Kettle

California's wildfire crisis has fundamentally broken the reinsurance market. Fifteen of the twenty largest wildfires in state history have occurred in the last two decades — 2018 and 2020 each setting records — yet the industry still relies on 30-year-old models that treat every fire as a simple ellipse defined by length and width. These models cannot identify which specific properties are at risk, so reinsurers effectively price as if a quarter of the state might burn. The result: a 60% drop in return on equity for the reinsurance industry over the past decade, with carriers either dramatically overpricing coverage or exiting California markets entirely — leaving homeowners uninsurable and threatening the broader mortgage market.

Kettle partnered with Tribe AI to rebuild wildfire risk modeling from the ground up using computer vision and satellite imagery. Rather than treating the problem as a single prediction task, the team decomposed it into two distinct models: an ignition model estimating the probability a fire starts at any given location, and a contagion model predicting how a fire will spread across the landscape. Engineers with Google Earth Engine and geospatial expertise integrated satellite imagery showing where historical fires actually burned — training the models on real burn behavior rather than idealized assumptions. A second researcher applied techniques from biomedical imaging to segment and highlight at-risk zones within satellite frames. Together, the models enumerate probabilistic burn scenarios across all addresses 12 months in advance, generating a full distribution of outcomes rather than a single-point forecast.

Tribe's work more than doubled the performance of Kettle's existing wildfire spread model — which was already significantly above the industry average. Key outcomes include:

• >2× improvement in spread model performance over the prior Kettle baseline
• Model generates probabilistic scenario enumerations across all lat/long coordinates for the upcoming fire season — not conditional 24-hour forecasts
• Addresses the core industry failure: distinguishing the 0.1% of structures actually at risk from the broader population that incumbents price as uniformly exposed
• Positioned Kettle to underwrite more accurately in the $10 billion wildfire reinsurance market, with one of the engineers subsequently joining Kettle full-time after the engagement

• Adjacent domain expertise often outperforms industry specialists for novel ML problems — satellite imagery and biomedical imaging skills transferred directly; traditional actuarial backgrounds would not have.
• Decompose complex predictions into sub-models: separating ignition probability from fire contagion made each component more tractable and independently improvable.
• Probabilistic enumeration beats point prediction for high-variance catastrophe risk — the goal is an accurate probability distribution across scenarios, not a single forecast.
• Training on observed outcomes, not parametric rules, is critical: using actual historical burn perimeters from satellite data captured fire behavior that ellipse-based assumptions systematically missed.
• Reinsurers should treat pricing model accuracy as a competitive moat, not a commodity — better predictions directly translate to the ability to write profitable business others cannot.