Balancing External vs. Internal Validity: An Application of Causal Forest in Finance.

This article focuses on evaluating the performance and applicability of causal forest, a nonparametric, machine learning–based estimator designed to recover heterogeneous treatment effects (HTEs) with low bias and variance in observational data where treatment is endogenous. Through extensive Monte Carlo simulations, the authors demonstrate that causal forest outperforms ordinary least squares (OLS) and regression discontinuity design (RDD) estimators in many settings, particularly by flexibly controlling for nonlinearities and mitigating bias from latent confounders when these are correlated with observed covariates. The paper applies causal forest to study the effect of loan covenant default on firm investment, revealing substantial heterogeneity: while the average treatment effect (ATE) of default on investment is small and statistically insignificant, a subset of financially weaker, smaller firms with high investment opportunities experience economically meaningful declines. The authors also show that RDD estimates in this context can be biased due to firms' manipulation around the default threshold, and that combining causal forest with RDD—using causal forest's centering step to identify subsamples less prone to manipulation—can improve inference. Overall, the study highlights causal forest as a valuable econometric tool that balances internal and external validity by providing generalizable, heterogeneous treatment effect estimates in settings where traditional pseudorandomization methods like RDD face limitations.