An Empirical Evaluation of Missing Value Imputation Strategies for Credit Default Prediction: Robustness across Missingness Rates and Mechanisms on U.S. Consumer Lending Data

Abstract

Missing values are pervasive in consumer credit data, arising from incomplete application forms, partial credit-bureau records, and heterogeneous coverage across upstream data feeds. The choice of imputation strategy can materially alter downstream default-prediction performance, yet systematic empirical evidence on how that choice interacts with missingness rate and missingness mechanism remains limited. This study reports an empirical evaluation of five widely used imputation strategies --- mean and mode filling, K-nearest-neighbor imputation, multivariate imputation by chained equations, iterative random-forest imputation, and the native missing-value handling of gradient-boosted trees --- on three public consumer-credit datasets that span U.S. originations and a major non-U.S. retail-lending portfolio: Lending Club, Home Credit Default Risk, and Give Me Some Credit. Injected missingness ranges from 10% to 50% under both missing completely at random and missing not at random mechanisms, with AUC as the primary ranking metric and the KS statistic and Brier score as secondary diagnostic checks for separation and calibration.To further isolate the effect of train-time imputation from train-test missingness mismatch, a matched-condition experiment is also conducted in which the test split carries the same injected missingness rate as the training split. Native tree-based handling and iterative random-forest imputation retain the most stable performance at elevated missingness rates, while simple filling degrades more visibly under non-random patterns, especially for applicants with limited observed credit history. The findings offer practical guidance on preprocessing decisions for credit data products.