You must build a binary classifier for fraud with a 0.2% positive rate and 10M rows × 500 features. Propose an end-to-end plan that covers: 1) data splitting with stratification and leakage prevention; 2) handling imbalance (class weights vs. focal loss, down/over-sampling, SMOTE variants, and when to use each); 3) appropriate metrics and why (PR curve, AUPRC, recall at fixed precision, cost-sensitive metrics) vs. why ROC-AUC is misleading; 4) threshold setting using cost matrices and calibration (Platt/Isotonic) and how you’d do post-deployment threshold tuning; 5) algorithm choices and justification (baseline logistic with class_weight, tree ensembles with balanced subsampling, anomaly detection fallback); 6) robust validation (time-based CV, group CV), data drift monitoring, and rejection rules for extreme edge cases; 7) a brief pseudocode of the training/evaluation loop that scales to this dataset.