| license: mit | |
| tags: | |
| - fintech | |
| - reconciliation | |
| - classification | |
| # FinTech Reconciliation Break Classifier (Pair-Level) | |
| This repository contains a trained model that predicts whether a bank↔broker reconciliation **pair** is a **TRUE_BREAK (1)** or **FALSE_BREAK (0)**. | |
| ## Features Used | |
| ['large_amt_mismatch', 'amt_diff', 'amt_pct', 'amt_diff_abs', 'amt_ratio', 'settlement_gap_days', 'is_weekend', 'is_settlement_weekend'] | |
| ## Best Run (selected by PR-AUC) | |
| - Experiment: exp_001_baseline | |
| - PR-AUC: 1.000 | |
| - ROC-AUC: 1.000 | |
| - F1: 0.998 | |
| ## Files | |
| - `model.joblib` : trained sklearn pipeline | |
| - `features.json`: selected feature list | |
| - `config.json` : training metadata + best metrics | |
| - `predict.py` : minimal inference helper | |