In this paper, we model and predict state-specific wildfire losses in the US using a combination of Bayesian dynamic models. In particular, the wildfire frequencies are modeled by a Bayesian multi-scale Dynamic Count Mixture Model (DCMM), which is capable of capturing a number of stylized features of wildfire data, including zero-inflation, over-dispersion compared to the Poisson distribution, and the time-varying patterns. Further, the DCMM is able to incorporate spatial dependence of different states, and thus improves the forecasting performance for individual states, especially those with low historical frequencies. The predictive distribution of future wildfire loss is then used to price wildfire catastrophe (CAT) bonds, the hedging effectiveness of which is evaluated for insurers in different states. We find that although using CAT bonds as a hedging tool may slightly increase the expected liability of an insurance portfolio due to bond premiums, the strategy can substantially reduce the variability risk and tail risk. Our studies suggest that CAT bond is a valuable tool of wild-fire risk mitigation for insurers. Finally, for index-based CAT bonds whose payoffs are linked to wildfire losses in a larger area than that the insurer operates in, their hedging efficiencies are still satisfying. Therefore, it may be beneficial for insurers, especially those operating in areas with less frequent yet more volatile wildfire losses, to issue index-based CAT bonds, which are likely to be less expensive but much more liquid than indemnity bonds written directly on their liabilities.

(This is joint work with Jianxi Su from the Department of Statistics, Purdue University, West Lafayette, U.S.A.)