There is a vast body of academic research on momentum as an explanatory factor in the cross-section of returns, beginning with the 1993 seminal paper of Narasimhan Jegadeesh and Sheridan Titman, Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency, published in the Journal of Finance. The research provides compelling evidence that buying instruments that have performed well (both relatively — cross-sectional momentum — and absolutely — time-series momentum) in the recent past and selling those that have performed poorly has produced abnormal returns in the short run.
Fotis Papailias, Jiadong Liu and Dimitrios Thomakos contribute to the momentum literature with their study Return Signal Momentum, published in the March 2021 issue of the Journal of Banking & Finance. They introduced financial market momentum based on the frequency of the signs of past returns, which they called “return signal momentum” (RSM). While a form of time-series momentum (TSM), RSM differs in two ways: (1) It takes into account each of the returns during the look-back period (rather than calculating the total period return as in TSM), and (2) it focuses on the signs of past returns regardless of their magnitude.
The researchers examined the performance of RSM strategies in 55 of the world’s most liquid commodity and financial futures markets (24 commodity futures, nine foreign exchange futures for nine countries against the U.S. dollar, nine equity indexes of developed countries, and 13 government bonds for six developed countries with various maturities). The earliest start date was 1970 and the latest was 1990. The sample ended in 2015. The RSM position signals were generated when the equally weighted average of past return signs exceeded a certain probability threshold. The authors noted that an intuitive strategy would be to use a threshold of 0.5—a long signal is generated when no less than 50 percent of the returns over the past 12 months are positive and goes short if less than 0.5.
Using the model validation technique known as “cross-validation” (sometimes referred to as rotation estimation or out-of-sample testing), Papailias, Liu and Thomakos also developed a time-varying threshold. And they analyzed fixed thresholds for the mean of 0.2, 0.3, 0.4, 0.5, 0.6, 0.7 and 0.8. The lookback period used to calculate RSM was 12 months, with signals reviewed monthly. Following is a summary of their findings:
71 percent of all assets examined had higher autocorrelation estimates for their RSM signals than for their TSM signals.
The persistence in the RSM approach is higher than that of the TSM approach and thus more predictable.
RSM is highly related to the global stock market index despite the fact that the 55 assets come from different asset classes. It is also related to TSM.
RSM generates superior profitability and lower risk characteristics relative to benchmarks such as a buy-and-hold strategy, the simple price moving average strategy and the TSM strategy.
RSM strategies perform better than the benchmarks when the threshold value for the mean is no larger than 0.5.
The Sharpe ratio of the best RSM strategy (a threshold of 0.4), at .0962, was 21 percent better than that of TSM’s, at 0.792.
All the RSM strategies with a fixed threshold less than 0.5 were associated with an annual return that was at least equal to 10 percent (minimum average monthly return of 0.103), with lower or similar volatility to the TSM. RSM portfolios also resulted in larger cumulative net profits and smaller maximum drawdown, indicating desirable risk/return characteristics. Specifically, the cumulative net profits of the RSM 0.4 portfolio were almost 18 percent larger than those of TSM, and the maximum drawdown (19.5 percent) was almost 44 percent smaller. The results were consistent when including estimates of transactions costs.
Controlling for TSM and CSM, the RSM effect does not come from the cross-sectional part of the asset pool, but mainly from the time-series part.
Using the probability of positive signs of the returns, which is an important indicator of RSM, shows more robust short-run continuation and long-run reversal.
Time-varying threshold correctly captures the market conditions, indicating that it has better market timing. A time-varying probability threshold that is based on cross-validation suggests that the threshold is negatively correlated with the market. In particular, when market expectations are positive, the time-varying threshold decreases, allowing the investor to take more long positions. When the market conditions deteriorate, the time-varying threshold increases, protecting the investor from the coming downtrend. However, the fixed threshold of 0.4 produced a higher Sharpe ratio (0.962 versus 0.916) and higher average monthly returns (0.119 versus 0.110), with virtually the same volatility, and much higher cumulative returns (27.164 versus 21.234) and much lower maximum drawdown (19.5 percent versus 26.8 percent).
The stronger the signal (the greater the RSM mean), the higher the performance/sign dependence. However, the higher the volatility, the lower the performance/sign dependence.
RSM and TSM strategies performed better when scaling for volatility than when unscaled. The reason is that sign predictability is negatively related to volatility. Thus, requiring an individual instrument’s weight to be inversely proportional to its volatility can further improve portfolio performance.
Their findings, which were statistically significant at the 5 percent confidence level, led Papailias, Liu and Thomakos to conclude: “Overall, our research indicates that market participants can successfully apply RSM as an alternative type of momentum for both speculation and risk management purposes.”
As you consider their results, the following words of caution are offered:
First, the authors did not publish any t-stats. Thus, we don’t know how significant the results were.
Second, when reviewing academic research findings, even from papers with long and large data samples covering multiple asset classes (such as the 55 futures contracts and four asset classes and as much as 45 years of data) that minimise the risks of data mining and have impressive t-stats, it’s important that a strategy also have intuitive reasons for believing the strategy will persist in the future. In this case, I’m referring to the finding that a 0.4 threshold was optimal. While a threshold of 0.5 is intuitive, there is no intuitive reason to believe that 0.4, or any other figure, would be more optimal. With that said, compared to the 0.4 threshold, the Sharpe ratio of the 0.5 threshold strategy was a strong 0.88 (versus 0.962), its monthly average return was 0.103 (versus 0.119), and its maximum drawdown was virtually identical (0.190 versus 0.195).
Third, with strategies that have high turnover, there is the potential for greater implementation drag. The authors assumed that the portfolio was rebalanced after the market close at the end of the month and before the market opened the next month. That, of course, isn’t possible. And while they did estimate transactions costs, they are hard to estimate.