A Second Look at the Role of Hedge Funds In a Balanced Portfolio The CFA Society of Victoria Victoria, BC September 21st, 2010 Jean L.P. Brunel, C.F.A Three main points … A highly heterogeneous universe Different optimization needs What about leverage A highly heterogeneous universe The term hedge fund is misleading as it does not cover a well-defined universe. Rather, it describes many differing strategies … A very wide risk spectrum Justified by a wide variety of strategies Looking for a better classification Recognizing differing return distributions A wide risk spectrum … Last 5 Year Data - Risk/Return Scatter 20.00% Average Returns Does this look as one set of strategies or quite a number of different ones? 15.00% 10.00% 5.00% 0.00% -5.00% -10.00% 0.00% 5.00% 10.00% 15.00% Volatility of Returns 20.00% 25.00% A wide risk spectrum … Last 15 Year Data - Risk/Return Scatter 20.00% 18.00% 16.00% Average Returns Moving from a 5-year to a 15-year analysis does not really change the picture that much … 14.00% 12.00% 10.00% 8.00% 6.00% 4.00% 2.00% 0.00% 0.00% 5.00% 10.00% 15.00% Volatility of Returns 20.00% 25.00% What do these managers do? Convertible Merger/Risk Statistical Fixed Income Pair Trades Market Neutral Equity Long/Short Sector Leverage Implied Leverage Implied Leverage Concentrated Portfolios Global Macro Managed Futures Leverage Concentration Model Market Market Market Valuation Valuation Valuation Valuation Model Model Model Model Return volatility < 6% Return volatility > 6% There seems to be two clusters … Last 5 Year Data - Risk/Return Scatter Absolute Return Strategies in Orange 20.00% 15.00% Average Returns It looks as if one can classify the various strategies according to whether they take fixed income- or equity-type risks … 10.00% 5.00% 0.00% -5.00% -10.00% 0.00% 5.00% 10.00% 15.00% Volatility of Returns 20.00% 25.00% There seems to be two clusters … Last 15 Year Data - Risk/Return Scatter Absolute Return Strategies in Orange 20.00% 18.00% Average Returns The 15-year picture confirms the insights gained from the shorter term time horizon … 16.00% 14.00% 12.00% 10.00% 8.00% 6.00% 4.00% 2.00% 0.00% 0.00% 5.00% 10.00% 15.00% Volatility of Returns 20.00% 25.00% These clusters make sense … An analysis of risk and return history within traditional and nontraditional clusters shows the grouping makes sense … The fixed income cluster makes sense: o absolute return and bonds: similar volatility o despite at times differing returns The equity cluster similarly makes sense: Cluster Risk/Return Averages Absolute Return Cluster Traditional Fixed Income Cluster Semi-Directional Cluster Traditional Equity Cluster Last 5 Years Return Volatility 7.49% 3.89% 7.15% 3.76% 6.91% 12.48% 9.85% 16.95% Last 15 Years Return Volatility 10.24% 4.63% 5.53% 4.31% 13.68% 13.86% 5.73% 17.20% In short … The term hedge fund is misleading as it does not cover a well-defined universe. Rather, it describes many differing strategies … The universe is indeed highly heterogeneous The strategy risk spectrum is very wide … … because managers do very different things It makes sense to classify hedge funds as: o o those that look like fixed income those that look like equities … and use that to build balanced portfolios Three main points … A highly heterogeneous universe Different optimization needs What about leverage Important differences … The returns on nontraditional strategies are often not normally distributed… Traditional returns are normally distributed That is not true for non-traditional returns: o o often showing a negative skew often substantial excess kurtosis Same return and volatility, and yet … The high “manager” risk incurred in nontraditional strategies disturbs the normal distributions we would typically expect … 60 50 Both means = 0.84% Arbitrage = 1.29% Normal = 1.22% Arbitrage Normal Distribution 40 30 20 10 0 -4.15% -10 -1.66% 0.84% 3.34% 5.83% First, consider negative skew … Negative skew means more points right of the mean, but also a wider range on the left (i.e. down) side of it as well 60 50 Both means = 0.84% Arbitrage = 1.29% Normal = 1.22% Arbitrage Normal Distribution 40 Arbitrage Skew = -2.71 Normal Skew = -0.10 30 20 10 0 -4.15% -10 -1.66% 0.84% 3.34% 5.83% Then, how about excess kurtosis? Excess kurtosis mean that the return distribution is “peaky” and that it has “fat tails” … 60 50 Both means = 0.84% Arbitrage = 1.29% Normal = 1.22% Arbitrage Normal Distribution 40 Arbitrage Skew = -2.71 Normal Skew= -0.10 30 Arbitrage Kurtosis = 9.73 Normal Kurtosis = -0.25 20 10 0 -4.15% -10 -1.66% 0.84% 3.34% 5.83% In plain English … A look at third and fourth statistical moments helps make sense of the high Sharpe ratio of nontraditional strategies … Strategies combining: o o negative skew and more highly positive kurtosis Have a higher risk of bad surprises: Which must be “compensated” by either: o o higher expected returns, or lower expected return volatility Which mean-variance optimization misses … Traditional optimization results … The traditional meanvariance model overallocates to absolute return strategies and ignores bonds … Note the very low allocations to bonds: Fixed Income - Like Universe Expected Return Expected Risk 4.53% 0.56% 6.59% 1.02% 9.21% 2.02% 11.63% 3.02% 11.85% 3.13% Target Risk 0.56% 1.00% 2.00% 3.00% 4.00% Portfolio Composition Cash Bonds Absolute Return Strategies Total 100% 0% 0% 100% 70% 3% 27% 100% 31% 9% 60% 100% 0% 5% 95% 100% 0% 0% 100% 100% Traditional optimization results … Similarly, it totally ignores traditional equities to “pile” into equity hedge strategies, despite the tail risk … Note the lack of allocation to traditional equities Equity - Like Universe 15.63% 8.58% 16.65% 9.00% 16.65% 9.05% 16.65% 9.05% 16.65% 9.05% Target Risk 8.58% 9.00% 10.00% 11.00% 12.00% Portfolio Composition Equity Equity Hedge Equity Non-Hedge Managed Futures Global Macro Total 0% 0% 0% 0% 100% 100% 0% 99% 0% 0% 1% 100% 0% 100% 0% 0% 0% 100% 0% 100% 0% 0% 0% 100% 0% 100% 0% 0% 0% 100% Expected Return Expected Risk Let us try and experiment … A simple experiment will helps us set early ground rules Let’s divide fixed income market history: o o periods when bond returns were positive periods when bond returns were negative Let’s divide equity market history: o o o periods when returns were high periods when returns were “normal” periods when returns were low Let’s re-run the traditional optimization: Traditional optimization results … In periods when bond returns are positive, a meanvariance optimization model will not shun bonds … Note that the model CAN allocate to bonds: Bond Returns Positive Fixed Income - Like Universe Expected Return Expected Risk 4.58% 0.60% 8.08% 1.02% 13.00% 2.03% 13.52% 2.57% Target Risk 0.60% 1.00% 2.00% 3.00% Portfolio Composition Cash Bonds Absolute Return Strategies Total 100% 0% 0% 100% 60% 26% 14% 100% 3% 62% 35% 100% 0% 100% 0% 100% Traditional optimization results … In periods when bond returns are negative, a meanvariance optimization model will seemingly shun bonds Note also that the model can ignore bonds: Bond Returns Negative Fixed Income - Like Universe Expected Return Expected Risk 4.16% 0.72% 4.96% 1.00% 6.09% 2.00% 7.10% 3.00% 8.07% 4.00% 8.87% 4.83% Target Risk 0.72% 1.00% 2.00% 3.00% 4.00% 5.00% Portfolio Composition Cash Bonds Absolute Return Strategies Total 100% 0% 0% 100% 83% 0% 17% 100% 60% 0% 40% 100% 38% 0% 62% 100% 17% 0% 83% 100% 0% 0% 100% 100% Traditional optimization results … In periods when equity returns are high, a meanvariance optimization model will not shun traditional equities … Note that the model CAN allocate to equities: S&P 500 Greater than 1.17% Equity - Like Universe 32.55% 6.76% 22.54% 7.75% 42.98% 8.75% 42.83% 8.94% 46.93% 7.58% Target Risk 6.76% 7.75% 8.75% 9.75% 10.75% Portfolio Composition Equity Equity Hedge Equity Non-Hedge Managed Futures Global Macro Total 63% 0% 0% 37% 0% 100% 37% 0% 0% 63% 0% 100% 3% 0% 97% 0% 0% 100% 0% 0% 100% 0% 0% 100% 100% 0% 0% 0% 0% 100% Expected Return Expected Risk Traditional optimization results … In periods when bond returns are normal, the meanvariance optimization model seems to ignore traditional equities … The model mostly ignores equities: S&P 500 Between 0.00% 1.17% Equity - Like Universe Expected Return Expected Risk 8.51% 1.02% 15.17% 3.50% 18.29% 6.00% 17.45% 8.50% 16.94% 10.60% Target Risk 1.02% 3.50% 6.00% 8.50% 11.00% Portfolio Composition Equity Equity Hedge Equity Non-Hedge Managed Futures Global Macro Total 100% 0% 0% 0% 0% 100% 20% 68% 10% 0% 2% 100% 0% 0% 76% 0% 24% 100% 0% 0% 25% 0% 75% 100% 0% 0% 0% 0% 100% 100% Traditional optimization results … In periods when equity returns are negative, the model does not want to hear about them … The model still ignores equities: S&P 500 Negative Equity - Like Universe Expected Return Expected Risk 1.08% 7.37% -12.02% 8.49% -16.21% 9.75% -22.31% 11.00% -27.83% 12.24% Target Risk 7.37% 8.48% 9.73% 10.98% 12.23% Portfolio Composition Equity Equity Hedge Equity Non-Hedge Managed Futures Global Macro Total 0% 0% 0% 0% 100% 100% 0% 0% 44% 0% 56% 100% 0% 58% 42% 0% 0% 100% 0% 30% 70% 0% 0% 100% 0% 6% 94% 0% 0% 100% What have we learned? The optimizer does not like losses!!! It can allocate to bonds: o When they offer competitive returns o But not when they are “normal” It can allocate to equities: o When they offer competitive returns o Or when they are the lowest risk choice These strategies do not always make sense Let us try a final experiment … Though this experiment is not a “solver,” but a calculator, it can help demonstrate the power of a more detailed model … Mean-variance optimization only uses: o o return and risk expectations, and … … covariance among each pair of assets Let’s design a different model: o o o o return and risk observations skew and kurtosis observations” implicit preferences for skew and kurtosis the same covariance matrix Let’s re-run the optimization: The goals for that model would be ... Rather than focusing on meanvariance, we calculate a “Z-Score” which incorporates all four moments … On the one hand: o o to capture as much return as possible while avoiding as much risk as possible At the same time, we would like: o o o to minimize the risk of negative surprises minimizing negative skew” minimizing excess kurtosis In “Greek” our “Z-Score” will be: o Max (E[r] - + l*skew - g*Kurtosis) Z-Score fixed income optimization: This model produces results that ignore absolute return strategies if the aversion to manager risk is set at a high level The model ignores absolute return strategies: Fixed Income - Like Universe landg0.01) Monthly Data Return Volatility Skew Kurtosis 0.37% 0.16% -0.24 -0.57 0.46% 0.41% -0.46 1.01 0.52% 0.66% -0.47 0.90 0.55% 0.81% -0.47 0.84 0.63% 1.13% -0.46 0.75 Target Risk 0.16% 0.41% 0.66% 0.81% 1.13% Portfolio Composition Cash Bonds Absolute Return Strategies Total 100% 0% 0% 100% 67% 33% 0% 100% 44% 56% 0% 100% 30% 70% 0% 100% 0% 100% 0% 100% Z-Score fixed income optimization: These results are much more intuitively satisfying, with a better balance between traditional and nontraditional strategies .. With a lesser manager risk aversion, the model allocates to absolute return strategies: Fixed Income - Like Universe landg= 0.005 Monthly Data Return Volatility Skew Kurtosis 0.37% 0.16% -0.24 -0.57 0.46% 0.41% -0.46 1.01 0.52% 0.66% -0.47 0.90 0.79% 0.81% -0.57 0.47 0.63% 1.13% -0.47 0.75 Target Risk 0.16% 0.41% 0.66% 0.81% 1.13% Portfolio Composition Cash Bonds Absolute Return Strategies Total 100% 0% 0% 100% 67% 33% 0% 100% 44% 56% 0% 100% 0% 55% 45% 100% 0% 100% 0% 100% A much better potential formulation This model has the potential to address our problem, but it still needs to be tested on balanced portfolios ... Neil Davies, Harry Kat and Sa Lu have proposed an interesting “solver” formulation: Minimize a b g Z (1 +d1 ) +(1 +d3 ) +(1 +d4 ) , Subject to E[X R] +xn +1 t +d1 Z1* , T T ~ ~ ~ 3 * E{X (R -E[R])} +d3 Z3 , ~ ~ -E{X T (R -E[R])}4 +d4 -Z4* , d1 , d3 , d4 0, T X VX 1; X 0; T xn +1 1 -I X Three main points … A highly heterogeneous universe Different optimization needs What about leverage Naïve expectations for L/S … We can dispense with the detailed analysis of statistical results and rather look at how similar or not these are to naïve expectations If systematic leverage is the key, on should o Find a relatively high R Square o A Beta coefficient greater than 1 o A negative Alpha coefficient Equity L/S vs. equity indexes … In fact, the R Squares are relatively low, the betas are very low and significant and the alphas are all positive and significant … 1995-2007 R Square Coefficient t-stat Coefficient t-stat Alpha t-stat Russell 3000 0.566 S&P 500 0.469 Russell 2000 0.753 0.4615 13.6955 0.4228 11.2878 0.4098 20.949 0.0075 5.202 0.008 4.979 0.0079 7.2946 S&P 500 + Russell 2000 0.762 0.084 2.3415 0.364 13.2614 0.0075 6.9984 Equity L/S vs. equity indexes … Again, the R Squares are relatively low, the betas are very low and significant and the alphas are all positive and significant … 2002-2007 R Square Coefficient t-stat Coefficient t-stat Alpha t-stat Russell 3000 0.637 S&P 500 0.567 Russell 2000 0.807 0.3857 10.0114 0.3672 8.6322 0.3167 15.4409 0.0048 3.4208 0.0053 3.4322 0.0041 3.9655 S&P 500 + Russell 2000 0.808 0.0282 0.5705 0.3001 8.3979 0.0041 3.9448 Leverage and manager alpha … Now the idea is to test the alpha of managers in rising and falling markets against the benchmark Managers can add value in two ways: o o Market timing: varying market exposure Bottom up security selection If managers are great market timers: o o positive and strong correlation in up markets negative and equally strong in down markets Caveat: multiple sources of alpha … In rising markets… Whatever relationship there is does appear quite weak and in the wrong direction: managers find it harder to add value in up markets… R Square Coefficient t-stat Coefficient t-stat Alpha t-stat Rising Markets – Russell 3000 1995-2007 S&P 500 + Russell Russell Russell S&P 500 3000 2000 2000 0.031 0.067 0.056 0.153 -0.1169 -1.7206 -0.1627 -2.6047 0.0995 2.3503 0.0091 3.2867 0.0106 4.1174 0.0013 0.5784 -0.2004 -3.2803 0.1265 3.0761 0.0069 2.5345 In rising markets… Whatever relationship there is now appear a bit stronger, but still weak and in the wrong direction … R Square Coefficient t-stat Coefficient t-stat Alpha t-stat Rising Markets – Russell 3000 2002-2007 S&P 500 + Russell Russell Russell S&P 500 3000 2000 2000 0.324 0.361 0.086 0.375 -0.2598 -4.2713 -0.2702 -4.6341 -0.087 -1.8962 0.0069 3.3859 0.0067 3.5639 0.0032 1.4415 -0.3144 -4.1332 0.0454 0.9082 0.0061 3.0766 How about falling markets? There appears to be virtually no relationship in view of the very low R Squares, and the direction is mostly wrong… R Square Coefficient t-stat Coefficient t-stat Alpha t-stat Falling Markets – Russell 3000 1995-2007 S&P 500 + Russell Russell Russell S&P 500 3000 2000 2000 0.000 0.000 0.008 0.186 0.0016 0.0255 -0.0078 -0.1380 0.0207 0.6137 0.0058 2.0665 0.0061 2.3723 0.0051 2.7529 -0.1591 -3.2775 0.0646 1.8333 0.0031 1.6127 How about falling markets? There appears to be a bit more of a relationship (still weak though) and the sign is in the right direction at least… R Square Coefficient t-stat Coefficient t-stat Alpha t-stat Falling Markets – Russell 3000 2002-2007 S&P 500 + Russell Russell Russell S&P 500 3000 2000 2000 0.2640 0.3100 0.0250 0.3320 -0.1660 -2.6818 -0.1709 -2.9949 -0.0407 -0.7215 0.0017 0.6915 0.0017 0.7166 0.0052 1.818 -0.1981 -2.9500 0.0441 0.7895 0.0026 0.9844 Naïve expectations for A/R … Though the test variables will be different, the naïve expectations we form are the same as in the case of long/short managers … If systematic leverage is the key, one should o Find a relatively high R Square o A Beta coefficient greater than 1 o A negative Alpha coefficient Absolute return vs. benchmarks … In fact, the R Squares are quite low, the betas are very low and mostly significant and the alphas are all positive and significant … R Squared Russell 3000 0.358 Coefficient t-stat Alpha t-stat 0.1248 8.9554 0.0076 12.6132 R Squared Russell 3000 0.356 Coefficient t-stat Alpha t-stat 0.1264 5.7537 0.0055 6.9108 90 Day Treasuries 0.064 1.457 3.1254 0.0039 2.2366 90 Day Treasuries 0.024 0.9158 1.2044 0.0043 2.3284 1995-2007 Salomon Merrill BIG High Yield 0.001 0.306 0.0221 0.3242 0.0087 10.5385 0.2498 7.9606 0.0071 11.1245 2002-2007 Salomon Merrill BIG High Yield 0.003 0.376 -0.0371 -0.4182 0.0064 6.1611 0.2331 6.0122 0.0043 5.2466 Average HY Spreads 0.038 -0.0007 -2.3737 0.0128 7.0065 Average HY Spreads 0.096 -0.0009 -2.5263 0.011 5.2646 Is there a static mix? We can test this by looking at whether absolute return strategy returns can be regressed against the same variables… 1995-2007 R Square Intercept Russell 3000 90-Day T. Bills Salomon BIG Merrill High Yield Average HY Spread R Square Intercept Russell 3000 90-Day T. Bills Salomon BIG Merrill High Yield 2002-2007 Five Independent Variables R Square 0.503 t-stats Intercept 0.0019 0.9882 Russell 3000 0.0766 5.0432 90-Day T. Bills 1.5497 4.3023 Salomon BIG -0.0749 -1.4093 Merrill High Yield 0.1791 5.2938 Average HY Spread 0 0.095 Four Independent Variables R Square 0.503 t-stats Intercept 0.002 1.5426 Russell 3000 0.0764 5.0865 90-Day T. Bills 1.5473 4.3215 Salomon BIG -0.074 -1.4219 Merrill High Yield 0.1785 5.3652 0.507 0.0039 0.0702 0.8588 -0.0001 0.1629 -0.0002 t-stats 1.4808 2.5818 1.3929 -0.0017 3.5209 -0.7348 0.502 0.0023 0.0744 1.0448 -0.0078 0.1648 t-stats 1.5944 2.8088 1.8659 -0.1083 3.5824 Leverage and manager alpha … Again, the idea is to test the alpha of managers in rising and falling markets against the benchmark Managers can add value in two ways: o o Market timing: varying market exposure Bottom up security selection If managers are great market timers: o o positive and strong correlation in up markets negative and equally strong in down markets Caveat: multiple sources of alpha … Alphas in rising markets … There is virtually no evident relationship and it is in the wrong direction for half of the variables and often not significant … R Squared Russell 3000 0.033 Coefficient t-stat Alpha t-stat 0.0497 1.8146 0.0068 6.0728 R Squared Russell 3000 0.128 Coefficient t-stat Alpha t-stat 0.0936 2.3603 0.0046 3.4483 Rising Markets 1995-2007 Salomon Merrill BIG High Yield 0.000 0.046 -0.0004 -0.0068 0.0085 11.9825 0.0901 2.1331 0.0073 8.8657 2002-2007 Salomon Merrill BIG High Yield 0.008 0.088 -0.0471 -0.5646 0.0072 8.1597 0.1069 1.9157 0.0056 4.9323 Average HY Spreads 0.009 -0.0003 -0.9205 0.0099 5.9715 Average HY Spreads 0.034 -0.0005 -1.1532 0.0092 4.5032 Alphas in falling markets … There is virtually no evident relationship and it is in the wrong direction more often than not. No statistical significance save HY bds R Squared Russell 3000 0.295 Coefficient t-stat Alpha t-stat 0.1660 4.4806 0.0086 4.9377 R Squared Russell 3000 0.154 Coefficient t-stat Alpha t-stat 0.1168 1.9097 0.0044 1.7806 Falling Markets 1995-2007 Salomon Merrill BIG High Yield 0.000 0.400 -0.0169 -0.1322 0.0027 1.7653 0.2712 5.6596 0.0041 3.8249 2002-2007 Salomon Merrill BIG High Yield 0.119 0.566 0.2436 1.6460 -0.0012 -0.6078 0.2500 5.1078 0.0013 1.1794 Average HY Spreads 0.003 -0.0002 -0.3644 0.0038 1.0720 Average HY Spreads 0.009 -0.0002 -0.4325 0.0024 0.5862 In short … It is hard to substantiate the notion that one can replace nontraditional managers with leveraged long only strategies … In most instances, the observed alpha is: o o not really related to market beta not readily replicable with a static mix More often than not, alpha is: o o Statistically unrelated to market timing The more recent past can be less clear-cut Three main points … A highly heterogeneous universe Different optimization needs What about leverage Questions? A Second Look at the Role of Hedge Funds In a Balanced Portfolio The CFA Society of Victoria Victoria, BC September 21st, 2010 Jean L.P. Brunel, C.F.A