Beta Based Portfolio Construction:

Transcription

1 ÖREBRO UNIVERSITY School of Business Economics, Master Thesis Supervisor: Niclas Krüger Examiner: Dan Johansson Fall 2017 Beta Based Portfolio Construction: Stock Selection Based on Upside- and Downside Market Risk Authors: Andreas Johansson Robin Petersson

2 Abstract In the traditional Capital Asset Pricing Model (CAPM), a stock s beta represents an average of its upside- and downside market risk. As a result, it may not be an accurate reflection of how the stock tends to move in market upturns and downturns respectively. The purpose of this study is to investigate if it is possible to achieve a higher risk-adjusted return by selecting stocks based on their upside- and downside beta when constructing financial portfolios. Using weekly Swedish stock data from to , and the Stockholm Benchmark Index () as a proxy for the market portfolio, upside- and downside stock betas are estimated over a rolling two-year period using a single-factor model. The upside- and downside betas, as well as the difference between the two measures, are ranked and the stocks are sorted into equally weighted quintile portfolios. Each portfolio is then rebalanced every six-month and tested outof-sample against an equally weighted benchmark index. The results indicate that although there seems to exist an upside-downside beta asymmetry, the portfolio performance seems to be largely explained by the traditional CAPM beta. Keywords: Capital Asset Pricing Model, conditional CAPM, upside beta, downside beta, upside risk, downside risk, portfolio construction.

3 Acknowledgements We would like to thank everyone who has helped us with this thesis by providing feedback and proof-reading the text. We would also like to thank our supervisor, Niclas Krüger, for always being available and willing to discuss any difficulties, as well as providing valuable comments and suggestions on how to move forward with the thesis. Andreas Johansson Robin Petersson

4 Table of Contents 1. Introduction 1 2. Theoretical Background The Conditional CAPM The Upside- and Downside Beta 4 3. Data Financial Market Data Data Frequency and Time Period 6 4. Empirical Model The Conditional CAPM and Portfolio Construction Performance Statistics 9 5. Results Discussion Conclusion 18 References 19 Appendix A 22 Appendix B 25 Appendix C 36 Appendix D 39

5 1. Introduction In the traditional Capital Asset Pricing Model (CAPM), a stock s beta represents an average of its upside- and downside market risk. It is also assumed to be static, i.e. unconditional upon the period of estimation. As a result, the beta may not be an accurate reflection of how the stock tends to move in market upturns and downturns respectively. The purpose of this study is to investigate if it is possible to achieve a higher risk-adjusted return by selecting stocks based on their upside- and downside beta when constructing financial portfolios. Modern portfolio theory is based on the findings of Markowitz (1952) who showed that efficient financial portfolios can be constructed using mean-variance optimization. The framework was later extended by several authors whom, independently, derived the CAPM (Sharpe, 1964; Lintner, 1965; Mossin, 1966). The model illustrates that if investors optimize their portfolios using mean-variance optimization, the aggregate market portfolio will be meanvariance efficient and only contain systematic risk. The risk-premium demanded by a rational investor therefore only depends on the asset s market risk, as measured by its beta in the CAPM. Since its inception, the CAPM has been subject to various forms of critique. Several studies have, for example, concluded that the assumption of normality is unrealistic when modelling asset returns. Markowitz (1959) himself suggested the use of semi-variance in mean-variance optimization since it better reflects an assets downside risk. Furthermore, Mandelbrot (1963) observed longer tails than under the condition of normality, suggesting that non-normal distributions with fatter tails better explain asset returns. Several empirical studies also indicate that the correlation between stocks tend to increase during market turmoil (see King & Wadhwani, 1990; Longin & Solnik, 1995; Forbes & Rigobon, 2002), possibly increasing an asset s beta. Bawa & Lindenberg (1977) suggested extending the CAPM to incorporate the possibility of asymmetric upside- and downside risk. A later empirical study by Ang, Chen & Xing (2005) uses a similar framework when analysing the risk-premium for downside risk in the U.S. stock market. They sort stocks by their CAPM-, upside- and downside beta 1, as well as the difference between these measures. The authors 1 Ang, Chen & Xing (2005) define a stock s upside (downside) beta as the covariance between the stock s return and the market return divided by the market variance, conditioned on market returns greater (less) than the average market return. 1

6 illustrate that high downside beta stocks have higher average returns, indicating that risk averse investors demand an additional risk premium for holding downside risk. Guy (2015) argues that if investors are risk-averse, they have incentives to take advantage of the upside- and downside beta asymmetry. By constructing portfolios that have a high upside risk and a low downside risk, an investor can earn a higher return in market upswings, while hedging some of the downside risk. Using monthly U.S. stock returns from the year , Guy estimates upside- and downside betas, conditioned on positive and negative market returns, over a rolling three-year period, and ranks the stocks by each measure. He then constructs a Target Date fund consisting, among other assets, of the lowest downside beta quintile, and the highest upside beta quintile, holds it for a year, and then rebalances it. The results indicate that an investor can achieve a higher risk-adjusted return using this approach of investing. This study builds on the work of Guy (2015), but differs in a few aspects. First, upside- and downside beta portfolios are tested using Swedish stock data over the period to Second, since a large difference between a stock s upside- and downside beta may indicate a greater upside risk relative to its downside risk, portfolios based on this difference are also tested to analyse its effect on risk-adjusted returns 2. Third, shorter estimation periods with more frequent data are used to better account for changing betas while also increasing the sample size. Fourth, since zero transaction costs is an unrealistic assumption for an actual investor, rebalancing costs are added to analyse their effects on the portfolio returns. Finally, the skewness and kurtosis of each portfolio are analysed to examine if the method of selecting stocks has any visible effects on the return distribution of the portfolios. The results indicate that although there seems to exist an upside-downside beta asymmetry, the portfolio performance seems to be largely explained by the traditional CAPM beta where low beta portfolios outperformed high beta portfolios. The rest of the thesis is organized as follows. Chapter 2 provides a theoretical background on the conditional CAPM and the upside- and downside stock beta. In Chapter 3, the data is introduced followed by the methodology of constructing portfolios in Chapter 4. Chapter 5 presents the results, which are discussed in Chapter 6. Chapter 7 concludes. 2 Ang, Chen & Xing (2005) found that stocks with a small realized difference between their upside- and downside beta had higher average returns, indicating that a risk premium was required for holding stocks with a large downside risk relative their upside risk. The returns were, however, not risk-adjusted. 2

7 2. Theoretical Background 2.1 The Conditional CAPM In the traditional CAPM, the expected return of an asset is a linear function of the market excess return since all unsystematic risk can be eliminated through the process of diversification. As a result, a rational investor will only demand a risk premium for bearing market risk. In mathematical form, the model can be expressed as: E r i = r f + β i E r M - r f (2.1) where β i = Cov r i - r f, r M - r f Var r M - r f (2.2) and E r i is the expected return of asset i, r f is the risk-free interest rate, β i is asset i s sensitivity to market risk, and E r M is the expected market return (Bodie, Kane & Marcus, 2014). Furthermore, investors are assumed to plan for a single homogenous time-period, causing the asset s beta and the market risk premium to be unconditional upon a given point in time, i.e. they are assumed to be constant. In reality, however, investors plan multiple periods ahead, suggesting that an asset s beta and the market risk premium are conditional upon a given point in time (Ghysels, 1998). The difference between the unconditional- and conditional CAPM can be illustrated using the expression for the covariance: Cov Y 1, Y 2 = E Y 1 Y 2 - E Y 1 E Y 2 (2.3) E Y 1 Y 2 = E Y 1 E Y 2 + Cov(Y 1, Y 2 ) (2.4) Substituting asset i s market beta, β i, for Y 1, and the market excess return, R M, for Y 2, equation (2.4) becomes: E β i R M = E β i E R M + Cov(β i, R M ) (2.5) 3

8 where E β i R M is the risk premium of asset i. If the unconditional model holds, the expected beta, E β i, and the market risk premium, E R M, are both constant, and their covariance, Cov(β i, R M ), is equal to zero. If, however, E β i and E R M covary positively over time, the unconditional beta will underestimate the asset s market risk, and if they covary negatively, it will overestimate the market risk (Barinov, 2014). As an illustration, consider the following example by Barinov (2014) using a single risky asset i. Assume that throughout the business cycle, an expansion lasts three times longer than a recession and that the asset s beta and the market risk premium is equal to two thirds (β i Exp. = " ) and four percent (E(R Exp. # M ) = 4%) respectively. During the recession, when the systematic risk increases, the asset s beta and the market risk premium rises to two (β i Rec. = 2) and twelve percent (E(R M Rec. ) = 12%) respectively. This suggests that the average market risk premium over the business cycle will be six percent ( 3 E(R Exp. ) M E(R Rec. ) = 6%), and the average beta 4 M according to the unconditional CAPM will be equal to one ( # β Exp. + % $ i β Rec. = 1). However, $ i since the market risk premium and asset i s beta covary positively, the actual risk premium of asset i is eight percent ( 3 β Exp. E(R Exp. 4 i M ) + 1 β Rec. E(R Rec. 4 i M ) = 8%), resulting in a two percent difference compared to the unconditional case. According to the conditional CAPM, this two percent difference is not a positive alpha, but a result of the positive correlation between asset i s beta and the market risk premium. 2.2 The Upside- and Downside Beta In most research on the conditional CAPM, the asset s market beta is conditioned on the time period it was estimated, e.g. a rolling one-year period, or other risk factors, e.g. market capitalization. Usually, however, they do not condition on upside- and downside market movements within the sample period. Since the return distribution of financial assets is often asymmetric, and their covariance tends to increase during market downturns, an asset s downside beta may be different from its upside beta. An asset s upside beta may thus be conditioned on market returns over some predetermined value c 3, and its downside beta may be 3 Ang, Chen, and Xing (2005) tested the robustness of their results by setting c equal to the previous year s average market return, the one-month T-bill rate, and zero, but found that the results remained largely unaffected by the choice of c. 4

9 conditioned on market returns below this value. Mathematically, asset i s upside beta, β i +, and downside beta, β ī, for period t can be expressed as (Ang, Chen & Xing, 2005): β + i,t = Cov R i,t, R M,t r M,t > c Var R M,t r M,t > c (2.6) - β i,t = Cov R i,t, R M,t r M,t c Var R M,t r M,t c (2.7) where R i,t is the excess return of asset i over the risk-free interest rate during period t, R M,t is the market excess return over the risk-free interest rate during period t, and r M,t is the market return during period t. By separating the upside beta from the downside beta, and incorporating the potential risk asymmetry, an investor may, in other words, be able to construct more efficient financial portfolios. 5

10 3. Data 3.1 Financial Market Data The data consists of weekly observations over the total return indices of stocks listed on NASDAQ OMXS from to Given data availability, stocks with listing (delisting) dates after (before) ( ) were also included to increase the sample size and to reduce the effect of survivorship bias, resulting in a sample size of 515 stocks. In addition, weekly observations over the Stockholm Benchmark Index () 4, and the 6-month Swedish Treasury bill (SSVX6M) rate were used as proxies for the market portfolio and the risk-free interest rate. Table 3.1 provides descriptive statistics over the listed-, delistedand total number of stocks, the, and the risk-free interest rate. It includes the number of return indices in each category, their average return, standard deviation, skewness, and kurtosis. The stock data was collected from Bloomberg (2017), and the T-bill data was collected from the Riksbank (2017). Table 3.1 Descriptive statistics over weekly returns. Category Number of Indices Avg. Return % (Max/Min) Listed Stocks (329 / -75) Delisted Stocks (485 / -98) All Stocks (485 / -98) Avg. SD % (Max/Min) 5.96 (19.22 / 1.11) 7.42 (22.63 / 2.63) 6.56 (22.63 / 1.11) Avg. Skewness (Max/Min) 0.99 (11.30 / -1.23) 1.55 (22.45 / -9.32) 1.22 (22.45 / -9.32) Avg. Kurtosis (Max/Min) ( / -0.03) ( / -0.47) ( / -0.47) SSVX6M Note: the table shows descriptive statistics over the weekly returns for the listed-, delisted- and total number of stocks, the, and the SSVX6M over the period to It includes the number of indices in each category, the average return, standard deviation, skewness, and kurtosis. Max- and min values are presented in parentheses. 3.2 Data Frequency and Time Period Weekly data was used in this study for several reasons. First, it contains less noise than daily data, making it more likely that the estimates represent pure signals. In addition, and although weekly data may contain more noise than monthly data, the occurrence of several market events 4 The is a free-float adjusted-, capitalization-weighted-, total return index, designed as an indicator of the Stockholm Stock Exchange (Bloomberg, 2017). 6

11 within a given month may result in different market responses averaging out, not capturing their full effects. Second, stock betas tend to change over time (see Blume, 1971; Fabozzi & Francis, 1978), creating the need to estimate betas over relatively short periods of time. Guy (2015) used monthly observations over a rolling three-year period, but to increase the number of observations, while shortening the estimation period to account for changing betas, weekly data over a rolling two-year period was chosen in this study. Third, as argued by Momcilovic, Begovic, & Tomasevic (2014), weekly returns seem to provide more stable in-sample estimates of stock betas, further supporting the use of weekly data. The starting date was chosen since it allowed the use of the as a proxy for the market portfolio from the first week of its inception, week 1 of 1996, up until week 26 of It also resulted in a sufficient amount of out-of-sample observations when testing the portfolios, as well as an adequate number of stocks for each of them to be sufficiently diversified. 7

12 4. Empirical Model 4.1 The Conditional CAPM and Portfolio Construction In order to construct portfolios based on upside- and downside betas, as well as the difference between the two measures, upside- and downside stock betas were estimated over a rolling twoyear period for each stock on the Stockholm Stock Exchange using ordinary least squares (OLS). To reduce the problem of survivorship bias, stocks that were listed and delisted over the sample period were also included, although with a couple of constraints. First, stocks that had not been listed for the entire two-year period were excluded due to their short history. Second, stocks that were delisted within the next six-month period were also excluded. This since it is often known a few months in advance if a stock will be delisted (Nasdaq 2016), and including these would increase the frequency and cost of rebalancing the portfolios. For each two-year period and stock, the upside- and downside beta was estimated according to: R i,t = α i + β i + R,t + ε i,t, r,t > 0 (4.1) R i,t = α i + β ī R,t + ε i,t, r,t 0 (4.2) where R i,t is stock i s excess return over the six-month T-bill rate during week t, α i is the intercept, β + i (β ī ) is the upside (downside) beta, R,t is the excess return over the sixmonth T-bill rate during week t, ε i,t is stock i s error term during week t, and r,t is the return of the during week t. The stocks were ranked lowest to highest according to their upside beta, downside beta, as well as the difference between their upside- and downside beta, resulting in three lists of stocks. Each list was then divided into equally weighted quintile portfolios, where the first quintile contained the 20 percent of stocks with the lowest beta values, the second quintile contained the 20 percent of stocks with the second lowest beta values, and so on. The portfolios were then held for a period of six months before new beta values were estimated over the most recent 8

13 two-year period, and new quintile portfolios were constructed. This resulted in 15 time-series of portfolio returns, one for each quintile portfolio 5. The portfolios were evaluated in three steps. First, to evaluate each portfolio s relative performance, a benchmark index consisting of all eligible stocks, the same rebalancing date, and weighting technique as the quintile portfolio had to be constructed. However, since positive and negative observations of the market return are required to estimate both the upside- and downside beta of a stock using OLS, the composition of stocks for each benchmark will differ at some point in time. For example, if the total return index of a stock has periods of no observations, and the market return is negative (positive) for all existing observations during a given two-year period, the stock s upside (downside) beta, and, hence, beta difference, cannot be estimated due to a lack of observations. Therefore, three benchmark portfolios had to be constructed, one for the upside beta portfolios, one for the downside beta portfolios, and one for the upside-minus-downside beta portfolios 6. Second, to analyse the effect of transaction costs, rebalancing costs of 0.5-, 1-, and 2 percent of the total portfolio value were added each (semi-annual) rebalancing date. Finally, in order to examine if the method of selecting stocks had any visible effects on the return distribution of the portfolios, the skewness and kurtosis were calculated to sort out potential trends. 4.2 Performance Statistics Performance statistics were used to evaluate each portfolio s performance against its benchmark, but since different statistics measure different return characteristics, the ordinal ranking of the portfolios may change depending on the measure being used. Therefore, several widely-used statistics were calculated to provide a more comprehensive overview of each portfolio s relative performance. First, to measure each portfolio s reward-to-volatility tradeoff, the Sharpe ratio was calculated. It is defined as the portfolio s average excess return, R p, divided by the standard deviation of the returns, σ p, as illustrated in equation (4.3): 5 That is, five portfolios based on upside betas, five portfolios based on downside betas, and five portfolios based on the difference between each stock s upside- and downside beta. 6 The, or other market indices, were not used for evaluation since they are usually capitalization weighted, their composition of stocks differ, as well as their rebalancing dates. Using an existing index could therefore lead to incorrect inferences. 9

14 Sharpe ratio = R p σ p (4.3) Second, to measure each portfolio s reward-to-market risk, the Treynor ratio was calculated, which is defined as the portfolio s average excess return divided by the portfolio s beta, β p, as illustrated in equation (4.4): Treynor ratio = R p β p (4.4) Third, to measure how much of the portfolio s return that is not explained by its exposure to the benchmark index, Jensen s alpha was estimated. It is defined as the portfolios average excess return minus its expected excess return predicted by the benchmark, R b, using a singlefactor model, as shown in equation (4.5): Jensen s alpha = R p - β p R b (4.5) Fourth, to risk-adjust each portfolio s excess return over its benchmark, the Information ratio was calculated. It is defined as the portfolio s alpha divided by the volatility of the excess return (i.e. the tracking error), σ e, as illustrated in equation (4.6): Information ratio = α p σ e (4.6) Finally, two measures of downside risk were calculated, the weekly five percent Value at Risk (VaR) and the weekly five percent Expected Shortfall (ES). The five percent VaR is defined as the fifth percentile of portfolio returns, indicating that five percent of portfolio returns are expected to fall below this value. The five percent ES is defined as the expected return given that it falls below the fifth percentile (Bodie, Kane & Marcus, 2014; Hull, 2012). 10

15 5. Results Table 5.1 presents descriptive statistics for each of the five upside- (U), downside- (D) and upside-minus-downside (UMD) beta portfolios, their respective benchmark, and the, over the out-of-sample period week 1 of 1998 to week 26 of It includes the ending value of one SEK invested at the beginning of the period, the geometric- and arithmetic averages, as well as the yearly standard deviation and tracking error 7. As the table illustrates, a similar pattern is visible for both the upside- and downside beta portfolios, where portfolios with a lower upside- or downside beta tend to have a higher geometric average return, as well as a lower standard deviation and tracking error. For the upside-minus-downside beta portfolios however, another pattern is visible. The geometric average return tends to increase, and the standard deviation and tracking error tend to fall, the more moderate the size of the difference between the upside- and downside beta. The return of each portfolio also seems to be explained by its traditional CAPM beta, shown in Table A.1 of Appendix A, where lower beta portfolios tend to generate a higher geometric return, as well as a lower standard deviation. Overall, portfolio D1 had the highest ending value (25.72SEK), geometric- and arithmetic averages (18.04% and 21.94% respectively), and the lowest standard deviation (27.94%), while portfolio UMD3 had the lowest tracking error (7.56%). The time-series of each portfolio is presented in Figure A.1 to A.3 of Appendix A. Table 5.2 presents the performance statistics, i.e. the annual Sharpe ratio, Treynor ratio, Jensen s alpha, Information ratio, and the weekly five percent VaR and ES, of each portfolio, their respective benchmark, and the. The same pattern as in Table 5.1 is visible, where lower upside- and downside beta portfolios, as well as portfolios with a more moderate difference between the two measures, tend to have a higher risk-adjusted return. As the table 7 The return characteristics have been annualized from a weekly basis, but since none of the portfolios follow a geometric Brownian motion (see Appendix C), the standard method of scaling, t, would underestimate the annual volatility. As a result, the Hurst exponent of each portfolio was estimated and used to annualize the returns. Using a similar method as Hamed (2006), each time-series (N=1018) was initially divided into overlapping sub-periods, starting at n = 10 observations. The length of each sub-period was then increased by a factor 2 until n = N 4 was reached, yielding ten series of overlapping sub-periods containing n equal to 10, 14, 20, 29, 41, 58, 81, 115, 163, and 230 observations. The classical rescaled range and standard deviation were then calculated according to equation (D.1) and (D.2) in Appendix D, and the Hurst exponent was estimated using OLS according to equation (D.3). The average Hurst exponent for the quintile portfolios was 0.68, and the highest and lowest Hurst exponent was 0.70 and 0.66 respectively. See Appendix D for more details on the method. 11

16 Table 5.1 Descriptive statistics over annual portfolio returns. Portfolio Ending value of 1SEK Geometric Average % Arithmetic Average % Standard Deviation % Tracking Error % U1 (Low β + ) U U U U5 (High β + ) Benchmark (U) n/a D1 (Low β - ) D D D D5 (High β - ) Benchmark (D) n/a UMD1 (Low β + - β - ) UMD UMD UMD UMD5 (High β + - β - ) Benchmark (UMD) n/a n/a Note: the table shows descriptive statistics for each of the five upside- (U), downside- (D) and upside-minus-downside (UMD) beta portfolios, as well as their respective benchmark and the, over the out-of-sample period week 1 of 1998 to week 26 of It includes the ending value of 1SEK invested at the beginning of the period, the geometric- and arithmetic averages, the standard deviation, and the tracking error. The values have been annualized from a weekly basis. illustrates, portfolios U1-U3 and D1-D2 all have better performance statistics than their respective benchmark, although portfolios UMD2-UMD3 also show some indication of outperforming their benchmark. Again, portfolio D1 seems to be the best performing portfolio with the highest Sharpe ratio (0.71), Treynor ratio (0.48), VaR (-2.43%), and ES (-4.36%), while portfolio U1 had the highest Alpha (4.82%), and portfolio UMD1 and UMD3 had the highest Information ratios (0.42). The results of adding rebalancing costs are presented in Appendix B. As Table B.1 and B.2 illustrates, with a 0.5 percent rebalancing cost, portfolios U1-U3, D1-D2, and UMD3 all outperformed their respective benchmark both in total- and risk-adjusted terms. With a 1 12

17 Table 5.2 Risk-adjusted portfolio returns. Portfolio Sharpe Ratio (vs. Benchmark) U1 (Low β + ) 0.58 (0.09) U (0.07) U (0.13) U (-0.09) U5 (High β + ) 0.35 (-0.14) Benchmark (U) 0.49 D1 (Low β - ) 0.71 (0.22) D (0.12) D (-0.01) D (-0.08) D5 (High β - ) 0.34 (-0.15) Benchmark (D) 0.49 UMD1 (Low β + - β - ) 0.42 (-0.07) UMD (0.01) UMD (0.06) UMD (0.02) UMD5 (High β + - β - ) 0.44 (-0.05) Benchmark (UMD) Treynor Ratio (vs. Benchmark) 0.43 (0.17) 0.33 (0.07) 0.31 (0.05) 0.20 (-0.05) 0.17 (-0.09) (0.22) 0.34 (0.08) 0.25 (-0.01) 0.21 (-0.05) 0.19 (-0.07) (0.01) 0.28 (0.02) 0.28 (0.02) 0.25 (-0.01) 0.24 (-0.02) Alpha % Information Ratio Weekly VaR(5%) Weekly ES(5%) n/a n/a n/a n/a n/a n/a n/a n/a Note: the table shows descriptive statistics for each of the five upside- (U), downside- (D) and upside-minus-downside (UMD) beta portfolios, as well as their respective benchmark and the, over the out-of-sample period week 1 of 1998 to week 26 of It includes the annual Sharpe ratio, Treynor ratio, Jensen s alpha, and Information ratio, as well as the weekly five percent VaR and ES. Significance levels are not reported due to a lack of symmetry in the data as shown in Table C.1 to C.3 in Appendix C. The highest values are highlighted in bold. 13

18 percent rebalancing cost, as illustrated in Table B.3 and B.4, portfolios U1, U3, and D1-D2 still outperformed their respective benchmark in both total- and risk- adjusted terms, while portfolio U2 only did so in risk-adjusted terms. Finally, with a 2 percent rebalancing cost, as illustrated in Table B.5 and B.6, portfolio D1 was the only portfolio outperforming its benchmark in either total- or risk-adjusted terms. The time-series of each portfolio with transaction costs included are presented in Figure B.1 to B.9. A Shapiro-Wilks test for normality, as well as tests for the skewness and kurtosis of each portfolio, are presented in Table C.1 to C.3 in Appendix C. As Table C.1 and C.2 illustrate, none of the portfolios, or the residuals of their excess returns, follow a normal distribution. This is further illustrated by the skewness and kurtosis statistics in Table C.3, which also show that there is no visible pattern between a portfolio s quintile number or performance, and its level of skewness and kurtosis. Most portfolios have a high level of skewness and kurtosis, although the returns of portfolios U5, D5 and UMD1 do seem to be symmetric as the skewness is not significantly different from zero. Overall, portfolio D1 has the highest skewness (-0.96), and portfolio U2 has the highest level of kurtosis (10.27) compared to a normal distribution. 14

19 6. Discussion The results indicate that lower upside- and downside beta portfolios, as well as portfolios with a more moderate difference between the two measures, tended to perform better. This is not in line with theory, or previous research, for several reasons. First, the upside beta portfolios were expected to perform better the higher their upside beta since, according to the conditional CAPM, they are sensitive to upside market risk, and the market has mainly had a positive return over the sample period. Furthermore, Guy (2015) argues that one reason his Target Date fund, which partially consisted of stocks selected solely by their high upside beta, performed well was because it exposed investors to a lot of upside risk. However, since he did not evaluate pure stock portfolios based on different upside betas, it is difficult to determine whether the high upside beta stocks in his Target Date fund performed better on average than the low upside beta stocks. Second, and although one may argue that low downside beta portfolios should perform better since they offer a hedge during market downturns, the results are not in line with the findings of Ang, Chen, and Xing (2005). They found that high downside beta portfolios generated higher arithmetic average returns since investors demand an additional risk-premium for holding downside risk. The results of this study, on the other hand, suggest that there are no major differences between the arithmetic averages of the downside beta portfolios, but that their geometric averages differ substantially and are inversely related to their downside beta. Third, since a high upside-minus-downside beta indicates a high upside risk relative the downside risk, portfolios with a large beta difference were expected to perform better, which is not the case according to the results. However, since neither the upside- or downside beta portfolios performed according to theory, it is not surprising that the upside-minus-downside beta portfolios did not either. Portfolios with a high beta difference will by construction have a large upside beta relative their downside beta, which may result in a low average return since the negative effect from a high upside beta cancels out the positive effect from a low downside beta. The same reasoning applies for portfolios with a low upside-minus-downside beta. For all three portfolio categories, the performance could potentially be explained by the traditional CAPM beta, which is inversely related to the returns. Although such a relationship is not predicted by the CAPM, it is a pattern well-documented by previous research, and could potentially be explained by behavioural factors such as lottery-ticket preferences (see Baker, Bradely & Wurgler, 2011; Bali, Brown, Murray & Tang, 2014). This would increase the relative demand for high beta portfolios, lowering their alphas, which is a pattern observed in the results. 15

20 The high beta portfolios of this study could therefore have a higher market risk premium than low beta portfolios, but their low alpha, combined with the observed volatility drain, could have lowered their realized returns enough for the low beta portfolios to outperform in both totaland risk-adjusted terms. Beta values extreme by chance could also be part of the explanation since stocks with a large difference between their estimated- and real beta are more likely to end up in the top- or bottom quintiles. The results of adding transaction costs indicate that the method of selecting stocks based on their upside-, downside-, or upside-minus-downside beta would have been profitable for an actual investor over the sample period. Even with a one percent rebalancing cost, several portfolios outperformed their respective benchmark in both total- and risk-adjusted terms. Portfolio D1 also outperformed its benchmark with a two percent rebalancing cost, although marginally in absolute terms. The high excess returns, combined with the large number of observations, also indicate that the results were not driven by mere chance and may persist some time into the future. Although the results seem to be largely driven by the portfolios traditional CAPM beta, the fact that all 15 portfolios had a negative upside-minus-downside beta, and that the returns of the upside- and downside beta portfolios differed, suggest that their market risk is asymmetric. However, the results of the skewness and kurtosis tests indicate that there are no distinct patterns between the quintile numbers and the skewness or kurtosis of the portfolios. The method of selecting stocks would therefore be inefficient in creating portfolios with a low (negative) skewness and kurtosis. The results rather suggest that the main difference between the return distributions of the portfolios is that they are located differently alongside each other. Overall, the findings are not in line with the traditional CAPM, which states that a financial asset should earn a risk-premium that is proportional to its market risk. They also suggest that an upside-downside risk asymmetry may need to be incorporated into the model for more efficient asset pricing, supporting the results of previous research. The credibility of these findings is affected by several factors. First, by including both listed and delisted stocks in the analysis, the sample size was increased, and the risk of survivorship bias was reduced, better reflecting the contemporary market faced by a Swedish investor. Second, no statistical tests of significance were performed since the necessary assumptions were not met, lowering the risk of incorrect inferences. Instead, several widely used performance statistics were calculated to 16

21 provide a broad overview of each portfolio s performance. Third, because the portfolio values did not follow a geometric Brownian motion, the standard method of scaling asset returns over time would underestimate the actual risk and was therefore not used. Instead, the Hurst exponent was estimated using rescaled range analysis, and although the method may result in some bias (see Appendix D), the annualized returns will be more accurate than under the assumption of a geometric Brownian motion. Finally, all market betas were estimated using a single-factor model, resulting in other risk factors being omitted. Ang, Chen & Xing (2005) included stocks historical market capitalization and book-to-market values as proxies for size and value. Attempts were made to control for additional risk factors, but data over these variables were inconsistent both over time and across stocks, which would substantially reduce the sample size and shorten the out-of-sample period. 17

22 7. Conclusion The purpose of this study was to investigate if it is possible to achieve a higher risk-adjusted return by selecting stocks based on their upside- and downside market risk, as measured by their CAPM beta conditioned on positive and negative market returns. Using weekly Swedish stock data from to , and the as a proxy for the market portfolio, upside- and downside stock betas were estimated over a rolling two-year period using a singlefactor model. The stocks were ranked lowest to highest according to their upside-, downside-, and upside-minus-downside beta at the end of each two-year period and divided into quintiles, resulting in fifteen equally weighted portfolios. Each portfolio was rebalanced every six-months and tested out-of-sample against an equally weighted benchmark index consisting of all eligible stocks for that portfolio. The results indicate that although an upside-downside beta asymmetry seems to exist, the portfolios performance seems to be largely explained by the traditional CAPM beta rather than a market risk asymmetry. There are two main conclusions from this study. First, the fact that all portfolios had a downside beta greater than their upside beta indicates that the market risk is asymmetric. This contradicts the traditional CAPM and suggests that a risk asymmetry may need to be incorporated into the model for more efficient asset pricing. Even if the skewness and kurtosis tests indicate that the method of selecting stocks was unable to produce portfolios with a low skewness and kurtosis, it may still be possible to exploit the risk asymmetry during portfolio construction, although with a different method. Second, whether the results of the low beta portfolios were driven by behavioral- or other effects, their high excess returns, combined with the large number of observations, indicate that the results were not driven by mere chance and may persist some time into the future. A Swedish investor could therefore relatively easy construct stock portfolios more efficient than the market using this approach of investing. Since the results seem to be largely driven by the stocks traditional CAPM beta, a suggestion for future research would be to attempt controlling for the traditional beta when constructing financial portfolios based on market risk asymmetries. In addition, stocks that are more likely to yield inaccurate beta estimates, i.e. stocks with infrequent data over the in-sample period, could be excluded to reduce their stochastic effect on the results. The remaining stocks could also be weighted according to their beta values to investigate if it would increase the portfolio returns further. 18

23 References Ang A., Chen J. & Xing Y. (2005). Downside Risk. The National Bureau of Economic Research. Working Paper No , pp Baker, M., Bradley, B. & Wurgler, J. (2011). Benchmarks as Limits to Arbitrage: Understanding the Low-Volatitlity Anomaly. Financial Analysts Journal. 67(1), pp Bali, T. G., Brown, S. J., Murray, S. & Tang, Y. (2014). Betting Against Beta or Demand for Lottery? Georgetown McDonough School of Business. Research Paper No , pp Barinov, A. (2014). Trading Strategies and Financial Models. Cumming: Conley Smith epublishing, LLC. Bassingthwaighte, J. B. & Raymond, G. M. (1994). Evaluating Rescaled Range Analysis for Time Series. Annals of Biomedical Engineering. 22(4), pp Bawa, V. J. & Lindenberg, E. B. (1977). Capital Market Equilibrium in a Mean-Lower Partial Moment Framework. Journal of Financial Economics. 5(2), pp Bloomberg L.P. (2017). Bloomberg Terminal (Version 08/04/17) [software]. Available: Blume, M. E. (1971). On the Assessment of Risk. The Journal of Finance. 26(1), pp Bodie, Z., Kane, A. & Marcus, A.J. (2012). Investments. 10 th ed., Berkshire: McGraw-Hill Education. Ellis, C. & Hudson, C. (2007). Scale-Adjusted Volatility and the Dow Jones index. Physica A. 378(2), pp Fabozzi, F. J. & Francis, J. C. (1978). Beta as a Random Coefficient. Journal of Financial and Quantitative Analysis. 13(1), pp

24 Fama, E. F. (1965). The Behaviour of Stock Market Prices. Journal of Business. 38(1), pp Forbes, K. & Rigobon, R. (2002). No Contagion, Only Interdependence: Measuring Stock Market Co-movements. Journal of Finance. 57(5), pp Ghysels, E. (1998). On Stable Factor Structures in the Pricing of Risk: Do Time-Varying Betas Help or Hurt? Journal of Finance. 53(2), pp Guy, A. (2015). Upside and Downside Beta Portfolio Construction: A Different Approach to Risk Measurement and Portfolio Construction. Risk governance and control: financial markets & institutions. 5(4), pp Hamed, K. H. (2006). Improved Finite-Sample Hurst Exponent Estimates Using Rescaled Range Analysis. Water Resources Research. 43(4), pp Hull, J. C. (2014). Options, Futures, and Other Derivatives. 8 th ed., Harlow: Pearson Education Limited. King, M. A. & Wadwhani, S. (1990). Transmission of Volatility Between Stock Markets. Review of Financial Studies. 3(1), pp Lintner, J. (1965). The Valuation of Risk Assets and the Selection of Risky Investments in Stock Portfolios and Capital Budgets. Review of Econometrics and Statistics. 47(1), pp Longin, F. & Solnik, B. (1995). Is the Correlation in International Equity Returns Constant: ? Journal of International Money and Finance. 14(1), pp Mandelbrot, B. B. (1963). The Variation of Certain Speculative Prices. The Journal of Business. 36(4), pp Mandelbrot, B. B. & Wallis, J. R. (1969). Computer Experiments with Fractional Gaussian Noises. Water Resources Research. 5(1), pp

25 Markowitz, H. (1952). Portfolio Selection. The Journal of Finance. 7(1), pp Markowitz, H. (1959). Portfolio Selection: Efficient Diversification of Investments. New York: John Wiley & Sons, Inc. Momcilovic, M., Begovic, S. V. & Tomasevic, S. (2014). Influence of Return Interval on Stock s Beta. Advances in Economics, Law and Political Science. pp Mossin, J. (1966). Equilibrium in a Capital Asset Market. Econometrica. 34(4), pp Nasdaq. (2016). Regelverk för Emittenter. Stockholm: Nasdaq. Available: Sharpe, W. F. (1964). Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk. Journal of Finance. 19(3), pp

26 Appendix A Table A.1 Portfolio beta values. Portfolio β β + β - β + - β - U1 (Low β + ) U U U U5 (High β + ) Benchmark (U) D1 (Low β - ) D D D D5 (High β - ) Benchmark (D) UMD1 (Low β + - β - ) UMD UMD UMD UMD5 (High β + - β - ) Benchmark (UMD) Note: the table shows beta estimates for each of the five upside- (U), downside- (D), and upside-minus-downside (UMD) beta portfolios, as well as for their respective benchmark, over the sample period week 1 of 1998 to week 26 of It includes the CAPM-, upside-, downside-, and upside-minus-downside beta. The was used as a proxy for the market portfolio. 22

27 Figure A.1 Time-series of upside beta portfolios U1 (Low β + ) U2 U3 Benchmark (U) Benchmark (U) Benchmark (U) U4 U5 (High β + ) Benchmark (U) Benchmark (U) Note: the figure shows the time-series of each upside beta portfolio (U), the upside beta benchmark portfolio, and the over the out-of-sample period week 1 of 1998 to week 26 of Each time-series has been indexed at 100 in week 52 of Figure A.2 Time-series of downside beta portfolios D1 (Low β - ) D2 D3 Benchmark (D) Benchmark (D) Benchmark (D) D4 D5 (High β - ) Benchmark (D) Benchmark (D) Note: the figure shows the time-series of each downside beta portfolio (D), the downside beta benchmark portfolio, and the over the out-of-sample period week 1 of 1998 to week 26 of Each time-series has been indexed at 100 in week 52 of

28 Figure A.3 Time-series of upside-minus-downside beta portfolios UMD1 (Low (β + - β - )) UMD2 UMD3 Benchmark (UMD) Benchmark (UMD) Benchmark (UMD) UMD4 UMD5 (High (β + - β - )) Benchmark (UMD) Benchmark (UMD) Note: the figure shows the time-series of each upside-minus-downside beta portfolio (UMD), the upside-minus-downside beta benchmark portfolio, and the over the out-of-sample period week 1 of 1998 to week 26 of Each time-series has been indexed at 100 in week 52 of

29 Appendix B Table B.1 Descriptive statistics over annual portfolio returns with a 0.5% rebalancing cost. Portfolio Ending value of 1SEK Geometric Average % Arithmetic Average % Standard Deviation % Tracking Error % U1 (Low β + ) U U U U5 (High β + ) Benchmark (U) n/a D1 (Low β - ) D D D D5 (High β - ) Benchmark (D) n/a UMD1 (Low β + - β - ) UMD UMD UMD UMD5 (High β + - β - ) Benchmark (UMD) n/a n/a Note: the table shows descriptive statistics for each of the five upside- (U), downside- (D) and upside-minus-downside (UMD) beta portfolios, as well as their respective benchmark, and the over the out-of-sample period week 1 of 1998 to week 26 of 2017 with a 0.5 percent rebalancing cost included. It includes the ending value of 1SEK invested at the beginning of the period, the geometric- and arithmetic averages, the standard deviation, and the tracking error. The values have been annualized from a weekly basis. 25

30 Table B.2 Risk-adjusted annual portfolio returns with a 0.5% rebalancing cost. Portfolio Sharpe Ratio (vs. Benchmark) U1 (Low β + ) 0.55 (0.06) U (0.03) U (0.04) U (-0.12) U5 (High β + ) 0.33 (-0.16) Benchmark (U) 0.49 D1 (Low β - ) 0.67 (0.18) D (0.08) D (-0.04) D (-0.11) D5 (High β - ) 0.32 (-0.17) Benchmark (D) 0.49 UMD1 (Low β + - β - ) 0.40 (-0.09) UMD (-0.02) UMD (0.03) UMD (-0.02) UMD5 (High β + - β - ) 0.42 (-0.07) Benchmark (UMD) Treynor Ratio (vs. Benchmark) 0.41 (0.12) 0.31 (0.05) 0.29 (0.03) 0.18 (-0.07) 0.16 (-0.10) (0.18) 0.32 (0.06) 0.24 (-0.02) 0.20 (-0.06) 0.18 (-0.10) (-0.03) 0.27 (0.01) 0.26 (0.02) 0.24 (-0.01) 0.22 (-0.04) Alpha % Information Ratio Weekly VaR(5%) Weekly ES(5%) n/a n/a n/a n/a n/a n/a n/a n/a Note: the table shows descriptive statistics for each of the five upside- (U), downside- (D) and upside-minus-downside (UMD) beta portfolios, as well as their respective benchmark, and the over the out-of-sample period week 1 of 1998 to week 26 of 2017 with a 0.5 percent rebalancing cost included. It includes the annual Sharpe ratio, Treynor ratio, Jensen s alpha, and Information ratio as well as the weekly five percent VaR and ES. Significance levels are not reported due to a lack of symmetry in the data as shown in Table C.1 to C.3 in Appendix C. The highest values are highlighted in bold. 26

31 Table B.3 Descriptive statistics over annual portfolio returns with a 1% rebalancing cost. Portfolio Ending value of 1SEK Geometric Average % Arithmetic Average % Standard Deviation % Tracking Error % U1 (Low β + ) U U U U5 (High β + ) Benchmark (U) n/a D1 (Low β - ) D D D D5 (High β - ) Benchmark (D) n/a UMD1 (Low β + - β - ) UMD UMD UMD UMD5 (High β + - β - ) Benchmark (UMD) n/a n/a Note: the table shows descriptive statistics for each of the five upside- (U), downside- (D) and upside-minus-downside (UMD) beta portfolios, as well as their respective benchmark, and the over the out-of-sample period week 1 of 1998 to week 26 of 2017 with a 1 percent rebalancing cost included. It includes the ending value of 1SEK invested at the beginning of the period, the geometric- and arithmetic averages, the standard deviation, and the tracking error. The values have been annualized from a weekly basis. 27