My previous chargeback posts are some of the most viewed and commented on my blog, and for good reason: merchants often get the short end of the stick when it comes to chargebacks. In the middle of last year it became clear that we finally needed to implement some sort of semi-automated fraud detection system, but not for the reason you might suspect.
Chargebacks weren’t up, but our order volume was, and that meant that the time we spent investigating orders for potential fraud was way up. Time that was now more valuable than ever given our growth. Our system for “flagging” and investigating orders was largely based upon investigating any orders that we deemed “suspicious”, and the criteria for “suspicious” had a large element of personal opinion. We have trained everyone to look out for fraud, so sometimes we’d have several people “investigating” the order at different steps of the process before it left our warehouse.
Over the next few months we developed and implemented a system to automate the process of flagging orders for us to review, and then we created a structured system for deciding whether or not the order was fraud.
The real challenge is identifying patterns for fraudulent orders. Before we could even get started we needed to see if there were any patterns, because if there weren’t any system we built wouldn’t be very good. We looked at about 20 different criteria for each chargeback we’ve received, as well as for each order we’ve canceled due to suspected fraud. Luckily, and somewhat surprisingly to me, we found a handful of criteria that highly correlated with chargebacks and very rarely occurred in non-fraudulent orders. Armed with that knowledge we could flag orders for closer investigation that met those specific criteria. (For obvious reasons I’m not going to list out exactly what criteria we looked at, but it’s mostly the same things that any fraud-prevention service like Kount would look at, in addition to a few custom ones we came up with for our business.)
Once we flag an order a report is created for us. The report includes every possible subjective and objective measure of fraud we could come up with. It is one persons job to review these orders and approve or reject them. We also have the ability to manually flag an order for review (if say a warehouse employee notices something odd) and to override the system and flag every order for a customer if they’ve been suspect in the past. I designed the system to be flexible – it’s easy to add or change criteria as time moves on and fraud changes, as it inevitably does. We’ve made a few tweaks to the formula already.
We’ll never eliminate chargebacks. We hope we’ve reduce them, but even that wasn’t really the goal of this project. The goal was to save a bundle of our collective time, and that we have definitely accomplished.
- Apple Pay and Google Pay Now Live – Step 2 of Our Braintree Integration
- We Recently Migrated Detailed Image to Braintree Payments: The Good & Bad From This Large, Unplanned Project
- How Twitter, WordPress, and Google Contacts Influenced Notify My Team’s UI
- PageSpeed vs reCAPTCHA and YouTube: Will Google Ever Get on the Same Page?