors. For example, recent research on Value at Risk (VaR) indicates that investors may be concerned strongly about the likely maximum loss, i.e. the left tail of the distribution of a portfolio’s expected returns. Of particular interest to us is the relationship between diversification and the other information contained in the distribution beyond variance and skewness. For example, investors may have a stronger preference for a lower expected extremeloss than for a higher skewness. As such, the relationship between expected extreme losses and diversification adds another dimension to the portfolio selection process. Indeed, Gaivoronski and Pflug (2000) conclude that investors who are concerned with VaR will not achieve comparable results using a portfolio selection methodology that relies on another risk measure such as variance; rather, VaR is a very different measure that deserves its own place in the portfolio selection process. To achieve our stated goals, we propose a parametric model to estimate the population distribution of stock returns. It is well known that stock returns are not normally distributed. 2 Therefore, to model the higher moments in stock returns, we develop the “Asymmetric Generalized t (AGT) distribution,” an asymmetric version of McDonald and Newey’s (1988) generalized t (GT) distribution. Although this is a simple three-parameter distribution, it is very flexible in that it permits very diverse levels of higher moments.2 This distribution nests several popular distributions often seen in the literature, including the normal and the Student’s t distributions. This general distribution allows us to estimate the population distribution of the returns and provides us with information on a wide range of distributional characteristics that have risk implications. We focus on the current 30 stocks in the Dow Jones Industrial Average (DJIA) for both computational tractability and its desired nature of a well-diversified portfolio. The results show that first, consistent with the previous studies using sample moments, our results show a trade-off between low variance and high skewness: diversification reduces the portfolio variance, but at the same time also reduces the skewness. Second, expected extreme losses become smaller when the portfolio size increases. As such, an investor’s strategy of diversification depends on the measures of risk for particular concerns. For example, investors who want to increase positive skewness would hold a less diversified portfolio, while those who care more about extreme losses would hold a more diversified portfolio. 2 Other flexible alternatives include the exponential generalized beta of the second kind (EGB2), transformations of normally-distributed variables discussed by Johnson (1949), and a family of modified Weibull distributions proposed by Malevergne and Sornette (2004). Wang et al. (2001) apply the EGB2 to GARCH models. The AGT distribution proposed in this paper is at least as equally flexible as the EGB2 distribution. Whereas the EGB2 distribution imposes limited ranges on higher moments, the AGT distribution has no such limits. The modified Weibull distribution is characterized by only two parameters, but it does not nest the Student’s t distribution, the most often used statistical distribution to capture the fat-tail behaviors in asset returns. 3 The next section outlines the model and the properties of the AGT distribution. Section 3 presents the empirical results by using the DJIA component stocks. Section 4 experiments with a more general pool of stocks to test the robustness of our conclusions. Section 5 offers the conclusion. 2. The Model and Methodology The conditional mean of the stock returns is modeled as a simple AR(m) process in an attempt to estimate the zero-mean, serially-uncorrelated residuals: yt = μ + m ∑φ y i t −i + εt , (1) i =1 where yt represents the daily stock returns and lag length m is chosen by the Ljung-Box Q tests as the minimum lag that renders serially-uncorrelated residuals (at the 5% significance level) up to 30 lags from an OLS regression. Note that the OLS regression is used only to determine the number of lagged returns to be included in the conditional mean. The coefficients in the conditional mean equation will be jointly estimated with the conditional variance equation using the full information maximum-likelihood (FIML) estimation. The conditional variance, denoted as ht ≡ E (ε t2 | t − 1) , follows a GARCH process. It is well documented in the finance literature that stock returns have asymmetric effects on predictable volatilities; see among others, Glosten, Jaganathan, and Runkle (1993) and Bekaert and Wu (2000). Therefore, we use an asymmetric GARCH model proposed by Glosten, Jaganathan, and Runkle (the GJR model), which is claimed to be the best parametric model among a wide range of predictable volatility models experimented by Engle and Ng (1993): ht = κ + α ⋅ ht −1 + β ⋅ ε t2−1 + γ ⋅ I t+−1 ⋅ ε t2−1 , where I t+−1 = 1 if ε t −1 >0 and I t+−1 = 0 otherwise. 4 (2) The relationship between the return residual and its conditional variance can be specified as ε t = ht vt , where vt is a zero-mean and unit-variance random variable. The specification of the distribution of the stochastic process {vt} determines the distribution of yt. The most commonlyused distribution for vt is the standard normal distribution, which is symmetric and has a kurtosis coefficient of three. While the GARCH specification makes some allowance for unconditional excess kurtosis, according to Bollerslev, Chou, and Kroner (1992), it is still unable to adequately model the fat-tailed properties of the stock returns. In response to the levels of kurtosis found in stock return data, Bollerslev (1987) combines a t-distribution and a GARCH (1,1) model. The tdistribution has fatter tails than are found in the normal distribution. McDonald and Newey (1988) propose the Generalized-t (GT) distribution, which is a more general distribution that can accommodate both leptokurtosis (thicker tailed than the normal distribution) and platykurtosis (thinner tailed than the normal distribution). The pdf of the GT distribution is: p f GT ( xt ; ω , p, q ) = p xt ⎞⎟ ⎛ 1 ⎞⎛⎜ 2ωq B ⎜⎜ , q ⎟⎟ 1 + qω p ⎟ ⎝ p ⎠⎜⎝ ⎠ 1 p q+ 1 p , where B(.) is the beta function. The parameters p and q are positive and p×q>n in order for the nth moment to exist. The mean is zero and the variance is ω 2 q 2 / p B (3 / p, q − 2 / p ) / B(1 / p, q ) . The kurtosis is [ B (1 / p, q ) / B (3 / p, q − 2 / p )]2 [ B (5 / p, q − 4 / p ) / B(1 / p, q)] with p×q>4. Special cases of the GT include the power exponential or Box-Tiao (BT) (q → ∞ ), Student’s t (p=2), normal (p=2 5 and q → ∞ ), and Laplace (p=1 and q → ∞ ). The apparent advantage of the GT distribution over the t-distribution is its flexibility in modeling higher moments.3 While the GT distribution is more flexible than the popular normal and student’s t distributions in modeling the fourth moment, it is still symmetric and unable to model the skewness in stock returns. A distribution with the ability to capture all of the first four moments could provide more flexibility for modeling returns. To obtain such a distribution, we develop an asymmetric version of the generalized t distribution, where we define: ⎧ ⎛ zt ⎞ ⎟ ⎪ f GT ⎜⎜ ⎟ ⎪⎪ ⎝1+ r ⎠ h( zt ) = ⎨ ⎪ ⎛ zt ⎞ ⎟ ⎪ f GT ⎜⎜ ⎟ ⎪⎩ ⎝1− r ⎠ for zt ≥ 0 , for zt < 0 , where -1<r<1. 4 This specification is a proper pdf and allows different rates of descent for zt>0 and zt<0, and therefore, it allows for skewness. Next, we scale zt by defining xt = zt / A, where A is a positive scaling constant for simplifying the notation. The density function of the asymmetric generalized t (AGT) distribution can be written as: 3 The GT distribution does not restrict the level of kurtosis. In addition, it can be shown that the square of the skewness is less than one plus the kurtosis [(SK)2 < KU + 1]. Therefore, a wider range for kurtosis also allows a wider range for skewness. 4 Hansen (1994) uses the same technique to develop an asymmetric t-distribution. In a different framework, Theodossiou (1998) also develops a skewed version of the GT distribution, but his distribution has four parameters and a more complicated pdf. 6 ⎧ ⎛ A ⋅ xt ⎪ f GT ⎜⎜ ⎪⎪ ⎝ 1+ r ∂z t = h( A ⋅ xt ) ⋅ A = ⎨ f AGT ( xt ) = h( z t ) ⋅ ∂xt ⎪ ⎛ A ⋅ xt ⎪ f GT ⎜⎜ ⎪⎩ ⎝ 1− r ⎞ ⎟⋅ A ⎟ ⎠ for xt ≥ 0 , ⎞ ⎟⋅ A ⎟ ⎠ for xt < 0 . Specifically, let A = [ p / 2ωq1/ p B(1 / p, q)]−1 , and the pdf becomes: ⎧ 1 ⎪⎧ ⎛ ⎪ ⎪⎪ ⎜ 2 B( p , q ) ⋅ x t ⎪ ⎨1 + ⎜ ⎪ ⎪ ⎜ p (1 + r ) ⎪ ⎪ ⎜⎝ ⎪⎩ f AGT ( x t ; p, q, r ) = ⎨ ⎪ 1 ⎪⎧ ⎛ ⎪ ⎪⎪ ⎜ 2 B( p , q) ⋅ xt ⎪ ⎨1 + ⎜ ⎪ ⎪ ⎜ p(1 − r ) ⎪ ⎪ ⎜⎝ ⎩⎩ ⎞ ⎟ ⎟ ⎟ ⎟ ⎠ p ⎞ ⎟ ⎟ ⎟ ⎟ ⎠ p ⎫ ⎪ ⎪ ⎬ ⎪ ⎪⎭ −q− ⎫ ⎪ ⎪ ⎬ ⎪ ⎪⎭ −q− 1 p for xt ≥ 0 , for xt < 0. 1 p As is apparent, the AGT distribution nests the GT, BT, Laplace, t, and normal distributions. (See the Technical Appendix for more details on this distribution.) Like in the GT distribution, p and q control the height and tails of the density. The additional parameter r controls the rate of descent of the density around x=0. Specifically, when r>0, the mode of the density is to the left of zero and the distribution skews to the right, and vice versa when r<0. When r=0, the distribution is symmetric. Figure 1 plots this density for several different parameterizations. Let vt = (xt-μ)/σ. The standard AGT distribution with zero mean and unit variance has the following probability density function: 7 1 ⎧ p −q− p ⎫ ⎧ 1 ⎪ ⎞ ⎛ ⎪ ⎪⎪ ⎜ 2 B ( p , q ) ⋅ σ ⋅ vt + μ ⎟ ⎪⎪ ⎟ ⎬ ⎪σ ⎨1 + ⎜ ⎟ ⎪ ⎜ p ( 1 r ) + ⎪ ⎪ ⎟ ⎪ ⎪ ⎪ ⎜⎝ ⎠ ⎭ ⎪ ⎩ f SAGT ( vt ; p, q, r ) = ⎨ 1 ⎪ p −q− p ⎫ ⎧ ⎪ 1 ⎞ ⎛ ⎪ ⎪ ⎜ 2 B( , q) ⋅ σ ⋅ vt + μ ⎟ ⎪ p ⎟ ⎪⎬ ⎪σ ⎪⎨1 + ⎜ ⎟ ⎪ ⎜ p(1 − r ) ⎪ ⎪ ⎟ ⎪ ⎪ ⎪ ⎜⎝ ⎠ ⎭ ⎩ ⎩ for σ ⋅ vt + μ ≥ 0 , (3) for σ ⋅ vt + μ < 0 . The conditional mean equation (1) and the GJR model (2) with the conditional distribution implied by (3) are jointly estimated by the full information maximum-likelihood (FIML) method. When estimating the model, we restrict p×q>2 for the variance to exist. 5 The likelihood function is obtained as: g AGT (ε t | ht ) = f SAGT ( v t ) × ∂v t ε 1 . = f SAGT ( t ) × ∂ε t ht ht 3. Empirical Results Data and Model Estimation 5 We use a logistic transformation θ = L + (U − L) to set constraints on the parameters. With 1 + exp(−ω ) this transformation, even if the parameter ω is allowed to vary over the entire real line, θ will be constrained to lie in the region [L, U]. Specifically, p and q are restricted to be between 0 and 50 and r is between -1 and 1. 8 We focus on the current 30 components of the DJIA for both computational tractability and the prominence of the index in the investment arena. 6 The DJIA is very widely quoted in investment news and is among the most scrutinized indicators of U.S. stock market performance. It includes a wide variety of industries and is composed of blue chip stocks that are typically industry leaders.7 Many investors choose to invest in these stocks in order to achieve diversification and long-term growth. We collect daily return data from the CRSP database from January 1990 to December 2002, for a total of 3280 observations for each stock. Panel A of Table 1 reports the sample mean and standard deviation of individual daily stock returns (in percentage). The average daily return ranges from 0.038% to 0.143%. The daily standard deviations are generally high, ranging from 1.457% to 2.915%. Panel B reports the estimated values of several key parameters in the model. In the conditional variance equation, as expected, the estimated α’s and β’s are statistically significant at the 5% level and show that the volatilities of the returns are highly persistent for all 30 stocks. In addition, the estimated GJR parameters ( γ ’s) are all negative and most of them are statistically significant, showing that the 6 Note that we are using the 30 stocks that are currently the components of the Dow Jones Index. Some of them were not in the Index in the earlier part of the sample. As of the date this paper is written, the latest change of the components in the DJIA was on April 8, 2004. Table 1 includes a list of the stock symbols. 7 The industries represented by the DJIA include materials, electronics, food/beverages/tobacco, financial services, aviation/aerospace, heavy equipment, chemicals, petroleum, automobiles, retail, computer hardware/software/services, pharmaceuticals, household supplies, telecommunications, and entertainment. 9 return shocks have asymmetric effects on predictable variance. Negative shocks tend to cause more volatilities than positive shocks do. Most of the estimated distributional parameters are significantly different from zero at the 5% level. Of particular interest is the estimate of r. Recall that r governs the asymmetry of the distribution. When r>0, the distribution is positively skewed. When r<0, the distribution is negatively skewed. All the 30 estimated r’s are positive, and 22 of them are significantly different from zero at the 5% level. To test the fit of the model, we conduct Newey’s (1985) GMM specification test by using orthogonality conditions implied by correct specifications. Correct specifications require that the standard AGT-distributed vt be i.i.d. and have zero mean and unit variance. Therefore, we test the moment conditions E(vt)=0 and E( vt2 )=1, as well as the serial correlation in vt at lags one through four. Panel C of Table 2 reports the t-statistics for the six selected orthogonality conditions, along with the chi-square statistic for the joint test. The chi-square test shows that our model fits very well, with all but one (stock code JPM) of the statistics being statistically insignificant at the conventional significance level. Considered individually, almost all of the six orthogonality conditions are insignificantly different from zero at the 5% level, with only four exceptions. Diversification and Risk Measures To investigate how the behaviors of portfolios’ distributional characteristics vary with increasing diversification, we construct portfolios containing different numbers of stocks in the following manner. First, a stock is randomly selected from the pool of the 30 DJIA stocks and the returns from this one-stock portfolio are used to estimate the model. Another stock randomly chosen from the remaining 29 stocks is then added to form a two-stock portfolio. The arithmetic average of the returns of these two stocks is used to estimate the model. We keep adding one stock 10 at a time to form 30 portfolios with sizes ranging from one to 30, where the portfolio with size n is a subset of the portfolio with size n+1. Several distributional characteristics with risk implications are next calculated from the estimations: the unconditional variance, the skewness parameter r, and the 5% and 95% quantiles.8 Since we do not assume the existence of higher moments beyond the variance, the usual measures of skewness and kurtosis may not exist. However, our parametric model provides all the information needed for the distribution. The parameter r measures the asymmetry of the distribution directly, and the 5% and 95% quantiles show the thickness of the tails. These risk measures are plotted against the portfolio sizes. We call these curves the “diversification structures of risk.” By doing this we are able to observe how a risk measure varies with the portfolio size (the degree of diversification). For example, if the plot is downward sloping, then the risk measure is decreasing with diversification. The whole process is repeated 100 times to yield 100 simulated samples. Therefore, there are 100 simulated curves for each risk measure. Note that the results for the 30-stock portfolio are identical across these 100 simulations. These curves are plotted in Figures 2(a)–2(d), which show a similar pattern: the absolute value of the slope is very big when the portfolio size is small, quickly declines as the portfolio size is approaching ten, and stays almost unchanged afterward. 8 There are two approaches to calculate the distributional characteristics of risks for a portfolio’s returns. The first method is to apply estimation schemes to the portfolio returns directly. This is the method employed in this paper and in Campbell et al. (2001). The second approach is to compute the distributional characteristics for the portfolio returns from a multivariate model, in which the dependence structure among individual stock returns needs to be identified beforehand. Examples of this are Guidolin and Timmermann (2006) and Malevergne and Sornette (2004). 11 The risk measures apparently are all non-linearly associated with portfolio sizes. Therefore, to model the diversification structure of risk, we propose the following specification: Ri ,n = ai + bi ⋅ 1 n + ci ⋅ + ei ,n , exp(n ) exp( n ) (4) where R is one of the four risk measures, i = 1,…,100 is the sample index, and n = 1, … , 30 is the portfolio size.9 According to this specification, the risk measure has an asymptote a. The second term 1 n decays rapidly as n increases. The third term also decays rapidly, but not as exp( n ) exp(n ) fast as the second term. Hence, the speed of decay depends on the relative magnitudes of the estimated b and c. This specification allows monotonic, humped, or S shapes, depending on the values of b and c. According to this structure, the “diversifiable” risk is b+c , and exp(1) b+c ⎡ 1 n ⎤ − ⎢b ⋅ + c⋅ shows the diversified risk by holding n stocks. exp(1) ⎣ exp( n ) exp( n ) ⎥⎦ In the first four columns of Panel (A) in Table 2, we report (for each risk measure) the average R-squared and estimated coefficients over 100 samples. The model fits quite well as the average R-squared’s are high and almost all the average coefficients, with only one exception, are statistically significant at the conventional level. Figures 3(a)-3(d) plot the estimated diversification structure of risk, along with its upper and lower one-standard deviation bounds, for each risk measure using the averages of the estimated coefficients. It can be seen that, first of all, there is a reduction tendency in return variance and return skewness as the portfolio size increases. This is consistent with the findings in previous studies using sample raw moments. Based on the fact that investors prefer lower return variance and more positive skewness, there exists a trade-off between 9 This specification has been used in the literature on yield curves. See, for example, Nelson and Siegel (1987). 12 lower variance and higher skewness when the portfolio is diversified. Second, both tails in the distribution are thicker when the portfolio is less diversified, as shown in the results for the 5% and 95% quantiles. The first four columns of Panel (B) in Table 2 report the diversifiable risk and the diversified risk by diversifying the portfolio with n stocks. The first, third, and fourth columns show that over 90% of the diversifiable risk can be diversified (the benefit of diversification) by holding five stocks in the portfolio. The second column shows that 90% of the diversifiable skewness is reduced (the cost of diversification) by holding five stocks. The gains or losses from diversification all vanish very quickly and are ignorable when the portfolio size is greater than ten. Value-at-Risk Analysis The goal of a quality risk system is to generate valid forecast distributions to enhance executive decision-making. Hence, most risk systems require models to generate valid ex-ante estimates of the forecast distribution, especially the left tail of the distribution. Recent research on Value at Risk (VaR) indicates that investors may be concerned strongly about the likely maximum loss, i.e. the left tail of the distribution of a portfolio’s expected returns. The evaluation of such large losses is a topic of increasing importance to investors operating in today’s tumultuous market environment. One major advantage of the proposed AGT distribution in this paper is its flexibility in estimating this risk measure. Of particular interest to us is the relationship between this maximum loss concern and the diversification strategy. The estimation results from the previous subsection are used to calculate the 1% and 5% conditional VaR’s for a one-day horizon. 10 Figures 4(a)–4(b) plot the estimated diversification 10 We also use weekly data to estimate the 1% and 5% conditional VaR’s at the weekly horizon. All qualitative patterns observed based on the daily data are well preserved in the weekly data. Specifically, the average expected maximum 1% and 5% weekly losses are 10.455% and 6.692%, 13 structure of risk along with its upper and lower one-standard deviation bounds. The last two columns of Table 2 report the regression results based on equation (4). It is shown that the expected extreme losses tend to be smaller as the portfolio size increases. The average expected maximum 1% and 5% daily losses are 4.805% and 3.030%, respectively, for a single-stock portfolio. The diversifiable losses are 1.493% and 0.945%, respectively. The average gains of diversifying the portfolio to a two-stock fund are 0.536% and 0.281%, respectively. More than 90% of the diversifiable losses are gone with a five-stock portfolio. The reduction in the extreme losses fades away as the portfolio size goes beyond five. 4. Alternative Data for Robustness Check Market Pool Even though the DJIA stocks are the most popular stocks held by investors and provide a well-diversified portfolio, one may argue that they still do not provide a representative picture of the market. To test the robustness of our conclusions, we replicate the empirical analysis for a more general set of stocks. This data set consists of all NYSE and AMEX firms that have complete return data (no missing value) during the sample period (January 1990 to December 2002). We follow the conventional method and exclude stocks that do not have a CRSP share code of 10 or 11, i.e., we only include ordinary common shares and exclude REITs, closed-end funds, primes, and scores. There are a total of 457 stocks in this pool. The random selection methods used before are applied to this new pool of data. We extend the maximum portfolio size to 50 to see whether our conclusions change when the portfolios include more than 30 stocks. The simulation results analogous to those in Figures 2-4 are plotted in respectively, for a single-stock portfolio. The diversifiable losses are 3.393% and 2.019%, respectively. More than 92% of the diversifiable losses are gone with a five-stock portfolio. 14 Figures 5-7. Table 3 shows that the results are analogous to those in Table 2. As expected, the simulations from the market pool yield a wider range of each risk measure. However, the qualitative results are consistent with those from the DJIA dataset. First, the absolute value of the slope is very big when the portfolio size is small, quickly declines as the portfolio size is approaching ten, and stays almost unchanged afterward, even beyond the portfolio size of 30. Our model of diversification structure of risk still fits quite well. Secondly, variance and skewness decline as the portfolio size increases, i.e., there exists a trade-off between lower variance and higher skewness when the portfolio is diversified. In addition, both tails in the distribution are thicker when the portfolio is less diversified, and the expected extreme losses (VaR) tend to be smaller as the portfolio size increases. Finally, 90% of the diversifiable risk can be diversified (the benefit of diversification) by holding five to six stocks, and 90% of the diversifiable skewness is reduced (the cost of diversification) by holding six stocks in the portfolio. The gains or losses from diversification all vanish very quickly and are ignorable when the portfolio size is greater than ten. Therefore, extending the sample to the market pool does not change our conclusions. Characteristics-Based Diversification Investors can achieve diversification by investing in the DJIA stocks, because the DJIA includes a wide variety of industries. However, these stocks are typically industry leaders and have high market capitalizations and trading volumes. Studies such as Campbell et al. (2001) suggest that returns on small firms be more volatile and therefore, to reduce the higher risk (measured by volatility) of small firms, it may be optimal to hold 20-30 stocks in a portfolio. This suggestion clearly is not supported by our simulations of industry-wise diversification. However, it would be interesting to see whether diversification strategies based on stock characteristics such as firm sizes 15 and liquidities would yield results different from those with industry-wise diversification. This section experiments these two alternative strategies to test the robustness of our conclusions. We first rank those 457 stocks used in the previous section according to their average sizes (market values) over the sample period. The seven stocks with the smallest sizes are dropped and the rest are put into 30 size-based groups, ranging from the smallest-size group to the biggest-size group, with each group including 15 stocks of similar sizes. We then construct portfolios containing different numbers of stocks in the following manner. First, a stock is randomly selected from the smallest-size group and the returns from this one-stock portfolio are used to estimate the model.11 Next, a stock is randomly selected from the second-to-the-smallest-size group and added to form a two-stock portfolio. The arithmetic average of the returns of these two stocks is used to estimate the model. We keep adding one stock at a time from a bigger size group to form 30 portfolios with portfolio sizes ranging from one to 30, where the portfolio with size n is a subset of the portfolio with size n+1. The thirtieth stock is randomly selected from the biggest-size group. Therefore, diversification is achieved by adding stocks from bigger size groups. The whole process is again repeated 100 times to yield 100 simulated samples, and the diversification structure of risk 11 We select stocks from size-ordered groups (from the smallest to the biggest), because we want to see the diversification effect when bigger-size stocks are added to the portfolio, in an attempt to compare our results with those from studies such as Campbell et al. (2001). A more general approach is to randomly select a size-group, randomly select another size-group from the other 29 groups, and so on. We also experiment with this more general approach and obtain similar results. The same sampling approach is also applied to the liquidity-based strategy experimented upon later. The results from both experiments do not change our conclusions and are available from the authors upon requests. 16 model is estimated using these 100 samples. Figures 8 (a)-(f) plots the estimated diversification structure of risk for the six risk measures. The second diversification strategy we experiment with in this section is based on the liquidity of the stocks, which is measured by the turnover ratios (trading volume divided by shares outstanding). We rank those 457 stocks according to their average turnover ratios over the sample period and conduct simulations similar to what we have done in the size-based diversification strategy. Figures 9 (a)-(f) plots the estimated diversification structure of risk for those six risk measures. Figures 8 and 9 show that our conclusions are not changed by using different diversification strategies based on stock characteristics: Variance, skewness, kurtosis, and extreme losses all decrease with increasing diversification, and the gains or losses from diversification all vanish very quickly and are ignorable when the portfolio size is greater than ten. 5. Discussion and Conclusion This paper addresses additional dimensions in the analysis of portfolio diversification and risk. In addition to the conventional measures of risk, namely variance and skewness, we propose a parametric model to estimate the whole distribution of asset returns and investigate the relationships between diversification and other distributional characteristics that have risk implications. Our results indicate that variance, kurtosis, and extreme losses decrease with increasing diversification. As such, the goal of an investor who wants to decrease variance and extreme losses would be to hold a more-diversified portfolio. However, portfolio skewness also decreases with increasing diversification. Therefore, an investor who desires higher skewness would choose to hold a less-diversified portfolio. These results indicate that an investor’s strategy of diversification depends on which risk measure is the main concern. 17 Our results also shed some light on the empirical puzzle that investors do not tend to hold fully diversified portfolios. A possible explanation is that most investors do not weigh downside risk equally with upside potential and prefer more positively skewed returns over low return variance and extreme losses. Furthermore, taking into account the transaction cost and information cost embedded in managing a more diversified portfolio, investors are likely not to hold more than ten stocks in their portfolios since the benefit of diversification beyond ten stocks is limited. The cost outweighs the benefit of a highly diversified portfolio. Possible extensions of this paper are considered. First, the literature on the risk-return tradeoffs using a three-moment CAPM indicates that stocks which decrease the skewness of a portfolio should have higher expected returns, that is, stocks with lower systematic skewness (i.e., coskewness with the market portfolio) outperform stocks with higher systematic skewness [see, for example, Kraus and Litzenberger (1976) and Harvey and Siddique (2000)]. In other words, investors prefer higher systematic skewness. Idiosyncratic skewness, on the other hand, does not affect expected returns. Therefore, it would be interesting to decompose the skewness measure into systematic skewness and idiosyncratic skewness and see whether the former decreases as the portfolio size increases, i.e., diversification decreases coskewness.12 Unfortunately, we are unable to do this decomposition using the current parametric model, because it would require a bivariate AGT distribution to jointly model the individual stock return and the market return, which is out of the scope of this paper. We leave this task to future studies. 12 In a more recent strand of the literature, Barberis and Huang (2005) and Kumar (2005) argue that models with cumulative prospect-theoretic preferences imply that idiosyncratic skewness should be priced as well. In this case, the relationship between diversification and idiosyncratic skewness is also an interesting topic. 18 Another possible extension of the paper is to analyze the actual holdings of stocks by investors in the market, rather than using the simulation method proposed in this paper. This would provide us with evidence from the real world, rather than from counterfactual simulations. For example, a unique dataset used by Barber and Odean (2000) and Mitton and Vorkink (2004), which consists of the investments of 78,000 households from January 1991 to December 1996, would serve this purpose. Since this dataset is not publicly available, we leave this extension to researchers who have access to it. Acknowledgement We are grateful to the editor, two anonymous referees, and the participants in the 2003 Conference of High-Frequency Financial Data in Taipei for helpful comments. Naturally, all remaining errors are ours. Yau acknowledges the research support from the National Science Council of the Republic of China (NSC93-2415-H-008-007). 19 Technical Appendix: Specifics of the Asymmetric Generalized t Distribution. The generalized t (GT) distribution in McDonald and Newey (1988) has the following pdf: p f GT ( xt ; ω , p, q ) = x ⎞ ⎛ 1 ⎞⎛ 2ωq B ⎜⎜ , q ⎟⎟⎜1 + t p ⎟ qω ⎟ ⎝ p ⎠⎜⎝ ⎠ 1 p p q+ 1 p , where B(.) is the beta function. To transform this symmetric distribution to an asymmetric one, we define: ⎧ ⎛ zt ⎞ ⎪ f GT ⎜ 1 + r ⎟ ⎠ ⎪ ⎝ h( zt ) = ⎨ ⎪ ⎛ | zt | ⎞ ⎟ ⎪ f GT ⎜ ⎩ ⎝1− r ⎠ for zt ≥ 0 , for zt < 0 , where -1<r<1. This specification allows different rates of descent for zt≥0 and zt<0. Next, we scale zt by defining xt = zt / A, where A is a scaling constant in order to simplify the notation. The density function of the asymmetric generalized t (AGT) distribution can be written as: ⎧ ⎛ A ⋅ xt ⎪ f GT ⎜⎜ ⎪⎪ ⎝ 1 + r ∂z t = h( A ⋅ xt ) ⋅ A = ⎨ f AGT ( xt ) = h( z t ) ⋅ ∂xt ⎪ ⎛ A ⋅ xt ⎪ f GT ⎜⎜ ⎪⎩ ⎝ 1− r ⎞ ⎟⎟ ⋅ A ⎠ for xt ≥ 0 , ⎞ ⎟⎟ ⋅ A ⎠ for xt < 0 . p⋅ A ⎧ ⎪ 1 ⎪ ⎛ ⎛ A⋅ x t ⎪ 2ωq p B ⎛⎜ 1 , q ⎞⎟⎜1 + ⎜ ⎟ ⎜ ⎜ ⎜ ⎪ ⎝ p ⎠⎝ ⎝ 1 + r ⎪ Specifically, f AGT ( xt ; A,ω , p, q, r ) = ⎨ ⎪ p⋅ A ⎪ 1 ⎪ ⎛ ⎪ 2ωq p B ⎛⎜ 1 , q ⎞⎟⎜1 + ⎛⎜ A ⋅ xt ⎜ p ⎟⎜ ⎜ 1 − r ⎪ ⎠⎝ ⎝ ⎝ ⎩ 20 p ⎞ 1 ⎟ ⎟ qω p ⎠ p ⎞ 1 ⎟ ⎟ qω p ⎠ ⎞ ⎟ ⎟ ⎠ q+ 1 p ⎞ ⎟ ⎟ ⎠ q+ 1 p for xt ≥ 0 , for xt < 0. 1 p 1 2ωq B( , q) p ; as such, the pdf becomes: Let A = p ⎧ ⎧ ⎛ 1 ⎪ ⎪ ⎜ 2 B( , q) ⋅ xt ⎪ p ⎪ ⎜ ⎪ f1 = ⎨1 + ⎪ ⎪ ⎜⎜ p(1 + r ) ⎪ ⎪⎩ ⎝ ⎪ f AGT ( x t ; p, q, r ) = ⎨ ⎪ ⎧ ⎛ ⎪ 1 ⎪ ⎜ 2 B( , q) ⋅ xt ⎪ p ⎪ f 2 = ⎪⎨1 + ⎜ ⎜ ⎪ ⎪ ⎜ p(1 − r ) ⎪ ⎪⎩ ⎝ ⎩ ⎞ ⎟ ⎟ ⎟ ⎟ ⎠ p ⎞ ⎟ ⎟ ⎟ ⎟ ⎠ p ⎫ ⎪ ⎪ ⎬ ⎪ ⎪⎭ −q− ⎫ ⎪ ⎪ ⎬ ⎪ ⎪⎭ 1 p −q− for xt ≥ 0 , for xt < 0 . 1 p The nth raw moment of xt is: ∞ Mn = ∫ ∞ ∫ −∞ ∞ 0 x n f AGT ( x )dx = x n f1 ( x )dx + ∫ ∫ −∞ 0 ∞ ∫ x n f 2 ( x )dx = x n f1 ( x )dx + ( −1) n x n f 2 ( x )dx. 0 0 n +1 ⎞ ⎛ n +1 n + 1 ⎞⎛⎜ ⎟⎟ p ⋅ k p ⎟ x (1 + k ⋅ x ) dx = B ⎜⎜ ,m − ⎟ p ⎠⎜ ⎝ p 0 ⎝ ⎠ ∞ Using the formula ∫ p −m n p −1 and letting p 1 1 ⎛ ⎞ ⎛ ⎞ ⎜ 2 B( , q) ⎟ ⎜ 2 B( , q) ⎟ p p ⎟ , k =⎜ ⎟ , and m = q + 1 , the nth raw moment becomes: k1 = ⎜ 2 ⎜ p(1 + r ) ⎟ ⎜ p(1 − r ) ⎟ p ⎜ ⎟ ⎜ ⎟ ⎝ ⎠ ⎝ ⎠ n +1 ⎛ n +1 n + 1 ⎞⎛ ⎞ ,m − M n = B ⎜⎜ ⎟⎟⎜ p ⋅ k1 p ⎟ p ⎠⎝ ⎠ ⎝ p = (1 + r ) n +1 n +1 ⎛ n +1 n + 1 ⎞⎛ ⎞ ,m − + ( −1) n B ⎜⎜ ⎟⎟⎜ p ⋅ k 2 p ⎟ p ⎠⎝ ⎠ ⎝ p n +1 n ⎡ 1 ⎤ p B( , q − ) ⎢2 B ( , q )⎥ p p ⎣ p ⎦ n = [( −1) (1 − r ) n −1 n +1 + (1 + r ) n +1 − n −1 + ( −1) (1 − r ) n n +1 n +1 n ⎡ 1 ⎤ ] p B( , q − ) ⎢2 B ( , q )⎥ p p ⎣ p ⎦ n 21 −1 n +1 n ⎡ 1 ⎤ p B( , q − ) ⎢2 B ( , q )⎥ p p ⎣ p ⎦ n − n −1 . − n −1 −2 ⎛2 1 ⎞ ⎡ ⎛ 1 ⎞⎤ Therefore, the mean is μ ≡ M 1 = 4 rp ⋅ B ⎜⎜ , q − ⎟⎟ ⎢2 B ⎜⎜ , q ⎟⎟⎥ ; the variance is σ 2 ≡ M 2 − μ 2 ; p ⎠ ⎣ ⎝ p ⎠⎦ ⎝p the skewness (SK) is: E ( x − μ )3 σ3 = E ( x 3 − 3μx 2 + 3μ 2 x − μ 3 ) σ3 = M 3 − 3μ (σ 2 + μ 2 ) + 3μ 2 μ − μ 3 σ3 = M 3 − 3μσ 2 − μ 3 σ3 ; and the kurtosis (KU) is: E ( x − μ )4 σ4 = E ( x 4 − 4 μx 3 + 6μ 2 x 2 − 4 μ 3 x + μ 4 ) = M 4 − 4 μM 3 + 6μ 2 M 2 − 4 μ 3 μ + μ 4 σ4 σ4 M − 4 μ ( M 3 − 3μσ 2 − μ 3 ) − 6μ 2σ 2 − μ 4 M 4 − 4 μ ⋅ SK ⋅ σ 3 − 6μ 2σ 2 − μ 4 . = 4 = σ4 σ4 ∞ Note that Pr( x ≥ 0) = ∫ 0 0 Similarly, Pr( x < 0) = ∫f −∞ 2 ( x )dx = ∞ ⎡ ⎤ 1 1 f1 ( x )dx = x f1 ( x )dx = (1 + r ) B ( , q ) ⎢2 B ( , q )⎥ p p ⎣ ⎦ 0 ∫ 0 −1 = 1+ r 2 . 1− r , which shows the asymmetry of the distribution when r ≠ 2 0. Furthermore, Pr(x≥0)+Pr(x<0)=1 confirms that fAGT is a proper pdf. 22 References Aggarwal, R. and R. Aggarwal. 1993. Security return distributions and market structure: Evidence from the NYSE/AMEX and the Nasdaq markets. Journal of Financial Research, 16, 209-220. Arrow, K. J., 1974. Essays in the Theory of Risk Bearing (Elsevier, New York). Barber, B. M., and T. Odean. 2000. Trading is hazardous to your wealth: The common stock investment performance of individual investors. Journal of Finance, 55, 773-806. Barberis, N. and M. Huang. 2005. Stocks as lotteries: the implications of probability weighting for security prices. Working paper, Yale University. Bekaert, G. and G. Wu. 2000. Asymmetric volatility and risk in equity markets. Financial Studies, 13, 1-42. Review of Bollerslev, T. 1987. A conditionally heteroskedastic time series model for speculative prices and rates of return. Review of Economics and Statistics, 69, 542-547. Bollerslev, T., R. Y. Chou, and K. F. Kroner. 1992. ARCH modeling in finance: a review of the theory and empirical evidence. Journal of Econometrics, 52, 5-59. Campbell, J. Y., M. Lettau, B. G. Malkiel, and Y. Xu. 2001. Have individual stocks become more volatile? An Empirical exploration of idiosyncratic risk, Journal of Finance, Vol 56, 1-43. Cooley, P. L. 1977. A multidimensional analysis of institutional investor perception of risk. Journal of Finance, 32, 67-78. Cromwell, N. O., W. R. L. Taylor, and J. A. Yoder. 2000. Diversification across mutual funds in a three-moment world. Applied Economics Letters, 7, 243-245. Engle, R. F. and V. K. Ng. 1993. Measuring and testing the impact of news on volatility. Journal of Finance, 48, 1749-1778. Evans, J. and S. Archer. 1968. Diversification and the reduction of dispersion: An empirical analysis. Journal of Finance, 23, 761-767. Fielitz, B., 1974. Indirect versus direct diversification. Financial Management, 3, 54-62. Gaivoronski, A. A. and Pflug, G., Value-at-risk in Portfolio Optimization: Properties and Computational Approach. Journal of Risk, 2004, 7, 1-31. Glosten, L. R., R. Jagannathan, and D. E. Runkle. 1993. On the relation between the expected value and the volatility of the normal excess return on stocks. Journal of Finance, 48, 1779-1801. Goetzmann, W. and A. Kumar. 2002. Equity portfolio diversification. Yale ICF Working Paper No. 00-59. 23 Greene, W. H. 2003. Econometric Analysis, fifth edition, New Jersey: Prentice Hall. Guidolin, M. and Timmermann, A., Term structure of risk under alternative econometric specifications. Journal of Econometrics, 2006, 131, 285-308. Harvey, C. R and A. Siddique. 2000. Conditional skewness in asset pricing tests. Journal of Finance, 55, 1263-1295. Hansen, B. E., Autoregressive conditional density estimation. International Economic Review, 1994, 35, 705-730. Johnson, N. L., 1949, “Systems of Frequency Curves Generated by Methods of Translation.” Biometrika, 36, 149-176. Kraus, A. and H. Litzenberger. 1976. Skewness preference and the valuation of risky assets, Journal of Finance, 31, 1085-1100. Kumar, A. 2005. Institutional skewness preferences and the idiosyncratic skewness premium. Working paper, University of Notre Dame. Lintner, J. 1965. Security prices, risk, and maximal gains from diversification. Journal of Finance, 20, 587-615. Malevergne, Y. and D. Sornette. 2004. VaR-efficient portfolios for a class of super- and subexponentially decaying assets returns distributions, Quantitative Finance, 4, 17-36. Malevergne, Y. and D. Sornette. 2005. Higher-moment portfolio theory, The Journal of Portfolio Management, 31(4), 49-55. McDonald, J. B., and W. K. Newey. 1988. Partially adaptive estimation of regression models via the generalized t distribution. Econometric Theory, 4, 428-457. Mitton, T. and Vorkink, K., Equilibrium underdiversification and the preference for skewness. Review of Financial Studies, 2006, forthcoming. Nelson C. R., and A. F. Siegel. 1987. Parsimonious modeling of yield curves. Journal of Business, 60 (4), 473-489. Newey, W. K. 1985. Generalized method of moments specification testing. Econometric, 29, 229-256. Journal of Newey, W. K. and K. D. West. 1987. A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix. Econometrica, 55, 703-708. Newey, W. K., and K. D. West. 1994. Automatic lag selection in covariance matrix estimation. The Review of Economic Studies, 61, 631-653. 24 Premaratne G. and A.S. Tay. 2002. How should we interpret evidence of time varying conditional skewness? Manuscript, Department of Economics, University of Singapore. Sharpe, W. F. 1964. Capital asset prices: A theory of market equilibrium under conditions of risk. Journal of Finance, 19, 425-442. Simkowitz, M. A. and W. L. Beedles. 1978. Diversification in a three-moment world. Journal of Financial and Quantitative Analysis, 13, 927-941. Theodossiou, P. 1998. Financial data and the skewed generalized t distribution. Management Science, 44, 1650-1661. Wang, K., Fawson, C., Barrett, C., and J.B. McDonald. 2001. A flexible parametric GARCH model with an application to exchange rates. Journal of Applied Econometrics, 16, 521-536. 25 Table 1: Estimated Coefficients and GMM Specification Test Statistics. (A) Sample Statistics Stock Mean Std Dev AA (B) Estimated Coefficients a (C) GMM Specification Tests b α β γ p q r (1) (2) t-Statistics (3) (4) χ (5) (6) 2 Statistic 0.058 2.100 0.943* 0.071* -0.043 2.054* 3.443* 0.132* 0.011 -0.175 0.810 0.645 0.697 -2.549* AIG 0.074 1.747 0.933* 0.078* -0.039* 2.028* 3.784* 0.090* -0.149 -0.210 0.752 0.293 0.438 0.434 1.246 AXP 0.072 2.256 0.924* 0.098* -0.062* 2.408* 2.378* 0.076* -0.234 -0.281 1.454 0.686 -1.948 -1.239 6.916 BA 0.043 2.039 0.928* 0.075* -0.025 2.312* 1.885* 0.052* -0.126 -0.066 0.728 0.781 1.142 -0.534 2.643 C 0.117 2.334 0.923* 0.091* -0.077* 2.257* 2.658* 0.065* -0.003 0.283 0.557 -0.013 -1.266 2.662 0.919 8.510 CAT 0.065 2.058 0.969* 0.043* -0.029* 2.124* 2.359* 0.126* -0.622 -0.632 1.276 -0.410 -1.415 -0.479 4.716 DD 0.052 1.854 0.966* 0.040* -0.021* 2.075* 3.334* 0.099* -0.403 -0.125 0.528 -0.001 -1.887 0.311 3.700 DIS 0.041 2.072 0.971* 0.040* -0.031* 2.409* 1.955* 0.089* -0.494 0.024 0.566 -1.786 -1.933 0.345 7.822 GE 0.070 1.749 0.942* 0.086* -0.068* 2.349* 3.092* 0.054 -0.535 -0.158 0.487 -0.123 0.137 -0.002 0.600 GM 0.038 2.068 0.917* 0.086* -0.053* 1.898* 5.224* 0.118* -0.245 -0.236 0.825 -2.106* -1.463 0.347 7.581 HD 0.107 2.314 0.927* 0.094* -0.073* 2.315* 2.609* 0.055* -0.649 0.268 0.827 0.160 -0.230 0.432 1.204 HON 0.064 2.226 0.866* 0.162* -0.112* 2.249* 1.968* 0.047 0.690 -1.259 -1.258 -1.308 5.616 HPQ 0.082 2.752 0.972* 0.040* -0.028* 2.446* 1.729* 0.060* -0.389 -0.009 0.610 IBM 0.067 2.126 0.927* 0.101* -0.083* 2.551* 1.474* 0.071* -0.061 -0.868 0.558 -1.156 0.285 INTC 0.125 2.915 0.950* 0.058* -0.038 3.007* 1.372* 0.047 0.237 0.410 -1.191 -0.637 0.046 JNJ 0.081 1.647 0.893* 0.110* -0.088* 2.345* 2.816* 0.052* -0.443 0.059 JPM 0.072 2.446 0.932* 0.114* -0.100* 2.001* 3.912* 0.068* 0.123 KO 0.499 0.526 -0.319 -1.028 1.220 1.783 3.576 0.128 -0.751 0.017 2.839 0.170 -0.589 -0.620 1.099 0.174 3.625* -0.297 -1.405 -0.088 13.980* 0.066 1.692 0.934* 0.083* -0.066* 2.381* 2.363* 0.078* -0.448 -0.108 1.299 0.013 -0.536 0.188 MCD 0.038 1.758 0.968* 0.032* -0.014 2.332* 2.336* 0.098* -0.580 -0.131 0.635 0.261 -1.458 -1.104 3.881 MMM 0.059 1.543 0.980* 0.021 0.278 2.998 MO -0.004 1.885* 2.864* 0.063* -0.302 -0.048 1.486 0.063 2.064 0.812* 0.163* -0.040 2.564* 1.319* 0.028 0.199 -0.633 2.157 -0.983 0.362 0.780 1.049 -0.130 1.428 4.419 MRK 0.070 1.781 0.938* 0.074* -0.062* 2.211* 2.894* 0.064* -0.759 0.005 0.514 0.363 -0.411 0.398 1.095 MSFT 0.143 2.393 0.870* 0.132* -0.087* 2.303* 2.688* 0.093* -0.231 -0.061 0.858 0.720 -1.588 -0.452 4.082 PFE 0.098 1.968 0.922* 0.084* -0.046* 2.284* 3.224* 0.051 -0.424 -0.107 0.955 -0.425 -2.092* -1.948 8.630 PG 0.072 1.731 0.916* 0.079* -0.047 2.351* 2.246* 0.048 -0.467 0.601 0.071 0.118 0.850 SBC 0.047 1.842 0.917* 0.099* -0.052* 1.779* 7.564* 0.056* -0.119 0.027 0.576 0.747 -0.572 0.443 1.531 UTX 0.073 1.874 0.907* 0.121* -0.085* 1.963* 3.523* 0.048 -0.178 0.513 0.931 0.552 -0.136 -0.397 1.619 VZ 0.043 1.799 0.944* 0.074* -0.045* 2.049* 3.886* 0.069* -0.067 0.055 0.306 0.114 -0.052 -0.447 0.292 WMT 0.090 2.053 0.944* 0.071* -0.042* 2.150* 3.681* 0.073* -0.052 -0.138 0.447 -0.442 0.037 0.213 XOM 0.056 1.457 0.934* 0.073* -0.034* 2.073* 4.350* 0.047 -0.031 0.077 1.154 0.008 -0.157 0.204 0.411 1.570 0.402 -0.639 The asterisk * indicates significance at the 5% level. a. The parameters are for equations (2) and (3). All computations were performed using the GAUSS MAXLIK module. The estimated standard errors were calculated with the robust standard errors corresponding to results summarized in Greene (2003, p.520). b. The null hypotheses are (1) E(vt) = 0, (2) E( v t2 -1) = 0, (3) E(vt ⋅ vt-1) = 0, (4) E(vt ⋅ vt-2) = 0, (5) E(vt ⋅ vt-3) = 0, and (6) E(vt ⋅ vt-4) = 0, where vt is the standardized residual from the AR-GARCH-AGT model. c. This is a joint test for the null hypothesis that the model is correctly specified based on the moment conditions. It has a χ2(6) distribution. The covariance matrix of the orthogonality conditions is calculated using the consistent estimator proposed by Newey and West (1987, 1994) with a Bartlett kernel. 26 c Table 2: Estimation Results for Equation (4): Ri ,n = ai + bi ⋅ 1 n + ci ⋅ + ei ,n ; DJIA exp( n ) exp( n ) Stocks Risk Measure Variance Skewness 5% quantile 95% quantile 1% VaR 5% VaR 0.683 0.651 (A) Estimation Results Avg R2 0.880 0.758 0.805 0.857 Average of estimated coefficients and their t-statistics (in parentheses) â 1.587 (116.4) -0.019 (-19.27) -1.975 (-225.5) 2.085 (243.5) 3.311 (246.1) 2.085 (224.2) b̂ 6.308 (9.917) -0.033 (-1.900) -1.860 (-8.706) 2.147 (9.422) 1.042 (2.110) 0.230 (0.726) ĉ 1.460 (3.891) 0.273 (16.78) -1.179 (-6.541) 1.525 (7.688) 3.017 (6.379) 2.340 (7.456) (B) Diversifiable and Diversified Risk ⎛ Average risk measure for n=1 (in percentages): ⎜ aˆ + ⎜ ⎝ 4.445 Diversifiable risk: 2.858 n 2 3 4 5 6 7 8 9 10 11 12 13 14-30 Diversified risk: 1.609 2.326 2.635 2.766 2.820 2.843 2.852 2.855 2.857 2.857 2.857 2.858 2.858 0.069 bˆ + cˆ ⎞ ⎟ exp(1) ⎟ ⎠ -3.093 3.436 4.805 3.030 -1.118 1.351 1.493 0.945 0.536 0.991 1.253 1.385 1.446 1.473 1.485 1.490 1.492 1.493 1.493 1.493 1.493 0.281 0.584 0.770 0.865 0.910 0.930 0.939 0.943 0.944 0.945 0.945 0.945 0.945 bˆ + cˆ exp(1) 0.088 ⎧⎪ bˆ + cˆ ⎡ 1 n ⎤ ⎫⎪ − ⎢ bˆ ⋅ + cˆ ⋅ ⎨ ⎬ exp(1) ⎣ exp( n ) exp( n ) ⎥⎦ ⎪ ⎩⎪ ⎭ 0.019 0.049 0.069 0.079 0.084 0.086 0.087 0.088 0.088 0.088 0.088 0.088 0.088 -0.547 -0.849 -0.997 -1.066 -1.096 -1.109 -1.114 -1.116 -1.117 -1.118 -1.118 -1.118 -1.118 0.648 1.016 1.200 1.285 1.323 1.339 1.346 1.349 1.350 1.351 1.351 1.351 1.351 Note: All risk measures, except for skewness, are in percentage. 27 Table 3: Estimation Results for Equation (4): Ri ,n = ai + bi ⋅ 1 n + ci ⋅ + ei ,n ; Market exp( n ) exp( n ) Pool Risk Measure Variance Skewness 5% quantile 95% quantile 1% VaR 5% VaR 0.676 0.647 (A) Estimation Results Avg R2 0.879 0.601 0.818 0.843 Average of estimated coefficients and their t-statistics (in parentheses) â 1.142 (70.427) -0.050 (-32.268) -1.687 (-139.910) 1.731 (151.558) 2.863 (106.910) 1.801 (105.085) b̂ 9.963 (5.973) -0.374 (-13.888) -1.828 (-4.556) 1.811 (3.962) -1.962 (-1.824) -1.719 (-2.613) ĉ 3.797 (4.607) 0.686 (25.250) -3.168 (-9.981) 4.104 (11.626) 8.838 (9.418) 5.797 (10.036) (B) Diversifiable and Diversified Risk ⎛ Average risk measure for n=1 (in percentages): ⎜ aˆ + ⎜ ⎝ 6.204 Diversifiable risk: 5.062 n Diversified risk: 0.065 bˆ + cˆ ⎞ ⎟ exp(1) ⎟ ⎠ -3.525 3.907 5.393 3.302 -1.838 2.176 2.530 1.500 0.403 1.307 1.918 2.245 2.403 2.475 2.507 2.520 2.526 2.528 2.529 2.530 0.164 0.720 1.107 1.317 1.418 1.465 1.485 1.494 1.498 1.499 1.500 1.500 bˆ + cˆ exp(1) 0.115 ⎧⎪ bˆ + cˆ ⎡ 1 n ⎤ ⎫⎪ − ⎢ bˆ ⋅ + cˆ ⋅ ⎨ ⎬ exp(1) ⎣ exp( n ) exp( n ) ⎥⎦ ⎪ ⎩⎪ ⎭ 2.686 -0.020 -0.733 0.820 3.999 0.031 -1.274 1.473 4 4.601 0.071 -1.572 1.842 5 4.867 -1.719 2.026 0.094 6 4.981 0.105 -1.786 2.111 7 5.029 0.111 -1.816 2.148 8 5.049 0.113 -1.829 2.165 9 5.057 0.114 -1.834 2.171 10 5.060 0.114 -1.837 2.174 11 5.061 0.115 -1.837 2.175 12 5.062 0.115 -1.838 2.176 13-50 5.062 0.115 -1.838 2.176 Note: All risk measures, except for skewness, are in percentage. 2 3 28 Figure 1 (a) Generalized Student t Density for p=2, q=100 ___ r = 0 ---- r = 0.5 …. r = -0.5 Figure 1 (b) Generalized Student t Density for r=0 ___ p=2 q = 3 (t distribution) ---- p=2 q = 100 (Normal) …. p=1 q = 100 (Laplace) 29 Figure 2: Risk Measures against Portfolio Sizes (100 Samples) (DJIA) Figure 3: Estimated Diversification Structure of Risk (DJIA) (a) (a) (b) (b) (c) (c) (d) (d) 30 Figure 4 (a): Estimated Diversification Structure of Risk (1-day horizon 1% VaR) (DJIA) Figure 4 (b): Estimated Diversification Structure of Risk (1-day horizon 5% VaR) (DJIA) 31 Figure 5: Risk Measures against Portfolio Sizes (100 Samples) (Market Pool) Figure 6: Estimated Diversification Structure of Risk (Market Pool) (a) (a) (b) (b) (c) (c) (d) (d) 32 Figure 7 (a): Estimated Diversification Structure of Risk (1-day horizon 1% VaR) (Market Pool) Figure 7 (b): Estimated Diversification Structure of Risk (1-day horizon 5% VaR) (Market Pool) 33 Figure 8: Estimated Diversification Structure of Risk (Size-Based Diversification) (a) Figure 9: Estimated Diversification Structure of Risk (Liquidity-Based Diversification) (a) (b) (b) (c) (c) (d) (d) (e) (e) (f) (f) 34