Abstract
In this paper, we document the importance of memory in machine learning (ML)-based models relying on firm characteristics for asset pricing. We find that predictive algorithms perform best when they are trained on long samples, with long-term returns as dependent variables. In addition, we report that persistent features play a prominent role in these models. When applied to portfolio choice, we find that investors are always better off predicting annual returns, even when rebalancing at lower frequencies (monthly or quarterly). Our results remain robust to transaction costs and risk scaling, thus providing useful indications to quantitative asset managers.