Reconciling interpretability and accuracy: Evidence from categorical factors

Linear models have long been favored in asset pricing for their interpretability in low-dimensional settings. In the contemporary high-dimensional landscape, however, non-linear machine learning models often deliver superior predictive accuracy, albeit at the expense of interpretability due to their “black-box” nature. This paper proposes and evaluates a novel type of explanatory variable, categorical factors, to reconcile this fundamental trade-off between interpretability and accuracy. These factors are constructed by first classifying a broad set of firm characteristics into distinct groups based on their financial nature and then extracting a single representative factor from each group using tailored feature selection techniques. Using US stock market data, we demonstrate that incorporating these categorical factors significantly enhances the predictive accuracy of linear models. Particularly, when the factors are extracted using non-linear feature selection methods, the performance of linear models can be elevated to a level comparable with or even better than that of advanced machine learning algorithms. As these categorical factors directly linked to distinct aspects of intuitive economic information, this key finding underscores their potential to integrate interpretability with predictive accuracy, thereby advancing both objectives in asset pricing. To further guide practitioners, we provide a framework for analyzing the relative importance of these factors, offering a practical tool for quantitative portfolio management. We also demonstrate how this framework can be used to track the temporal evolution of the factors’ predictive power, which is critical for adapting investment strategies to evolving market regimes.