Building a Fraud Detection System That Actually Works

The dirty secret of payment fraud detection is that most systems are optimized for the wrong metric. They measure catch rate — what percentage of fraudulent transactions they block. But the real cost isn’t fraud. It’s false positives.

The False Positive Tax

Every legitimate transaction your fraud system blocks is a customer who won’t come back. Studies show that 33% of customers who experience a false decline will never attempt to purchase from that merchant again. That’s not a security problem. That’s a revenue problem.

At Meridian, we built our ML pipeline around a different objective: maximize merchant revenue while keeping fraud below their risk threshold.

The Architecture

Our system processes each transaction through three layers:

Velocity checks — Pattern matching against the card’s recent history
Device intelligence — Browser fingerprinting, IP reputation, and behavioral signals
Ensemble model — Gradient-boosted trees trained on 2.1B historical transactions

The key insight is that these layers run in parallel, not sequentially. A transaction gets a risk score from each layer simultaneously, and the final decision is a weighted combination.

Results

For merchants who migrated to Meridian’s fraud system:

False positive rate: Reduced from 2.4% to 0.3%
Fraud loss rate: Maintained at 0.001%
Revenue recovered: Average $47K/month per merchant from previously blocked legitimate transactions

Open-Sourcing Our Approach

We’ve published our feature engineering methodology and a sanitized training dataset on GitHub. The model weights themselves remain proprietary, but the architecture and approach are fully documented.