Of flagged transactions, the share that are fraud.
MODEL EVALUATION / SYNTHETIC TRANSACTIONS
Where should we
draw the line?
Change the review threshold and see how many suspicious transactions are caught, missed, or flagged unnecessarily.
—held-out transactionsNo real customer data
DECISION CONTROL
Review threshold
More reviews · fewer missesFewer reviews · more misses
Transactions with a model score at or above this threshold are flagged for review. A flag is not a confirmed fraud finding.
Of fraud transactions, the share flagged.
Transactions requiring a human review.
Fraud transactions below the threshold.
OUTCOMES
Confusion matrix
—Fraud flagged
—Legitimate flagged
—Fraud missed
—Legitimate cleared
Changing the threshold changes the decisions shown here; it does not retrain the model.
MODEL CARD
How this was built
- Model
- Scaled logistic regression, class weighted
- Split
- —
- Average precision
- —
- ROC AUC
- —
Threshold selected for F1 on the validation set. The metrics above use a separate test set. This synthetic benchmark is a demonstration, not a production fraud system.
REVIEW QUEUE
Flagged transactions
| Transaction | Score | Amount | Distance | Last hour | New device | Cross border | Actual label |
|---|
Showing the highest-scored flagged transactions.