Predicting Profitability Deterioration in U.S. Commercial Banks Using Explainable Machine Learning
TL;DR
An explainable machine-learning framework for predicting next-quarter profitability deterioration in U.S. commercial banks can support early-warning prioritization but should complement, not replace, financial analysis and supervisory judgment.
Abstract
This study develops an explainable machine-learning framework for predicting next-quarter profitability deterioration in U.S. commercial banks. Public regulatory records yield 277,959 bank-quarter observations from 7,344 banks. Deterioration is defined as a decline of at least one percentage point in annualized quarterly return on assets. L2 logistic regression, Random Forest, and XGBoost are evaluated through an expanding window covering 15 out-of-sample quarters from 2022Q1 to 2025Q3. XGBoost achieves mean ROC AUC of 0.8196, average precision (AP) of 0.3281, Brier score of 0.0303, and top-decile event capture of 54.09 percent. A restricted five-knot cubic-spline logistic model using only current ROA attains mean ROC AUC of 0.7593 and AP of 0.2247; full-predictor XGBoost improves these measures by 0.0603 and 0.1035 and raises top-decile capture by 7.71 percentage points (all paired p < 0.001). When the original out-of-sample scores are evaluated against persistent deterioration over two quarters, XGBoost attains mean ROC AUC of 0.9202 and AP of 0.3989. SHAP analysis consistently ranks current profitability, nonperforming loans, and bank size as the three leading predictors, followed by the compensation expense ratio and the net interest income-to-assets ratio. The framework can support early-warning prioritization but should complement, not replace, financial analysis and supervisory judgment.