ReggieJOE/steg-fraud-detection-zindi

Zindi Africa — Fraud Detection in Electricity and Gas Consumption Challenge. Achieved ROC-AUC of 0.874 # STEG Fraud Detection — Zindi Africa ## Challenge Detect fraudulent manipulation of electricity and gas meters using 14 years of billing history for 135,000 clients in Tunisia. **Metric:** ROC-AUC **My Score:** 0.874 **Leaderboard Rank:** #161 ## Approach - Cleaned and merged two datasets: client profiles + 4.5M invoice records - Engineered 80+ features per client: consumption stats, monthly patterns, yearly trends, zero-consumption ratios, account duration, fraud rate encoding - Trained XGBoost + LightGBM ensemble with 5-fold cross-validation - Submitted raw probabilities (not binary labels) for better ROC-AUC scoring ## Key Learning Submitting raw probabilities instead of hard 0/1 predictions jumped my score from 0.669 to 0.871 in a single submission — ROC-AUC is rank-based. ## Files - `STEG_Fraud_Detection.ipynb` — Full pipeline notebook - `submission_v5.csv` — Best submission file - `STEG_Fraud_Detection.pptx` — Project presentation ## Tools Python · XGBoost · LightGBM · Scikit-learn · Pandas · Google Colab