Explainable Predictive Business Intelligence for Banking: Integrating Machine Learning, SHAP, and Large Language Models for Customer Response Prediction

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This study develops an integrated predictive business intelligence (BI) framework that combines machine learning (ML), explainable artificial intelligence (XAI), and large language models (LLMs) for customer response prediction in banking. Using the UCI Bank Marketing dataset of 45,211 observations, we compare Logistic Regression, Decision Tree, Random Forest, and XGBoost using accuracy, precision, recall, F1-score, and ROC-AUC. Among the evaluated models, Random Forest achieved the strongest overall performance, with 90.0% accuracy, 56.0% precision, 53.0% recall, 55.0% F1-score, and 91.8% ROC-AUC. XGBoost produced the highest recall at 78.0%, with 86.0% accuracy, 45.0% precision, 57.0% F1-score, and 91.4% ROC-AUC. Logistic Regression achieved 82.0% accuracy and 90.3% ROC-AUC, whereas Decision Tree achieved 87.0% accuracy and 67.3% ROC-AUC. We apply SHapley Additive exPlanations (SHAP) to identify the factors contributing to model predictions and integrate an LLM to convert structured prediction and explanation outputs into natural-language business insights. The resulting architecture connects predictive modeling, model explainability, BI visualization, and managerial interpretation within a unified decision-support workflow. The findings indicate that Random Forest provides a balanced predictive performance for the benchmark task, while XGBoost offers greater sensitivity to potential positive cases. The framework demonstrates how combining ML with XAI and LLM-based interpretation can improve the accessibility and transparency of predictive BI. However, practical deployment requires institution-specific validation, particularly because the benchmark data originate from a Portuguese banking campaign and include variables that may introduce temporal leakage. The study therefore positions the proposed framework as a research and deployment architecture rather than evidence of direct performance in U.S. banking environments.