ikimina-reliability-index

Model card - Ikimina Reliability Index

Artifacts: Use the Files and Versions tab on this repository.

What this is: CPU XGBoost (tree_method="hist") + CalibratedClassifierCV (sigmoid, cv=4), saved as calibrated_model.joblib (scikit-learn / joblib — not Transformers). Load in Python with joblib.load. Use model_meta.json with it: 14 feature names in order (feature_columns), score formula, holdout metrics, tier labels.

Output: Reliability index in [0, 100] from default probability on label defaulted_within_6m (synthetic, 6-month horizon). Tiers: 0–40 high risk, 41–70 watch, 71–100 low risk.

Data & split: Synthetic data (generate_data.py, seed 42), 500 members × 12 months. Holdout: last 100 ascending member_id. Inputs match ikimina_members.csv + group fields (group_id, size, avg_contrib_xaf, urban_flag, founded_year).

Other files (optional): month_imputes.json, group_reliability.json — short-history imputation and train-only group blend if your app uses them; core scoring uses the joblib + meta.

Runtime: CPython 3.10–3.12 recommended; needs xgboost, scikit-learn, joblib, numpy, pandas.

Use / limits: Prioritization / exploration only — not an automatic lending decision. Simulated labels; no fairness audit on protected attributes (not in schema). Hyperparameters hand-tuned, not full HPO.

Privacy: Scores only with member consent and minimum necessary identifiers.