Hybrid Forecasting Models for Credit Default and Financial Risk: A Comparative Empirical Study

Abstract

This study presents a comparative evaluation of six major machine learning and statistical modeling techniques—K-Nearest Neighbors (KNN), Logistic Regression (LR), Discriminant Analysis (DA), Naïve Bayes (NB), Artificial Neural Networks (ANN), and Classification Trees (CT)—for the prediction of credit default probability among credit card clients. Leveraging real-world Taiwanese credit data alongside custom pre-processed market datasets and Python-based simulations, the study applies a novel evaluation framework using lift charts and regression diagnostics to evaluate both classification performance and probability estimation fidelity. Experimental validation is carried out through Python models and augmented by a custom smoothing technique to approximate latent default probabilities. Empirical evidence indicates the superior generalizability and probabilistic alignment of neural networks, highlighting their promise in robust credit scoring systems. The results are substantiated through extensive figures, tables, and cross-model evaluations, providing a reproducible blueprint for risk forecasting across financial sectors.