Causal Machine Learning for Identifying Market Stress Drivers: A Leakage-Safe Walk-Forward Investigation

Abstract

The early identification of market stress drivers is essential for maintaining financial stability, yet conventional machine learning approaches suffer from persistent causal confusion and evaluation leakage that undermine their practical credibility. This paper develops a leakage-safe walk-forward framework that integrates causal machine learning with rigorous temporal validation to isolate structural drivers of market drawdowns. We argue that the standard cross-validation protocols commonly adopted in financial machine learning violate the temporal ordering of information, leading to inflated performance metrics and spurious causal attributions. By contrast, our proposed architecture enforces strict temporal independence across training, validation, and test partitions while embedding causal inference methods—including double machine learning and heterogeneous treatment effect estimation—within a walk-forward loop. The framework is designed to separate persistent market stress signals from transient volatility, thereby enabling more robust early-warning systems. We examine the system-level trade-offs between predictive accuracy, interpretability, and computational sustainability, and discuss the governance implications of deploying causal models in high-stakes financial regulation. Through a series of conceptual case illustrations drawn from recent literature on leakage-safe evaluation and causal discovery, we demonstrate how the proposed methodology mitigates overfitting to spurious correlations and enhances the economic credibility of stress detection. The paper concludes with forward-looking recommendations for the design of causal, leakage-resilient infrastructures in financial surveillance.