Website CV Paper
Fields
Econometrics, Finance
Job market paper
Boosting Portfolio Performance with Missingness-Aware Deep Learning
Firm characteristics are key signals for return prediction in machine-learning-based asset pricing, but widespread missingness can hinder their use. While recent studies developed more sophisticated imputation methods, evidence about their value for asset pricing remains mixed. We investigate whether information associated with missing characteristics can be better exploited when the prediction model distinguishes it from information observed directly. We propose a new class of deep learning asset return prediction models – which we call missingness-aware – and show that explicitly incorporating information associated with missing characteristics – which characteristics are missing and how many – improves out-of-sample return prediction and long-short portfolio performance.