Credit risk prediction using behavioral features in an AI chatbot interface

Abstract

The following article presents an innovative credit risk assessment system that combines artificial intelligence with real-time interaction through a Streamlit-deployed chatbot. The aims of this article are demonstrated using a dataset of 3,080 clients from Tunisian banks in 2024, including financial and behavioral variables. Random Forest and XGBoost models were trained to predict loan defaults with up to 90% accuracy. The focus was on detecting high-risk profiles, achieving perfect recall (100%) and 76% precision for this class. The SHAP method ensures decision transparency by identifying key predictive variables. This system enables bank advisors to instantly obtain a risk score with personalized explanations, enhancing the speed, efficiency, and trustworthiness of credit evaluations, while outperforming traditional approaches.