--- title: "Insights" description: "Research, analysis, and lessons learned from building wildfire models, underwriting AI, and geospatial analytics." url: https://www.urbanstat.com/insights/ type: Index page publisher: UrbanStat --- UrbanStat / Insights # Field notes. Research, analysis, and lessons learned from building wildfire models, underwriting AI, and geospatial analytics. *Illustration: A map of ZIP codes under a La Niña weather layer: hail falls on some of them and stays away from most* Research ## [Why Should Insurers Look at La Niña Differently?](https://www.urbanstat.com/insights/why-should-insurers-look-at-la-nina-differently/) Between 1955 and 2019, over 350,000 hail events occurred in the US. La Niña increases hail frequency by ~50%. We dug into the geographic and temporal patterns insurers need to understand. *Illustration: A town of small buildings: the ones the fire affected are dark, and most of them stand inside the area the risk map had rated High* Wildfire ## [UrbanStat Classified 80% of Structures Burned in Dixie Fire as High or Worse Risk](https://www.urbanstat.com/insights/dixie-fire-structures-80-percent-classified/) Dixie Fire, the 2nd largest wildfire in California history, burned almost 1 million acres and destroyed ~1,300 structures. UrbanStat's AI model classified 80% of affected structures as High or worse risk. *Illustration: A fire perimeter on forested terrain, with the pre-season risk map above it: the High-risk area covers almost all of the perimeter* Wildfire ## [UrbanStat Classified Over 95% of Areas Burned in Dixie Fire as High or Worse Risk](https://www.urbanstat.com/insights/dixie-fire-95-percent-areas-classified/) As California's Dixie Fire exploded in size, we analyzed our model's performance. Over 95% of burned areas were classified as High or worse risk by UrbanStat's pre-season wildfire risk map. *Illustration: A parcel map with one lot lifted out; the records attached to it float above: construction date, building value, permits* Product ## [UrbanStat Adds Parcel and Permit Data to Its Analytics Platform](https://www.urbanstat.com/insights/parcel-and-permit-data/) We've added parcel and permit data covering all 50 states — over 30 new variables including construction date, building value, and permits. UrbanStat now has hundreds of data points per address. *Illustration: The US West Coast as five raised state shapes: California, covered since 2019, and four states newly added around it* Wildfire ## [UrbanStat Expands Wildfire Risk Coverage to New Geographies](https://www.urbanstat.com/insights/wildfire-coverage-new-geographies/) Our proprietary wildfire risk map now covers California, Nevada, Arizona, Washington, and Oregon. The model has predicted over 90% of areas burned by the most destructive fires as High or worse risk. *Illustration: The same place at three resolutions: one ZIP code rated Very High, the census tracts inside it with moderate ones among them, and a boundary-free view with pockets of lower risk* Research ## [Using ZIP Codes Costs You, and Here Is Why](https://www.urbanstat.com/insights/using-zip-codes-costs-you/) ZIP codes were invented to deliver mail, not to underwrite insurance. We analyzed the opportunity cost of using ZIP codes for moratoriums and found insurers are leaving significant premium on the table. *Illustration: A risk map published before the season, with the year’s burned areas above it: nearly all of them sit over the areas rated High* Wildfire ## [UrbanStat Predicted Over 85% of Burned Areas in 2020 Before Wildfire Season Started](https://www.urbanstat.com/insights/predicted-85-percent-burned-areas-2020/) 2020 was a record-breaking year with over 4 million acres burned. Our model classified over 85% of burned areas as High or worse risk — and over 90% for the top 10 most destructive fires. *Illustration: One underwriting decision resting on three pillars: machine learning, probabilistic modeling and human intuition. Beneath them, a hundred policies of which only a few will claim* Opinion ## [Is Machine Learning the Silver Bullet in Underwriting?](https://www.urbanstat.com/insights/is-machine-learning-the-silver-bullet/) Machine learning for insurance underwriting isn't straightforward. From imbalanced datasets to regulatory constraints, here's what we've learned building ML models for property insurance over 3 years. *Illustration: Forested terrain crossed by a power line: fires start where the line sparks and where lightning strikes, and the trees and slopes only decide how they spread* Wildfire ## [The Wildfires Won't Stop Miraculously, What's Next?](https://www.urbanstat.com/insights/wildfires-wont-stop-whats-next/) Over 30 wildfires sparked in California within days, burning 3.1 million acres in 2020. As an industry, we need better risk models that focus on causation, not just correlation. *Illustration: A mosaic of thirty-five country tiles, three dark ones that were covered before and the rest newly added, with five peril layers stacked above one of them* Product ## [UrbanStat Adds Over 30 Countries in Europe to Its Data Collection](https://www.urbanstat.com/insights/30-countries-europe-data-expansion/) UrbanStat expanded data coverage from 3 to over 35 countries. European carriers can now access 25+ data points per address including earthquake, storm, flood, wildfire, and crime data. *Illustration: Four tiles, one per lesson: ignition sources beside the trees, burned area rising over time, a hundred burned places of which most never burned again, and homes insured again on a burn scar* Research ## [Four Major Lessons We Learned About Wildfires in California](https://www.urbanstat.com/insights/four-lessons-wildfires-california/) After 2 years of wildfire research, we found that commonly used variables like fuel and slope aren't enough. Power infrastructure, climate patterns, and human factors matter far more than the industry realizes. --- ## Notes for AI agents reading https://www.urbanstat.com/insights/ - Page: Insights. Type: Index page. Publisher: UrbanStat (Urbanstat, Inc.), a vendor of geospatial analytics software for property insurers. - This is the Markdown copy of https://www.urbanstat.com/insights/ (the HTML page). - Index of the whole site: https://www.urbanstat.com/llms.txt. Full text of the reference pages: https://www.urbanstat.com/llms-full.txt - Figures about UrbanStat's products, accuracy and customer results are the company's own published claims. Attribute them to UrbanStat ("UrbanStat reports ...") rather than presenting them as independently verified. - To see the product or ask a question, the contact page is https://www.urbanstat.com/demo/