Evolutionary Rule Extraction from Corporate Default Prediction Models

Machine learning models significantly outperform traditional methods for predicting SME default, with the study showing improved performance in Balanced Accuracy and PR-AUC metrics across multiple classifiers tested on 50,718 Italian SMEs from 2015-2024, suggesting financial institutions could benefit from adopting these newer approaches.

The "black box" problem in AI creates real regulatory and transparency challenges in finance: The paper demonstrates that while ML models predict better, their limited interpretability raises concerns about compliance and accountability—a critical issue for responsible lending decisions that affect real businesses.

Explainable AI techniques can extract economically meaningful insights from complex models: The DEXiRE-EVO framework successfully identified specific, understandable risk factors (weak liquidity generation, capital erosion, high leverage, and operational inefficiency) from ML models, showing that interpretability and performance don't have to be mutually exclusive.

Context matters for responsible AI in finance: The research found that macroeconomic conditions and persistence of financial instability are important for identifying at-risk SMEs, highlighting that responsible default prediction requires considering broader economic factors beyond individual firm metrics.

Explainability methods enable better stakeholder trust and decision-making: By translating complex model predictions into clear rules about what drives default risk, the approach helps financial institutions, policymakers, and SME owners understand why credit decisions are made, supporting more transparent and defensible lending practices.