Intrinsic (Time Series) Momentum Does Not Really Exist?

Does rigorous re-examination of time series (intrinsic or absolute) asset return momentum confirm its statistical and economic significance? In their April 2018 paper entitled “Time-Series Momentum: Is it There?”, Dashan Huang, Jiangyuan Li, Liyao Wang and Guofu Zhou conduct a three-stage review of evidence for predictability of next-month returns based on past 12-month returns for a broad set of asset futures/forwards:

1. They first run a time series regression of monthly returns versus past 12-month returns for each asset to check predictability for individual assets.
2. They then run pooled time series regressions for asset returns scaled by respective volatilities as done in prior research, overall and by asset class, noting that pooled regressions can inflate conventional t-statistics and thereby incorrectly reject the null hypothesis. To correct for this predictability inflation, they apply three kinds of bootstrapping simulations.
3. Finally, they consider a simple alternative explanation of the profitability of an intrinsic momentum strategy tested in prior research that each month buys (sells) assets with positive (negative) past 12-month returns, with the portfolio weight for each asset 40% divided by its past annualized volatility (asset-level target volatility 40%).

Their asset sample consists of 55 contract series spanning commodity futures (24), equity index futures (9), government bond futures (13) and currency forwards (9). They construct returns for an asset by each day calculating excess return for the nearest or next-nearest contract and compounding to compute monthly excess return. Using daily excess returns for the 55 contract series during January 1985 through December 2015, they find that:

• For individual asset regressions of next-month returns versus past 12-month returns:
  • Using in-sample testing, only 8 of 55 assets (dispersed across asset classes) exhibit predictability at a 10% significance threshold. Only five have R-squared statistics greater than 0.01, and 17 have a negative relationship between future and past returns.
  • Using the first 15 years of the sample period for training and the last 16 years for out-of-sample testing, 45 of 55 assets have a negative relationship between future and past returns. Among the 10 with positive relationships, only three exhibit predictability at a 10% significance threshold. Results are similar for past return lookback intervals of one, three and six months.
• For pooled regressions:
  • Using in-sample testing, conventional t-statistics indicate highly reliable positive relationships between next-month returns and past 12-month returns overall and by asset class. However, bootstrapping simulations show that these reliabilities derive from the pooling methodology, not from inherent return predictability. Moreover, volatility scaling used in the pooled regression contributes substantially to t-statistics. Using raw returns generates much weaker results.
  • Using out-of-sample testing, pooled regressions improve predictability for some commodity and equity futures.
• From an investing perspective:
  • Excluding volatility scaling from the intrinsic momentum strategy outlined above, only eight of 55 assets generate excess profitability at a 5% significance level.
  • On an asset-by-asset basis, an alternative strategy that each month buys (sells) an asset if its inception-to-date historical average return is positive (negative or zero) performs about the same as intrinsic momentum without volatility scaling based on both average return and Sharpe ratio. Average returns differ significantly for only seven of 55 assets.
  • Portfolios of assets formed using intrinsic momentum or inception-to-date average return with volatility scaling as above, past 12-month return weighting or equal weighting perform about the same.

In summary, statistical and economic evidence for intrinsic (absolute or time series) momentum is weak, with a strategy based simply on the sign of inception-to-date average return about as good.

Cautions regarding findings include:

• Return calculations are gross, not net. Accounting for costs of monthly portfolio reformation would reduce returns.
• All assets considered are futures/forwards. Results may differ for other kinds of assets.
• Testing strategy alternatives on the same (or correlated) data introduces snooping bias, such that the best-performing alternative overstates expectations.

See other relevant research summaries for other analyses and perspectives.