The Pricing Model Was Working Against the Portfolio

A pricing analysis for a major national property and casualty insurer. We're sharing it because the pattern isn't unique to them.

A major national commercial P&C insurer came to us with what they thought was a routine data question: was their pricing model capturing the full risk picture across their commercial property book?

The models ran clean. The answer did not.

The Property Claims Picture

We built three models to predict where claims were most likely to come from, then checked how well they performed against policies the models had never seen before.

The idea: if you could only look closely at 25% of the book, which 25% should it be? If the model has no skill, any random group of 25% will contain about 25% of the eventual claims. If the model has skill, it should do better than that.

Major storm claims — riskiest 25% of policies contained 40% of all storm claims

Liability claims — riskiest 25% of policies contained 41% of all liability claims

Everyday property claims — riskiest 25% of policies contained 48% of all property claims

A random 25% of their book would contain 25% of claims. The models found significantly more than that.

The everyday-property result was the sharpest: the riskiest 25% of the book held nearly twice the claim exposure of any random 25% selected from the same book. Worth paying attention to. What made it consequential was what the pricing looked like underneath.

Loss ratio is how much an insurer pays out in claims per dollar of premium collected. A 100% loss ratio means they're breaking even on claims alone, before expenses. Here's what the numbers showed for their everyday property book:

Riskiest 25% of the property book — loss ratio 98.1%

Safest 25% of the property book — loss ratio 33.5%

Premium increase needed to make the riskiest group sustainable +96%

The 96% figure reflects both current mispricing and years of compounding. It's not a rounding error.

The safest group was running at 33.5%: profitable, and almost certainly over-priced relative to any competitor with more precise risk data. In a competitive market, over-priced accounts shop at renewal. Their best customers were the ones with the most reason to leave.

The Liability Picture

The liability results ran backwards from what the pricing model assumed.

Accounts priced as low-risk — loss ratio 75.4%

Accounts priced as high-risk — loss ratio 58.8%

The accounts being charged less (because the model thought they were safer) were actually filing more claims per dollar of premium than the ones being charged more. The pricing was inverted.

When we looked at what was actually predicting liability claims, it came down to things like this: businesses with any OSHA inspection record filed liability claims at twice the rate of uninspected ones (14% vs. 7%). Industry type mattered significantly too; manufacturing and agriculture/forestry accounts filed claims at a rate above 17%, compared to 6% for financial services businesses.

These signals were available. They weren't in the pricing model.

Where the Gap Comes From

The insurer's existing vendor data was built to answer a different question: who is this business, and are they creditworthy? That's genuinely useful. Standard business credit data gives you identity, financials, and payment history at scale.

Predicting physical property risk requires a different layer of information entirely: what does the property look like, what's around it, what does the business actually do, and what does their safety record show? Those signals aren't in a credit report. Swarmalytics maintains 16,000+ data points per commercial property across the US, updated continuously. That's what the models ran on.

The bottom line: the pricing model was charging too little for the accounts most likely to generate losses, and too much for the accounts least likely to. It’s both problems at the same time.

Why This Doesn't Show Up On Its Own

An insurer that discovers a large segment of its commercial book is mispriced doesn't put that in a press release. They fix it quietly. The incentive to disclose is zero; the incentive to move fast is everything.

When every insurer in the market is working from the same sources, and everyone carries similar results, nobody sees the gap. They're all benchmarking against each other's blind spot. You only see it when someone runs the analysis with different data.

This insurer didn't know the pattern existed until the models ran. That's not a failure on their part. The information they needed wasn't in the data they had.

Is your book showing this pattern too?

If you're running a commercial P&C book and want to know what's there, we can show you. The same methodology. Your own data, at a different resolution.