tags:
- xgboost
- tabular-classification
- finance
- credit-scoring
Ikimina Digital Trust Scorer
Model Details
- Model Type: XGBoost Classifier (CPU-only)
- Objective: Predict the probability of default (within 6 months) for members of Rwandan informal savings groups (Ikimina) and calibrate it into a 0-100 Reliability Index.
- Architecture:
n_estimators=50, max_depth=3, learning_rate=0.1
Intended Use
This model is designed to be queried via a USSD interface by Ikimina secretaries. It outputs a Reliability Index to help groups make informed decisions before approving pool loans, mitigating the risk of capital loss.
Training Data & Features
The model was trained on synthetic tabular transaction records simulating 12 months of Ikimina financial behavior. It utilizes 9 engineered features, heavily weighting behavioral metrics:
feat_total_missed (Most Important: 43%)feat_repayment_ratio (22%)feat_tenure_monthsfeat_recent_miss_score (Recency-weighted)feat_max_on_time_streak
Evaluation Metrics (100-Member Holdout)
- ROC-AUC Score: 0.982
- Brier Score: 0.031
- Calibration: Highly calibrated across three business tiers: High Risk (0-40), Watch (41-70), and Low Risk (71-100).
