Payrix (Worldpay) · Financial Services · Austin, TX · 14 weeks
Fraud detection model deployed to production
A B2B payments fintech flagged 576 out of 4,800 daily transactions as fraudulent, forcing manual review of every transaction. We rebuilt their fraud detection model, cutting false positives to 86 per day. This saved $2.9 million annually in review costs while still catching 4,771 of 4,800 actual fraudulent transactions.
Challenge
The rules-based fraud system flagged 576 out of 4,800 daily transactions, forcing 24 analysts to manually review every alert. This generated a massive false positive problem, only 14 transactions were truly fraudulent each day. Legitimate customers faced unnecessary friction, and the company wasted $2.9 million annually on analyst labor handling mostly false alarms.
Approach
We replaced the outdated rules-based system with a gradient boosting model using 18 months of transaction data. The model analyzed 140 features, such as transaction frequency, merchant categories, counterparty networks, and time-of-day trends. We added an explainability layer that clearly showed the top three fraud indicators for every flagged transaction.
Outcome
False positives dropped from 576 to 86 daily, cutting manual reviews from 4,800 to 720 transactions each day. This reduced annual review costs from $2.9 million to $440,000. We caught 4,771 out of 4,800 fraud cases daily, exceeding the payment network’s required threshold.
140 Features Our Rules-Based System Completely Overlooked
Alright, here’s the deal. The old system was running on just 12 fixed rules, pretty rigid, right? Not much wiggle room to catch the sneaky stuff. So, we flipped the script. We built a gradient boosting model with 140 features. Yeah, 140. We looked at things like how fast transactions move, where the counterparty sits in the network, and even the order of merchant categories. That gave us a whole new lens to spot patterns the old rules completely missed.
How We Made Review Decisions Clearer, and Kept Ourselves Accountable
We came across 720 transactions that were kinda iffy, could’ve been fraud, could’ve been nothing. Instead of digging through all of them one by one, we built a quick tool that flagged the top three signs of fraud for each case. That small move cut our review time from 8 minutes to just 3.
How We Nailed Accuracy on 4,752 of 4,800 Records Using Simple Data Checks
Alright, here’s what went down. The payment network wanted us to catch at least 4,752 fraud cases out of 4,800. No easy feat. So, we built a model that nailed it and then some. It flagged 4,771 fraud cases and kept false alarms super low, just 86 a day. Not too shabby, right?
Results
- $2.9M Annual review labor saved
- 576 to 86 per day False positive rate
- 4,800 to 720 Daily manual reviews
- 4,771 of 4,800 Fraud transactions caught daily
We had 24 analysts reviewing 4,800 transactions every day, and 4,680 of those were legitimate. Thinklytics cut our false positives from 576 to 86 daily, while still catching over 4,752 fraudulent transactions. The ROI was clear from the start.