Essays on Empirical Asset Pricing

Over the past decade, the integration of machine learning (ML) models into financial re-search has led to significant advances. Despite these advances, ML models suffer from a lack of interpretability and theoretical foundation, limiting financial insight. Structural models, in contrast to ML models, are inherently theory-driven. Despite the broad range of method-ologies that “structural models” encompasses, they share a core characteristic: they are born from theory and are designed to offer explicit predictions and insights into the phenomena they represent. This paper proposes a flexible framework, a Deep Structural Model (DSM), for combining the two methodologies, in an attempt to get the best of both worlds: the high predictability of machine learning and the economic intuition and interpretation of structural models.
The specific structural model examined in this paper is a modified version of the classic Merton (1974) model wherein the assets of the firm follow a geometric brownian motion. The asset drift is the sum of the risk-free rate, a term representing mispricing, and systematic risk compensation, while asset volatility contains a systematic and idiosyncratic component. This model jointly estimates the conditional expected equity returns and (co)variances and enables the analysis of the importance of mispricing relative to systematic risk compensation, as well as the effect of firm leverage on expected equity returns.