Application of Artificial Intelligence in Predicting Banking Crises: Analysis of Audited Data in Accordance with National Standards in Iran’s Banking System

This study aims to develop and validate an artificial intelligence-based early warning system for predicting banking crises in Iran by integrating audited financial data, audit quality indicators, and macroeconomic variables, thereby improving the accuracy and interpretability of crisis forecasting for regulatory and supervisory applications. A comprehensive dataset of ۷۹۲ bank-year observations from ۲۲ Iranian banks over the period ۲۰۰۴–۲۰۲۳ was analysed. The study employed multiple predictive models, including XGBoost, Random Forest, LSTM, Transformer, and classical logistic regression. Key variables included the capital adequacy ratio (CAR), non-performing loans (NPL), return on assets (ROA), liquidity ratios, audit opinion types, audit tenure, disclosure quality, GDP growth, inflation, and exchange rate volatility. Model performance was evaluated using ۵-fold cross-validation and accuracy, sensitivity, specificity, F۱-score, AUC-ROC, and AUC-PR metrics. SHAP (SHapley Additive exPlanations) analysis was used to ensure model interpretability and to identify key risk drivers. XGBoost achieved the highest predictive accuracy (۹۶.۱۰%), AUC-ROC (۰.۹۸۲), and F۱-score (۰.۹۶), outperforming logistic regression by reducing Type I errors by ۵۹.۸% and Type II errors by ۶۱.۱%. SHAP analysis showed that CAR (۲۰.۸%), NPL (۱۷.۶%), and ROA (۱۵.۶%) were the most influential predictors, while audit quality variables contributed ۱۳.۵% to overall model performance. The NLP module achieved ۹۳.۲% accuracy in classifying audit opinions and an AUC-ROC of ۰.۹۴۷ using textual features alone. Artificial intelligence, particularly XGBoost combined with SHAP interpretability and NLP-enhanced audit analysis, provides a robust and actionable framework for predicting banking crises in Iran, offering significant improvements over classical statistical methods for proactive regulatory intervention.

کلیدواژه ها:

Audited Financial Data ، Audit Quality ، Artificial intelligence ، Banking Crisis Prediction ، Xgboost ، SHAP Interpretability