Trends’ Signal Strength and the Performance of CTAs

Abstract

We propose a new asset-based factor that relies on aggregating momentum signals over different horizons. Aggregating signals this way captures assets’ trend signal strength, thereby addressing a limitation in existing time series momentum strategies. Our factor mimics a trend-following manager that increases exposure to markets where trends develop and decreases exposure to markets where trends fade. Taking into account a number of practical implementation issues, we found that our proposed factor performs better at replicating the stylized facts of Commodity Trading Advisors’ returns than previous methods and allows a more meaningful assessment of fund alpha.

Editor's Note: The original version had a typographical error in Equation 5, which has been corrected in this version.

Disclosure: The authors report no conflicts of interest.

Editor’s Note

Submitted 9 November 2017

Accepted 31 August 2018 by Stephen J. Brown

View correction statement:

Correction

Acknowledgments

We thank the people at RPM Risk & Portfolio Management AB—in particular, Alexander Mende—for many stimulating discussions and comments. We also thank Dries Heyman, Koen Inghelbrecht, and Kathryn Kaminski for helpful comments. Last but not least, we thank two anonymous reviewers and the Financial Analysts Journal executive editor for stimulating comments and suggestions. This work was supported by funding from EC Grant Agreement n. 324440 (Futures) Marie Curie Action Industry-Academia Partnership and Pathways Seventh Framework Programme.

Notes

¹ Or, recently, in relation to the factor model specification of Buraschi, Kosowski, and Trojani (2014), who proposed a factor that tries to capture correlation risk.

² For a thorough discussion, see Baltas and Kosowski (2013) and the references in it.

³ Although reporting to hedge fund databases is voluntary, Joenväärä, Kosowski, and Tolonen (2012), in an analysis of the various publicly available hedge fund databases, concluded that BarclayHedge is the most comprehensive, especially for CTAs. Data from January 1994 were used to mitigate a potential survivorship bias; most databases collect information on defunct programs only from 1994 on (see Joenväärä et al.).

⁴ By keeping track of the number of months that were backfilled when a fund was first included in the BarclayHedge database, RPM Risk & Portfolio Management tracked backfill bias in the database for the 2005–10 period. For that sample period, the median (average) backfill bias was 12 (14) months, which is consistent with the correction proposed by Kosowski et al. (2007).

⁵ Results for a trading strategy that also included horizons between 1 and 9 days were qualitatively unchanged from results reported here and are available upon request.

⁶ Recently related to our work, Lempérière, Deremble, Seager, Potters, and Bouchaud (2014) and Dudler, Gmür, and Malamud (2015) suggested TSMOM specifications that account for the signal strength.

⁷ RPM Risk & Portfolio Management is a fund of funds specializing in CTA strategies and liquid global macro managers. It is based in Stockholm.

⁸ Effective spread estimators (Roll 1984; Smith and Whaley 1994) and approaches to estimate the bid–ask spread directly from the order book (Locke and Venkatesh 1997) have also been proposed. Szakmary, Shen, and Sharma (2010) and Locke and Venkatesh pointed out, however, that these estimates are close to the tick size. Because estimating the bid–ask spread from the order book was beyond the scope of the current study, we stuck to the simplification that the tick size is a good proxy for the bid–ask spread.

⁹ See Frazzini, Israel, and Moskowitz (2012) for details on estimating nonlinear transaction cost functions and their implications for exploiting asset-pricing anomalies.

¹⁰ When a futures contract is rolled over to a further-dated contract, the strategy closes the nearby contract and opens a position in the new contract. The date of the contract rollover coincides with the rollover used for the construction of the continuous futures (see the online supplemental material, available at www.cfapubs.org/doi/suppl/10.1080/0015198X.2018.1547052, for the continuous contract construction as well as the estimated transaction costs). On such days, turnover is usually much higher than on other days. Daily turnover of the strategy is fairly limited except in the case of short-rate futures. These contracts exhibit low levels of volatility (0.01% average daily volatility) in comparison with other contracts (1.2% average daily volatility) and thus require a large notional position to obtain similar levels of risk. Each short-rate futures contract included generated an average daily turnover of 22%–23%, whereas the average turnover for the other contracts is just 0.3%.

¹¹ The time series momentum factor (of MOP) is available from AQR’s website: www.aqr.com/Insights/Datasets/Time-Series-Momentum-Factors-Monthly.The BK monthly, weekly, and daily futures-based trend-following benchmarks are available from Robert Kosowski’s website: www.imperial.ac.uk/people/r.kosowski/research.html.

¹² The BarclayHedge and SG manager-based indexes are equal weighted. As a consequence, these indexes overweight CTAs that target high levels of volatility. The manager-based indexes are rebalanced once a year. The BarclayHedge CTA Index is a broad index of CTAs, some of which are not trend followers or systematic. The SG CTA Index includes only the 20 largest CTAs that are open to investment and report performance and AUM on a daily basis.

¹³ Following the work of MOP; Asness, Moskowitz, and Pedersen (2013); and Koijen et al. (2018), we regressed ATSMOM’s returns on a number of macroeconomic, liquidity, volatility, and sentiment variables. The results, reported in the online supplemental material (available at www.cfapubs.org/doi/suppl/10.1080/0015198X.2018.1547052), indicate that variation in these variables does not explain ATSMOM’s excess returns.

¹⁴ We refrained from using an incremental F-test because of potential multicollinearity issues. The pairwise Pearson correlation between our factor and the average of BK’s factors is 0.8.

¹⁵ A Marchenko–Pastur distribution is the asymptotic behavior of singular values of large rectangular random matrixes. See also, for example, Süss (2012).

¹⁶ Results are available in the online supplemental material (available at www.cfapubs.org/doi/suppl/10.1080/0015198X.2018.1547052).

¹⁷ Principal components are indifferent to scaling because they are extracted in such a way as to show zero pairwise correlation.

¹⁸ We thank an anonymous referee for this valuable insight.

¹⁹ Note that dropping funds that stopped reporting before turning two years old unavoidably introduced some survivorship bias.

²⁰ We thank an anonymous referee for this valuable insight.