Fraud probability
Live demo · real public dataset · nothing calls a model
Fraud Detection.
A fraud model built to UK bank model-validation practice, with a plain-English reason for every declined payment.
The score is not the interesting part. The judgement calls are: why the leaderboard winner was not chosen, where the cost-optimal threshold broke on later data, and what the explanations can and cannot tell a customer.
Every figure below is read from the pipeline's own output files. For the full interactive version, open the Streamlit dashboard ↗ (it can take about 30 seconds to wake).
At a glance · out-of-time test set · deployment threshold
Fraud caught
of
Alert precision
of alerts are real fraud
Alerts per 10,000
transactions sent to review
KS / AUPRC
/
ranking quality
01 · Score a transaction
One payment, one decision, one reason
Each sample is a real transaction from the final, out-of-time test window, which the model never saw in training. Its score and SHAP values were computed once, with the same code the API runs; this page replays them.
Why
SHAP contribution · log-odds
Toward fraud — may be cited as a reason
Away from fraud — never cited
What the customer receives
Want to change the amount or time of day and watch it flip? Open the dashboard ↗
02 · Threshold & cost
Where to draw the line
A missed fraud and a false alarm do not cost the same, so the threshold is set by expected cost, not F1. It was chosen on the validate set and then frozen before the test set was scored.
Fraud caught
False alarms
Total cost
The two thresholds, frozen
| Threshold | Set | Fraud caught | Precision | Cost |
|---|
Stated honestly
03 · Champion
The leaderboard is not the decision
Three models trained under identical out-of-time conditions. On the test set, Random Forest ranks best — the table shows it.
| Model | Validate | Test | ||||
|---|---|---|---|---|---|---|
| KS | AUPRC | F1 | KS | AUPRC | F1 | |
Bold = best in column
Then apply the deployment threshold.
04 · Drift & monitoring
What a monitor should catch
Each split is a later slice of time, so differences between them are real temporal shift.
Dashed lines: 0.10 moderate · 0.25 significant shift
Retraining trigger · Airflow drift_check
05 · The full dashboard
Explore it yourself
- Score a Transaction Edit amount, time of day and the top risk indicators, and watch the decision change.
- Threshold & Cost The same cost trade-off, with the full curves.
- Drift and Champion Rationale PSI, KS curves, calibration and feature importance.
Real public dataset (anonymised card transactions) · nothing you do calls a paid service.