Case Study: Using Historical Transaction Data to Reduce Chargebacks

blog.maxmind.com

9 points·antitamper··3 comments

I've worked in fraud prevention for large e-commerce sites for the last decade. Surprised these types of stories are still noteworthy enough to trigger a blog post by Maxmind.

However I'd love if this triggered some comments and questions from people who are interested in this kind of thing.

yeah, i'd figure most large sites have automated processes for this too, often with at least some machine learning stuff.

doing the analysis and cancellations manually in excel, without backtesting, feels super irresponsible (she tested it in production!!)

"How could Jenn be sure that the transactions she was canceling were fraudulent transactions? Jenn anticipated that legitimate customers would complain; however, she never heard from anyone whose transactions were canceled for this reason."

wat

A large number of e-commerce shops still rely on heuristics (rules of thumb) to manage their fraud risk. Accertify is a major software vendor in the e-commerce space and their framework doesn't incorporate machine learning. Their sales pitch is that they give tons of flexibility to write a rule that adds or subtracts point from the score. My response was always how do I know how many points this rule should be? Once you look under the hood at most shops you see that everything is + or - 1 million and essentially the scoring engine becomes a simple pass/fail engine.

The super-majors like Amazon are heavy into machine learning and rebuild some of their models daily. But even with their prowess they still had a thousand people doing manual investigation a couple of years ago. One subtle signal that you have triggered their fraud system is that they will ask you to renter your card number.

Reply on news.ycombinator.com