Microsoft paired machine learning with optimization to cut e-commerce fraud

Every fraud filter has to balance letting bad orders through against rejecting legitimate ones, and a filter tuned to stop fraud tends to block good customers too.

Microsoft's system combined sequential machine learning to spot both known and emerging fraud with an optimization layer that weighed fraud loss, false positives and the revenue at stake. That let it raise the bank acceptance rate while lowering the fraud loss rate.

Over 2016 to 2018 the project reduced Microsoft's fraud loss by 0.52 percentage points, generating about $75 million in annual savings and over $1 billion in additional revenue from more legitimate purchases going through. The work was a finalist for the 2019 INFORMS Franz Edelman Award.

Flag fraud and keep the good order moving

Modeling both fraud and good-sale loss exposes the real cost of each wrong decision.

Sequential models catch new fraud patterns that static rules miss.

Raising the acceptance rate of legitimate purchases grows revenue while shrinking fraud losses.

An explicit optimization replaces a hand-tuned threshold that was over-rejecting.

A fraud system judged on losses alone over-rejects; adding the cost of a lost good sale makes the whole decision cheaper.

The system was deployed across Microsoft's storefronts and became a reference for how online retailers pair machine learning with operations research in fraud management. It was named a 2019 Edelman finalist and documented in an INFORMS journal paper.