From Prediction to Scientific Discovery: A Case Study in Causal Inference for Financial Economics

This paper presents a case study on bridging the translational gap between advanced causal machine learning and scientific practice in financial economics, directly addressing the core questions of the CauScien workshop. We tackle a fundamental scientific question: what are the causal drivers of stock market troughs? Moving beyond the "black box" prediction paradigm, we implement a novel, two-stage comparative causal analysis designed for a complex, real-world setting. We first establish a baseline using Double/Debiased Machine Learning (DML) for a standard partially linear model. Recognizing the limitations of this assumption in a non-linear domain, we then employ a more flexible DML specification to estimate the Average Partial Effect (APE), which is better suited to our binary, interactive setting. The comparison reveals that conclusions about economic causality are critically sensitive to model specification. The more flexible APE model corrects the economic interpretation of key indicators and uncovers robust causal roles for the volatility in options-implied risk appetite and market liquidity—relationships obscured or misrepresented by the simpler linear model. By integrating these findings with intermediary asset pricing theories, we demonstrate how translating modern causal inference methods to a complex social science domain can yield new scientific insights.