Multi-Factor Timing – A time series clustering approach
IIIABSTRACTUnsupervised learning tools in conjunction with macroeconomic forecasts can improve factor rotation strategies relativeto traditional logit-based models and static equal-weight portfolios. Applying a popular subsequence time series clustering algorithm to identify common themes, or ‘motifs’, within six European factor return time series, I link these motifs to a variety of lagged macro-economic state variables usingordered logit regressions. The resulting, out-of-sample trading strategy generates strong, beta-and leverage-adjusted excess returns and substantially improved Sharpe, Omega and MAR ratios between 2011 and 2019, even after implementation costs. It also shows significant and exploitable alpha in Fama & French’s six-factor model. A plain macro model simply calibrated on the factors’ past standardized return time series, however, underperforms a static equal-weight portfolio and doesn’t generate significant alphas. Overall, the findings hint at the importance ofselectivity when calibrating logit or probit regression-based prediction models.