xgboost-fraud-detection-calibrated

Calibrated XGBoost Fraud Detection Model

This repository contains a production-ready, highly optimized, and calibrated XGBoost model designed for out-of-time financial fraud detection under severe class imbalance.

Model Summary

• Framework: XGBoost + Scikit-Learn (Isotonic Regression Calibration)
• Validation Strategy: Out-of-Time (OOT) Temporal Split
• Target Metric Optimized: Precision-Recall AUC (PR-AUC)

Performance Metrics (Out-of-Time Test Set)

• OOT ROC-AUC: 0.97274
• OOT PR-AUC (AUPRC): 0.72298
• Expected Calibration Error (ECE): 0.000%

Usage

To use this model in your environment, download the consolidated .joblib artifact which bundles the model weights, the probability calibration mapping, and the exact production feature order.

import joblib
import xgboost as xgb
from huggingface_hub import hf_hub_download

# 1. Download from Hub
model_path = hf_hub_download(
    repo_id="xNub/xgboost-fraud-detection-calibrated", 
    filename="calibrated_fraud_model.joblib"
)

# 2. Load the pipeline components
artifacts = joblib.load(model_path)
model = artifacts["model"]
calibrator = artifacts["calibrator"]
features = artifacts["features"]

# 3. Predict & Calibrate
# dmatrix = xgb.DMatrix(df[features])
# raw_predictions = model.predict(dmatrix)
# calibrated_probabilities = calibrator.predict(raw_predictions)