Machine Learning, Regularization, and Reporting Bias

We study how investors' use of machine learning methods affects corporate financial reporting decisions. Our model consists of a representative investor who employs a regularized regression (e.g., lasso) to infer value from earnings reports, along with a firm manager who may bias earnings reports. The investor's use of regularized regression serves as a commitment to underutilize earnings reports, and consequently, induces a lower price-earnings association, a smaller reporting bias, but greater price efficiency. The two-account version of our model suggests that regularization gives rise to a spillover effect: factors that influence the reporting bias in one account indirectly impact the manager's incentive to manipulate the other account. The model offers policy implications on algorithmic transparency and provides testable predictions