Research shows that some categories of human-ignited wildfires might be forecastable, due to their temporal clustering, with the possibility that resources could be pre-deployed to help reduce the incidence of such wildfires. We estimated several kinds of incendiary and other human-ignited wildfire forecast models at the weekly time step for tribal land units in the United States, evaluating their forecast skill out of sample. Analyses show that an Autoregressive Conditional Poisson (ACP) model of both incendiary and non-incendiary human-ignited wildfires is more accurate out of sample compared to alternatives, and the simplest of the ACP models performed the best. Additionally, an ensemble of these and simpler, less analytically intensive approaches performed even better. Wildfire hotspot forecast models using all model types were evaluated in a simulation mode to assess the net benefits of forecasts in the context of law enforcement resource reallocations. Our analyses show that such hotspot tools could yield large positive net benefits for the tribes in terms of suppression expenditures averted for incendiary wildfires but that the hotspot tools were less likely to be beneficial for addressing outbreaks of non-incendiary human-ignited wildfires.
Keywords: arson; autoregressive; human-caused; incendiary; law enforcement; wildfire hotspotting.