Evolutionary Rule Extraction from Corporate Default Prediction Models

Researchers developed DEXiRE-EVO, an evolutionary rule extraction framework combining machine learning with explainable AI to predict SME defaults in Italy. The approach outperforms traditional logistic regression while maintaining interpretability, identifying key risk factors like weak liquidity, high leverage, and operational inefficiency across 50,718 firms from 2015-2024.

This research addresses a critical gap in financial risk modeling: the tension between predictive accuracy and regulatory transparency. Machine learning models have historically excelled at forecasting corporate defaults but operate as black boxes, creating compliance challenges for financial institutions. The DEXiRE-EVO framework resolves this by layering explainability onto ML classifiers through evolutionary algorithms and contextual importance analysis, demonstrating that superior performance and interpretability are not mutually exclusive.

The findings emerge from an exceptionally large Italian SME dataset spanning nine years, providing robust evidence during multiple economic cycles including pandemic disruption. By identifying economically coherent default indicators—liquidity stress, capital depletion, leverage ratios, and operational metrics—the research validates that ML-extracted patterns align with established financial theory rather than artifact noise. This alignment strengthens confidence in both the models and the regulatory acceptance potential.

For financial institutions, this represents a practical toolkit for automating credit decisions while maintaining audit trails and explainability documentation required by regulators like ECB and national banking authorities. SME lending is systemically important for EU economies, where traditional credit scoring often excludes small firms due to sparse financial data. Transparent ML-driven alternatives could expand access to capital while reducing default losses.

The broader implication extends to enterprise AI adoption: regulatory pressure increasingly demands explainability regardless of industry. This research demonstrates that evolutionary computation and contextual analysis can satisfy both accuracy and transparency requirements without sacrificing either. Future implementations may deploy similar frameworks across lending, insurance, and investment management, where regulatory compliance and predictive performance directly impact profitability.

• →DEXiRE-EVO combines machine learning with evolutionary rule extraction to achieve superior default prediction while maintaining interpretability.
• →ML classifiers significantly outperformed traditional logistic regression on balanced accuracy and PR-AUC metrics across 50,718 Italian SMEs.
• →Extracted rules identified economically meaningful risk factors including weak liquidity generation, capital erosion, high leverage, and operational inefficiency.
• →The framework bridges the regulatory compliance gap by providing explainable decision paths for automated credit risk assessment.
• →Contextual macroeconomic conditions and persistence of financial instability emerged as critical secondary indicators in default identification.