Shadow mode runs a fraud model on live transactions while withholding its outputs from decisions. For an autoencoder, this allows suspicious transactions flagged by reconstruction error to be validated and thresholds tuned without directly affecting customers. section 8.4.6
Use shadow results to:
- Set alert thresholds against investigation capacity and acceptable customer friction. section 8.4.6
- Monitor alert volume, useful findings, and score distributions; recalibrate as data and fraud patterns change. section 8.4.6
Before a real-time model goes live, validate it on live traffic in shadow mode for at least two weeks. Pair this check with monitoring, A/B testing readiness, and a tested rollback plan. section 8.4.7
