Conditional safe haven behavior of bitcoin during market stress using explainable artificial intelligence

Abstract This paper employs an Explainable AI framework to analyze how Bitcoin functions as a conditional safe-haven asset during periods of global financial distress. To do this, daily prices from January 2020 to March 2025 for Bitcoin, Gold, and the four Major stock markets (S DAX 40, and CAC 40) are analyzed. Safe-haven divergence is defined as the event in which Bitcoin records a positive daily log return while at least three of four major equity indices, the S&P 500, FTSE 100, DAX 40, and CAC 40, simultaneously experience negative daily log returns. The data were analyzed using supervised classifiers trained on lagged and rolling predictive data, and the models were assessed for predictive performance by splitting the data into time-ordered training and testing sets to avoid look-ahead bias. The calculated model performance metrics include ROC-AUC, PR-AUC, Precision, Recall, and matched imbalance. XGBoost Models outperformed baseline models in detecting divergence events, but their overall predictive performance was moderate. SHAP results indicate that lagged returns and directional changes in U.S. and European Stock Markets are the main Predictors, whereas gold returns and the Stock Market Volatility Index (VIX, MOVE) make minimal contributions. The findings indicate that Bitcoin functions as a conditional safe-haven asset only under certain market-stress conditions, not as a continuous-crisis asset. These findings provide practical insights for investors and portfolio managers on Bitcoin’s conditional role during periods of market stress. The study contributes by introducing a divergence-based, explainable AI framework that enables both prediction and interpretation of conditional safe-haven behavior. Unlike prior studies that rely on correlation-based or quantile regression approaches, this study’s primary innovation lies in combining a novel divergence-based event definition with XGBoost classification and SHAP explainability, enabling simultaneous prediction and economic interpretation of safe-haven episodes across multiple crises from COVID-19 through the Russia–Ukraine conflict and beyond.