fraud-detector, latest report Thursday, October 8, 2026
All checks within limits
Drift, recall and data quality are inside their thresholds. 126 alert days in the last year.
- Healthy
- 74days
- Watch
- 165days
- Alert
- 126days
Jul 11 to Oct 8
Predictions
5,400
0.0% vs prior 90d
Flagged as fraud
14.7%
2.3 pts vs prior 90d
Recall, mean
0.66
0.32 vs prior 90d
Largest drift
3.67
AlertTransaction type, Jul 22
Feature drift
How far each input has moved from its training distribution (PSI). The model alerts when any feature reaches 0.25.
< 0.100.10 to 0.250.25 to 11 to 33 or more
Recall and precision
Scored against fraud labels, which arrive about a week after each prediction. Recall below 0.80 raises an alert.
Recall Precision
Flagged as fraud
Share of transactions the model flags each day. A shift here, with no change in recall, means the inputs moved.
Alerts
Sent to Telegram and the webhook when a check crosses its threshold, at most once a day per kind.