Machine Learning and Predicted Returns for Event Studies in Securities Litigation

We investigate the use of machine learning (ML) and other robustestimation techniques in event studies conducted on single securities for the purpose of securities litigation. Single-firm event studies are widely used in civil litigation, with billions of dollars in settlements hinging on the outcome of the exercise. We find that regularization (equivalently, penalized estimation) can yield noticeable improvements in both the variance of event-date abnormal returns and significance-test power. Thus we believe that there is a role for ML methods in event studies used in securities litigation. At the same time, we find that ML-induced performance improvements are smaller than those based on other good practices. Most important are (i) the use of a peer index based on returns for firms in similar industries (how this is computed appears to be less important than that some version be included), and (ii) for significance testing, using the SQ test proposed in Gelbach et al. (2013), because it is robust to the considerable non-normality present in abnormal returns.