Spatio-Temporal Modeling of Wildfire Risks in the U.S. Forest Sector

In the US forestry industry, wildfire has always been one of the leading causes of damage. This topic is of growing interest as wildfire has caused huge losses for landowners, residents and governments in recent years. While individual wildfire behavior is well studied (e.g. Butry 2009; Holmes 2010), a lot of new literature on broadscale wildfire risks (e.g. by county) is emerging (e.g. Butry et al. 2001; Prestemon et al. 2002). The papers of the latter category have provided useful suggestions for government wildfire management and policies. Although wildfire insurance for real estate owners is popular, the possibility to develop a forestry production insurance scheme accounting for wildfire risks has not yet been investigated. The purpose of our paper is to comprehensively evaluate broadscale wildfire risks in a spatio-temporal autoregressive scenario and to design an actuarially fair wildfire insurance scheme in the U.S. forest sector. Our research builds upon an extensive literature that has investigated crop insurance modeling. Wildfire risks are closely linked to environmental conditions. Weather, forestland size, aspects of human activity have been proved to be crucial causal factors for wildfire (Prestemon et al. 2002; Prestemon and Butry 2005; Mercer et al. 2007). In light of these factors, we carefully study wildfires ignited by different sources, such as by arson and lightning, and identify their underlying causes. We find that the decomposition of forestland ecosystem and socio-economic conditions have significant impacts on wildfire, as well as weather. Our models provide a good fit to data on frequency and propensity for fires to exist (e.g. R-square ranges from 0.4 to 0.8) and therefore provide important fundamental information on risks for the development of insurance contracts. A number of databases relevant to this topic are used. With the Florida wildfire frequency and loss size database, a complete survey of four measurements of annual wildfire risks is implemented. These four measurements are annual wildfire frequency, burned area, fire per acre and burned ratio at county level. In addition, the national forestry inventory and analysis (FIA) database, Regional Economic Information Systems (REIS) database and the national weather database have supplied forestland ecosystem, socioeconomic, and weather condition information respectively. With our spatio-temporal lattice models, impacts of environmental factors on wildfire and implications of wildfire management policies are assessed. Forestland size, private owners’ share of forestland, population and drought would positively contribute to wildfire risks significantly. Cold weather and high employment are found to be helpful in lessening wildfire risks. Among the forestland ecosystem, oak / pine & oak / hickory forestland would reduce wildfire risks while longleaf / slash & loblolly / shortleaf pine forestland would have a mixed impact. An interesting finding is that oak / gum / cypress forestland would reduce wildfire frequency, but would enhance wildfire propensity at the same time. Hurricanes could intensify wildfire risks in the same year, but would significantly decrease the next year’s wildfire risks. Meanwhile, cross sample validation verifies that our method can forecast wildfire risks adequately well. Since our approach does not incorporate any fixed-effect indicator or trend as in the panel data analysis (Prestemon et al. 2002), it offers a universal tool to evaluate and predict wildfire risks. Hence, given environmental information of a location, a corresponding actuarially fair insurance rate can be calculated.