Market Regime-Aware Explainability for AI-Driven Financial Decisions: A Counterfactual and Uncertainty-Based Framework

Abstract

AI models now play a key role in financial decisions, including buy, sell, and hold recommendations, risk monitoring, market direction prediction, and portfolio rebalancing. In financial domains, though, high predictive precision isn’t enough, because financial decision-makers want to know why a model made a decision, whether the reasons behind the decisions are sound given the current market conditions, and what type of uncertainty is present in the model's output. Current explainable AI (XAI) methods in finance often provide static feature-attribution explanations, whereas financial markets are dynamic and non-stationary, exhibiting latent regimes such as bullish, bearish, sideways, and high-volatility phases. We present RACU-XAI, a framework for generating market regime-aware frameworks that combines market regime detection, Artificial Intelligence-related financial decision modeling, regime-conditioned explanations, counterfactual explanation generation, and uncertainty quantification. The framework is validated on daily market data for the S&P 500, NASDAQ-100, and Bitcoin/USD, using engineered price, technical, risk, and regime features. We proposed a three-class classification problem over a five-trading-day horizon for the main decision task. The findings indicate that RACU-XAI achieves the best predictive performance on S&P 500 data, with an accuracy of 0.724, an F1 score of 0.712, and an AUC of 0.769. Its performance improves financial performance after transaction costs, and the cumulative return is 61.4%, the Sharpe ratio is 1.31, and the maximum drawdown is -12.6%. Regime-aware SHAP analysis shows that the explanatory drivers vary widely across market conditions, and counterfactual explanations indicate the minimum changes required to shift the decision from Sell to Hold or from Buy to Hold. Finally, uncertainty estimates are strongly associated with error rates, suggesting that uncertainty could serve as a suitable risk-filtering mechanism. The results indicated that conditioning explanations on market regimes in AI-based decision support systems can reinforce transparency, stability, and risk consciousness in financial markets.