Mountain Fire – Observing the predictive power of the Bellwether wildfire tool

Modeling complex natural systems like wildfires in a way that provides actionable insights for insurance underwriting is a challenge. Advanced machine learning tools provide amazing new capabilities, but in the adoption phase of these new tools there can be reluctance to trust the results, especially when the approaches are novel and feel like a black box. Comparing the predicted versus actual wildfire perimeter gives us an opportunity to validate the predictive power of these new tools and increases confidence in the results. The recent Mountain Fire in Ventura County, California is an example where we see agreement between predicted wildfire probability and actual wildfire burn perimeter.

The Mountain Fire ignited the morning of 6 November 2024 and quickly exploded to over 14,000 acres (5,600 hectares) in under 12 hours, fueled by 25+ MPH (40+ KMH) winds and humidity in the teens. The fire worked through a mix of open space, exurban ranches, $1M+ suburban homes, and agricultural lands. In total, the fire burned nearly 20,000 acres and destroyed 182 structures.

The risk assessment platform CatNet® provides forward-looking wildfire risk intelligence for the US and Canada from Bellwether, a team at Alphabet's innovation engine X, the moonshot factory. Bellwether's advanced machine learning map harnesses nearly 600 layers of geospatial data, including detailed information on vegetation canopy, wind patterns, road networks, and past burn events of all kinds to forecast wildfire risk one year and five years into the future. Examining the fire perimeter in conjunction with the year-ahead wildfire probability from Bellwether, we see a high degree of agreement between actual versus predicted. Of particular interest is the burn to the southwest that is disconnected from the original fire by a full 2 miles (3.4 kilometers), which is separated by fields of irrigated agriculture.

Even though the Bellwether map is not explicitly modeling the spread of the Mountain Fire incident or ember spread, it implicitly captures these risks by assessing the probability of a burn occurring in 2024 on a pixel-by-pixel basis with data available up to 2023. Learnings from past wildfire events anticipate conditions observed in the Mountain Fire, specifically that a fire could ignite in one area and skip over lower probability areas and ignite again in disconnected higher probability areas. We have observed similar findings where a wind-driven wildfire "jumped" over a similar distance in the McDougall Creek Fire near Kelowna, British Columbia, Canada in 2023. 

This congruence between predicted versus actual is a powerful example of how the Bellwether wildfire tool available in CatNet® can help you understand the wildfire risk to communities and properties, be it for underwriting and pricing for property insurance, establishing community wildfire mitigation strategy, or understanding the potential impact to a real estate or mortgage portfolio.

Learn more about CatNet® and the Bellwether wildfire probability and contact us.