We examine which company characteristics determine the economic value of machine learning portfolios. Three results stand out. First, the importance of variables throughout the sample is simply too narrow and provides little reliable guidance, highlighting the necessity for evaluation outside the sample based on economic criteria. Second, traditional models are dominated by microcaps, which drive up returns and concentrate profits in high-cost stocks; Excluding microcaps is important for meaningful conclusions. Third, some predictors have negative meaning and consistently degrade performance; Removing them will improve risk-adjusted returns and make clear which features are necessary. These results show that machine learning can only provide reliable insights into asset pricing under economic constraints.
