SHAP-Guided Adversarial Training for Robust Interpretability in Asset Pricing

During the post-COVID volatility regime, many traditional and machine-learning asset pricing models failed to provide stable explanations of return drivers. Shifts in factor attributions undermined confidence in reported risk exposures and led to costly portfolio adjustments, such as the misclassification of quality factors during the 2020 market shock. While predictive accuracy is often the primary benchmark, our objective is to improve the stability of model explanations over time so they remain both economically relevant and reliable for decision-making. To our knowledge, this is the first framework to embed attribution metrics directly into adversarial training for non-differentiable financial models. We develop a SHAP-guided adversarial training approach that constructs perturbations along SHAP-derived directions, simulating adversarial stress without requiring gradient access. This design directly regularizes attribution-relevant input sensitivity during training. Perturbation strength is calibrated through an epsilon-sweep, revealing a non-monotonic trade-off between interpretability and performance: while epsilon = 0.05 maximizes attributional stability, a more moderate epsilon = 0.005 balances robustness with economic signal retention. We evaluate the framework using monthly U.S. stock data from CRSP spanning 1990–2024, partitioned into training (1990–2010), validation (2011–2019), and test (2020–2024) periods. Model performance is assessed along three dimensions: predictive accuracy (MSE, IC), economic relevance (Q10–Q1 portfolio spreads), and attributional robustness (SHAP variance, rank correlation, and directional mean drift). A naive random-sign adversarial baseline is included for comparison. Empirically, SHAP-guided models deliver substantial gains in attributional coherence: SHAP variance falls by over 90%, and feature rank correlation across regimes improves from 0.20 to 0.90. These improvements transform highly unstable explanations into consistent, regime-robust factor profiles. The explanatory focus shifts from fragile, regime-dependent signals like momentum toward persistent, risk-consistent features such as volatility and max down, aligning with investor behavior during crises. These results demonstrate that attributional stability can be explicitly optimized without sacrificing predictive performance, and that it carries clear economic value. In volatile markets, maintaining consistent explanations is as critical as predictive fit for ensuring reliable portfolio allocation, credible economic interpretation, and effective model governance. By design, the SHAP-guided adversarial training framework is model-agnostic and readily transferable to other asset classes and predictive tasks where interpretability is essential, enabling deployment in real-world investment and risk management workflows without imposing explicit economic priors.