Conference Agenda

Machine learning factor models of the cross section of stock returns have produced spectacular results, which are explained by two different ingredients: expanding the information set and allowing for more flexible function forms. We disentangle the value-added of each ingredient, considering a variety of empirical settings: from highly stylised to realistic. We show that the benefit of both ingredients declines when moving from standard settings in the literature to more realistic settings that exclude microcaps, remove look-ahead bias on yet-to-be-published factors, account for transaction costs, and exclude short positions. While the value of nonlinearity disappears even before imposing transaction costs or short-sale constraints, the value of an expanded information set is more persistent. Our feature importance analysis reveals that characteristics like firm size and short-term reversal - crucial predictors in standard settings - lose most of their value once investability constraints are considered. These findings challenge claims about the universal benefits of machine learning sophistication, demonstrating that real-world implementation constraints fundamentally alter which model ingredients improve portfolio performance.

Machine learning (ML) methods are highly flexible, but their ability to approximate the true data-generating process is fundamentally constrained by finite samples. We characterize a universal lower bound, the Limits-to-Learning Gap (LLG), quantifying the unavoidable discrepancy between a model’s empirical fit and the population benchmark. Recovering the true population R2 , therefore, requires correcting observed predictive performance by this bound. Using a broad set of variables, including excess returns, yields, credit spreads, and valuation ratios, we find that the implied LLGs are large. This indicates that standard ML approaches can substantially understate true predictability in financial data. We also derive LLG-based refinements to the classic Hansen and Jagannathan (1991) bounds, analyze implications for parameter learning in general-equilibrium settings, and show that the LLG provides a natural mechanism for generating excess volatility.