Factor Timing with Machine Learning

Can machine learning exploit interactions between many equity factors and many potential factor return predictors to create an attractive factor timing strategy? In their August 2024 paper entitled “Optimal Factor Timing in a High-Dimensional Setting”, Robert Lehnherr, Manan Mehta and Stefan Nagel apply machine learning with mean-variance optimization to time equity factors when the numbers of factors and potential factor return predictors are large. They consider both a small set of four (size, book-to-market, profitability and investment) and a much larger set of 131 factors. They focus on a small set of 11 potential predictors (five economic variables and six specific to the small set of factors) but consider also a much larger set augmenting those 11 with many other factor specific variables. They simplify outputs by suppressing the weakest signals. Their machine learning process each year uses an expanding window of at least 20 years for training, one year for validation and one year for testing. They focus on Sharpe ratio as the essential performance metric. Their benchmarks are annually rebalanced: (1) equal weighting of factors; and, (2) straightforward mean-variance optimization of factors that ignores interactions between factors and potential predictors. To estimate net performance, they apply 0.1% portfolio reformation frictions at the portfolio of factors (not factor) level. Using monthly factor portfolio returns and economic variable values during January 1965 through December 2022, with 1986 the first year of portfolio testing, they find that: (more…)