Forecasting intentional wildfires using temporal and spatiotemporal autocorrelations
Abstract
We report daily time series models containing both temporal and spatiotemporal lags, which are applied to forecasting intentional wildfires in Galicia, Spain. Models are estimated independently for each of the 19 forest districts in Galicia using a 1999–2003 training dataset and evaluated out-of-sample with a 2004–06 dataset. Poisson autoregressive models of order P – PAR(P) models – significantly out-perform competing alternative models over both in-sample and out-of-sample datasets, reducing out-of-sample root-mean-squared errors by an average of 15%. PAR(P) and static Poisson models included covariates deriving from crime theory, including the temporal and spatiotemporal autoregressive time series components. Estimates indicate highly significant autoregressive components, lasting up to 3 days, and spatiotemporal autoregression, lasting up to 2 days. Models also applied to predict the effect of increased arrest rates for illegal intentional firesetting indicate that the direct long-run effect of an additional firesetting arrest, summed across forest districts in Galicia, is –139.6 intentional wildfires, equivalent to a long-run elasticity of –0.94.