We Found Insurance Fraud in Our Crash Data

Summary

We Found Insurance Fraud in Our Crash Data April 22, 2025 When we set out to build geospatial risk scores for vehicle crashes at Matrisk AI, we never expected that a side by side look at Vehicle Identification Numbers and crash timelines would hint at possible insurance fraud. But data sometimes surprises you. Below, I’ll walk through how we stumbled upon this discovery, what we found, and why it might matter for anyone insuring vehicles. A curious hunch Our main focus has been risk scores: Where are crashes most frequent? What are the road conditions? Which stretches of highway see severe outcomes? Yet, a handful of states in our dataset disclose VINs for each crash, and that simple addition changed everything. Why VINs are a big deal They let us track exactly which vehicle is in multiple crashes, not just a vague “car in two accidents.” They allow us to approximate when that vehicle might have switched insurance carriers. They open the door to spotting suspicious patterns (like the same VIN appearing in an unreasonable cluster of accidents). I still remember the moment this hit home. Some years ago, a police officer casually told me, “You should get a dashcam, insurance fraud is common around here.” His offhand comment stuck with me, but life moved on. Fast forward to our modern data pipeline, and suddenly those words didn’t seem so casual after all. Linking crashes together After filtering out invalid VINs, we narrowed the dataset to roughly ~15 million crashes. (We also removed all drug and alcohol related crashes, since it’s unlikely someone committing insurance fraud would be under the influence.) From there, our pipeline: Counts how often the same VIN appears in a short interval (e.g., 6–12 months). Flags overlapping or “back to back” insurance coverage for the same VIN. Identifies repeated patterns of single vehicle collisions, nighttime crashes, and reported injuries. We’re not conducting a law enforcement sting, but we do want to spot anomalies that might ...