Pooling and winsorizing machine learning forecasts to predict stock returns with high-dimensional data

We evaluate US market return predictability using a novel data set of several hundred ag- gregated firm-level characteristics. We apply LASSO, Elastic Net, Random Forest, Neural Net, Extreme Gradient Boosting, and Light Gradient Boosting Machine methods and find these models experience large prediction errors that lead to forecast failures. However, winsorizing and pooling machine learning model forecasts provides consistent out-of-sample predictability. To assess robustness, we apply machine learning methods to high-dimensional data for Canada, China, Germany and the UK as well as the Goyal–Welch data. All machine learning models we consider, except for the ensemble pooled methods, fail to significantly predict returns across our samples, highlighting the importance of pooling, evaluating additional economies, and the fragility of individual machine learning methods. Our results shed light on the sparsity versus density debate as the degree of sparsity and variable importance evolves over time.