FRAUD ANALYTICS · QUANTITATIVE RISK
Evaluate fraud risk without hiding the false-positive trade-off
A reproducible modelling prototype that generates fictional transaction data, learns a risk ranking, selects a decision threshold on validation data, and measures performance on a later holdout.
Decision problem
Fraud review systems face a trade-off: a low threshold may find more suspicious activity but overwhelm reviewers, while a high threshold may reduce alerts and miss more fraud. Model evaluation should make this trade-off visible rather than optimizing overall accuracy alone.
Workflow
Synthetic transaction generation → chronological split → model fitting → validation threshold selection → holdout scoring → metrics export.
Methods and metrics
- Balanced logistic regression as a transparent baseline.
- Temporal training, validation, and test partitions to reduce look-ahead risk.
- Threshold selected on validation data, not on the final test set.
- Average precision, ROC AUC, precision, recall, F1, false-positive rate, alert rate, and confusion counts.
Decision use
The outputs show how to compare model ranking quality and the resulting review workload. A real deployment should set the operating threshold using fraud-loss and review-cost assumptions, then monitor stability and calibration over time.
Reproducibility
The source script builds the synthetic dataset from a fixed random seed and writes the holdout metrics and predictions to an output directory.
Install the listed requirements and run the command in the project README to repeat the analysis.
Evidence boundary
All observations and labels in the demo are synthetic. The results validate the code path, not expected fraud detection in a bank or payment platform. There is no claim of production performance, client impact, or operational deployment. Before real use, the method would require a representative authorised dataset, leakage analysis, calibration, cost-sensitive thresholding, drift monitoring, and privacy/fairness review.