DEXiRE-EVO: Evolutionary AI pulls interpretable rules from SME default models

A team of researchers has developed DEXiRE-EVO, an evolutionary rule extraction framework that addresses the interpretability gap in machine learning-based credit risk models. By combining multi-objective optimization with Contextual Importance and Utility (CIU) explainability, the framework extracts transparent, human-readable rules from complex models. The study analyzed a panel of 50,718 Italian SMEs over the period 2015-2024, comparing traditional logistic regression with several ML classifiers including gradient boosting and neural networks. The ML models achieved significantly higher Balanced Accuracy and PR-AUC than logistic regression, confirming that data-driven approaches can improve default prediction—but at the cost of interpretability.

DEXiRE-EVO bridges this trade-off by evolving a compact set of rules that reveal economically meaningful patterns. The extracted rules highlight that weak internal liquidity generation, internal capital erosion, high leverage, and operational inefficiency are primary drivers of SME financial distress. Additionally, contextual macroeconomic conditions and persistence of financial instability help identify high-risk firms. The work demonstrates that combining ML with evolutionary rule extraction yields both strong predictive performance and transparency, supporting more trustworthy, data-driven decision-making in lending and regulatory compliance.