Random Forests for Wildfire Insurance Applications

Homeowners’ insurance in wildfire-prone areas can be a very risky business that some insurers may not be willing to undertake. We create an actuarial spatial model for the likelihood of wildfire occurrence over a fine grid map of North America. Several models are used, such as generalized linear models and tree-based machine learning algorithms. A detailed analysis and comparison of the models show a best fit using random forests. Sensitivity tests help in assessing the effect of future changes in the covariates of the model. A downscaling exercise is performed, focusing on some high-risk states and provinces. The model provides the foundation for actuaries to price, reserve, and manage the financial risk from severe wildfires.