A SHAP-Based Comparative Analysis of Machine Learning Model Interpretability in Financial Classification Tasks
A SHAP-Based Comparative Analysis of Machine Learning Model Interpretability in Financial Classification Tasks

Summary/Abstract: As artificial intelligence technologies become increasingly prevalent across the financial sector, the interpretability of machine learning models has become a critical concern for regulatory authorities and financial institutions. This study employs SHAP (SHapley Additive exPlanations) to systematically compare the predictive performance and interpretability of five mainstream machine learning models in financial classification tasks. Using a real financial dataset containing 24 financial indicators to train logistic regression, five machine learning models - logistic regression, random forest, XGBoost, LightGBM, and support vector machine - are trained on this dataset. SHAP is then applied to analyse the feature importance patterns across models. Empirical results demonstrate that LightGBM achieves the best predictive performance (accuracy 95.90%, Area Under the Curve (AUC) 99.18%), while XGBoost shows advantages in terms of interpretability. SHAP analysis identifies those prior earnings per share is the most critical feature, and the Top-K overlap analysis reveals a high degree of consistency among tree-based models in feature importance recognition. This study provides scientific basis for financial institutions to select appropriate explainable AI models, and holds significant importance for enhancing transparency and trustworthiness in financial AI applications.

• Details
• Contents