Predicting Consumer Credit Reporting Dispute Outcomes Using Distilbert and XGboost Ensemble Learning

Among the most common complaints received by the Consumer Financial Protec-tion Bureau (CFPB) are those related to credit reports, with over 500 000 complaints filed in 2024, exceeding the capacity for manual reviews. This paper developed a hybrid approach using XGBoost and DistilBERT to predict at first contact whether a credit-reporting complaint will receive corrective action or an explanation of why corrective action would not be taken. DistilBERT’s 768-dimensional narrative em-beddings are fused with categorical metadata (company, product, issue, sub-issue, state), and the model was optimized through randomized search with 5-fold cross-validation. Instead of SMOTE, class imbalance was handled through cost-sensitive weighting. A feature-ablation study showed that nearly all of the improvement over text-only models came from company identity; added on its own, it improved F1 by .142. SHAP values indicated that company identity was the greatest contributing factor, and its direction tracked each bureau’s observed relief rate (Experian 11.46%, Equifax 84.23%, TransUnion 96.65%). We refer to this as the textual ceiling: narrative content alone explains only part of the variation in outcomes. A disaggregated fair-ness assessment found significantly lower relief recall rates for Older Americans and Servicemembers, and these differences remained after controlling for company iden-tity, although protected-status variables had been excluded during training. Across N = 1417351 records, the enhanced model achieved a macro-F1 score of .7983 and a ROC-AUC of .8791 on the full-scale test data (n = 212543), performing significantly better than text-only baseline models (macro-F1 ≈ .639; McNemar’s χ2 = 2390.52, p < 10−300).