The authors are using a Global Tactical Asset Allocation strategy to tilt their portfolio toward asset classes that are more favorable. Typically this is done through a "building block" model, where each of the asset classes uses a different model to build the overall allocation strategy. There are a few problems with this: securities within asset classes aren't compared to securities in other asset classes, it takes a lot of time to build all of the different models, and there has to be a good risk-management process in place to mitigate unnecessary risk. As such, the authors propose a single model, they name the Global Tactical Cross-Asset Allocation (GTCAA) strategy, that is not subject to those limitations. This allocation approach selects asset classes, rather than securities within asset classes, as has been done in previous studies. If they find this approach to work, it could be a challenge to market efficiency.
The authors then summarize the results of a few previous studies in favor of momentum strategies; for example, Jegadeesh and Titman's 6-month results, Fama and French 12-1 momentum strategy, and Rouwenhorst's international markets, and Pirron's futures market studies. Based on those previous studies, they will look at the 1-month return strategy, 12-1 momentum strategy, and value strategies across asset classes. In their study, they found these value and momentum strategies to exhibit statistically and economically significant returns of 7-8% over the 1986 - 2007 period. In addition, when they combined the value and momentum strategies, they find excess returns of 12% over the same period. These strategies also outperformed in the 1974-1985 out-of-sample period, overcame transaction costs, and overcame typical risk factors such as the Fama-French and Carhart 4-factor models. These results are significant, because it provides a single model for practitioners to use easily; also they use variables that can be used across all the asset classes.
Data and Methodology
The authors use 12 asset classes; these include 3 US equity, 3 international equity, 3 US bonds, 2 international bonds, and 1 month libor over the 1985 - 2007 period. They selected these asset classes to achieve ease of data retrieval, ease of modeling, liquidity, large capitalization, and lack of correlation with other asset classes in the study. For each of the asset classes, they found the excess returns in local currency and subtracted the local risk-free rate to simulate the return of a typical futures contract. Over the 1985 - 2007 period, emerging markets returned the highest at 10.8% and Japan equity returned the worst at 0.7%. When matched against their standard deviations, all assets seemed to have similar Sharpe ratios, excluding Japan equity which seemed to have very high volatility.
To form portfolios of these asset classes, at the beginning of each month the authors rank them according to their momentum or value scores and put them into quartiles (with 3 asset classes in each quartile). In doing so, they will calculate the average 1-month returns of each quartile as well as the return of the top quartile minus the return of the bottom quartile. They will use the 1-month momentum and 12-1 momentum strategies, and they will use yield measures for the value strategy. In addition, they will use a combination strategy which allocates 25% to the 1-month momentum strategy, 25% to the 12-1 momentum strategy, and 50% to the value strategy.
The authors realize this strategy is somewhat simplistic, so they make a few adjustments to the yields to adjust for risks; otherwise, for example, high yield bonds will most likely always get high allocations over risk-less bonds. So, they subtract 1% yield for government bonds, 2% from US investment grade bonds, 6% from US high-yield bonds, 1% from emerging market equities, and 2% from US REITS.
Main Results
Over the 1986 - 2007 period, each of the individual strategies' top quartiles outperformed the bottom quartiles by 7-8%, and the combination strategy's top quartile outperformed the bottom quartile by 12%. There was also a somewhat monotonic relationship of these returns across the quartiles. Each of these results were statistically significant. The information ratios were about 0.60 for the individual strategies and 1.19 for the combination strategies. As would be expected, the returns of the momentum strategies were positively correlated with each other; however, the momentum strategies were negatively correlated with the value strategy. And that diversification is why the combined strategy performs so well.
The authors also look at a chart of the returns for each strategy over time. Each of the 3 individual strategies perform similarly over the period; however, the combination strategy's return was stable and significantly outperformed the individual strategies over the period.
Next, the authors wanted to determine whether these returns were caused by biases to certain asset classes; so for the combined strategy, they look at the percentage of time that each asset class is allocated to a particular quartile. The US REIT seems to be the most frequent asset class in the top quartile, and UK equity in the bottom quartile; however, no asset class seems to have too big or too small of an allocation in a particular quartile. Therefore, the authors find this be evidence there is no bias toward any particular asset class.
Next, the authors analyze the loadings toward the Fama/French and Carhart 4 factors (i.e., market, size, value, and momentum factors). The returns of the 1-month momentum strategy seem to be explained by the size factor and alpha. The returns of the 12-1 momentum strategy seem to be explained by the market and momentum factors. The returns of the value strategy seem to be explained by market, size, momentum, and alpha; although interestingly, those returns have a negative relationship with the market factor and the momentum factor. Finally, the returns of the combination strategy seem to be explained by alpha, size, and value factors, but the alpha figure is a significant 11%.
Robustness Tests
The authors now analyze the transaction costs of utilizing these strategies. The 1-month strategy has the highest turnover at 1675%, while the valuation strategy has the lowest turnover at 234%; since the 1-month strategy had the lowest returns and the highest transaction costs, its net returns are the lowest. However, the high returns of the combination strategy allowed it to have the highest net returns of all the strategies, when the transaction costs are estimated to be below 0.40%. When the transaction costs exceed 0.40%, the value strategy has the highest net returns, because of its lower turnover. Even at a transaction cost of 0.50%, the combination strategy has excess returns of 4.6%.
Next, the authors replicate the prior results over the 1974 - 1985 period to see if the strategies work over an out-of-sample period. Due to lack of data, there are only 8 asset classes in this analysis. The results are quite similar to those found in the 1986 - 2007 period, with the top quartile returns exceeding those of the bottom quartiles, with high information ratios, and t-statistics across all strategies; these measures were, however, a bit lower than was found in the 1986 - 2007 period. As was found before, the combination strategy significantly outperformed the individual strategies.
Next, the authors understand that some assets are more volatile than others; so the more volatile ones may have more extreme ranks (i.e., end up in the top quartile or bottom quartile more often) than the less volatile asset classes. So, they re-perform the tests with adjusted allocations based on tilting the weights to obtain a 10% volatility for each of the asset classes. The results are similar to the returns we found in the prior sections, with the top quartile outperforming the bottom quartile, the information ratios significant, and the t-statistics significant; as we found before, the combination strategy continues to outperform the individual strategies.
Finally, the authors are concerned that the returns of the strategies are principally because of a single or few asset classes. So the authors provide a chart of the average returns of each asset class when allocated to each quartile. In the top quartile, no asset class averages a return less than 0.3%, and the average return across all asset classes is 0.8%; in the bottom quartile, no asset class averages a return greater than 0.2% (except US mid-cap equities). The authors find this to be evidence that no individual asset class causes the majority of returns in the strategies; the returns seem to be spread across all the asset classes.
Discussion
These results could certainly be due to data mining or chance; however, the authors don't expect this to be the case, and expect these results to continue going forward. The authors also caution, that it could be possible that these returns could just be compensation for a risk that was not modeled; the authors also think that explanation is likely not true, either, because the alpha was 12%, so that presents a very high hurdle to be consumed by risk measures. Also, there didn't seem to be any risk-differences between top and bottom quartile portfolios (e.g., volatility, skewness, etc.). Further, the authors look at how the strategies would have performed in different regimes (e.g., high/low interest rate environment, high/low term spread environment, high/low credit spread environment, and high/low volatility strategy). They find the returns for each strategy to be similar across all 4 strategies in different regimes; however, the value and combination strategies seem to perform differently in different credit spread and interest rate regimes.
These findings provide a challenge to the efficient market hypothesis; however, the authors note that it is challenging to use a risk-model that works across asset classes (because different asset classes have different risks). The authors propose the results of these strategies may be due to behavioral effects that make it difficult for the smart money to arbitrage away these results. For example, practitioners may find cross-asset allocation to be too challenging or the strategy and valuation measures may be too simplistic. Secondly, typical asset managers specialize in individual asset classes; so they are more concerned with selecting individual securities within the asset class, rather than how the asset class as a whole will perform. Finally, allocations by end-investors may be primarily driven by long-term considerations (e.g., pension funds' ALM), fixed allocation mechanisms (e.g., 401ks), herding behavior, or recent performance. The authors believe these constraints will continue going forward, so the results of these strategies will continue as well.
The authors note that hedge funds have the greatest ability to capture this alpha and arbitrage away the results. So to understand whether they currently are doing so, the authors regress the returns of the strategies against the returns of various hedge fund strategies. They find the returns of the 12-1 momentum strategy seem to explain the returns of global macro, long-short equity, managed futures, and multi-strategy. The returns of the combined strategy seem to be related to the managed futures and multi strategy returns. But for the most part, the individual strategies' returns don't seem to be related to hedge fund strategies' returns, so maybe these returns are not being arbitraged away by hedge funds.
Summary, implications and extensions
In summary, the individual strategies (i.e., 1-month momentum, 12-1 momentum, and value) all earned significant return premiums of 7-8% over the 1986 - 2007 period, and the combined strategy (i.e., 25% 1-month momentum, 25% 12-1 momentum, and 50% value) earned an alpha of 12%. Even after adjusting for the Fama/French (i.e., market, size, and value) and Carhart factors (i.e., momentum), there still remains a significant excess return for the strategies. The authors argue against risk-based explanations and instead suggest the market to be macro-inefficient; this is because there is not enough smart money to arbitrage away these alphas due to various constraints. Future researchers could extend this study by expanding the number of asset classes, expanding the number of predictor variables, or introducing portfolio optimization to the allocations.
No comments:
Post a Comment