Cost Sensitive Machine Learning for Credit Risk Assessment and Lending Decisions

Abstract

Machine learning can improve credit risk prediction, yet predictive gains do not automatically translate into better lending decisions. The financial value of a score depends on which errors it makes, whether its probability is calibrated, and how its operating threshold reflects the costs of approval and rejection. This study develops an analytical framework linking these quantities for statistical and ensemble credit scoring models. Published findings on classifier benchmarking, imbalanced credit data, behavioral information, cost-sensitive learning, and distributional effects establish the research context. The original analysis derives feasible loss bounds from accuracy and class prevalence, identifies conditions for economic dominance between classifiers, and examines sensitivity to error costs and probability uncertainty. Deterministic scenarios use adverse class prevalences from 5% to 50%, relative error penalties from one to ten, and classifier accuracies from 70% to 95%. No borrower-level models are newly trained. At 85% accuracy, 30% adverse prevalence, and a five-to-one error penalty, feasible loss ranges from 150 to 750 units per 1,000 applications. Two feasible classifiers with accuracies of 85% and 80% can incur losses of 750 and 200 units, respectively, demonstrating a possible reversal between statistical and economic rankings. Expected loss minimization gives a probability threshold of one sixth at a five-to-one penalty ratio. A general regret expression also shows why probability uncertainty near that threshold deserves particular attention. The framework explains how findings from credit scoring research can inform model selection without treating results from different populations as directly comparable. It supports evaluating the combined prediction and decision process through discrimination, calibration, financial consequences, and distributional outcomes.