A DEPLOYABLE REAL-TIME SERVICE FOR EARLY CARD-FRAUD SCORING WITH FEW LABELS AND A SELF-SUPERVISED NOVELTY INDICATOR

Abstract

Fraud detection models must start working when only a few confirmed fraud

labels exist, and they must score each payment within the time budget of card authorisation. This paper describes the software design of a real-time fraud scoring service that meets both requirements. The risk model was selected on a validation period, without using test data, from candidates that included self-supervised representations; the selected model is a random forest on 31 decision-time features. The self-supervised component is kept as a separate novelty indicator. On 284 807 real card transactions, a model trained on the labels of only the first three hours of operation reached an area under the precision–recall curve of 0.770 on transactions more than 33 hours later. The service has authenticated REST endpoints, schema validation, checksum-protected model artifacts and mid-rank calibration of scores, and its 99-th percentile HTTP latency was 9.6 ms.