other

Fake Promise-to-Pay Detection

Payment-derived labels and supervised classification for detecting unreliable promise-to-pay commitments.

XGBoost
Classification
Leakage-Aware Validation

Unified 10 synthetic collections datasets into 3,696 promise-to-pay records. Derived labels from observed payments and validated them against human annotations (Cramér’s V = 0.56, p < 0.0001). Compared six classifiers with leakage-aware preprocessing and validation-based threshold selection. The final XGBoost model reached 0.83 PR-AUC and 0.85 F1 on held-out data, against an approximately 0.71 prevalence baseline. These are résumé-reported results on synthetic collections data, not production debt-collection outcomes. Context: Open IIT Data Analytics.