Predicting fires for policy making: Improving accuracy of fire brigade allocation in the Brazilian Amazon

The positioning of federal fire brigades in the Brazilian Amazon is based on an oversimplified prediction of fire occurrences, where inaccuracies can affect the policy's efficiency. To mitigate this issue, this paper attempts to improve fire prediction. Firstly, a panel dataset was built at municipal level from socioeconomic and environmental data. The dataset is unparalleled in both the number of variables (48) and in geographical (whole Amazon) and temporal breadth (2008 to 2014). Secondly, econometric models were estimated to predict fire occurrences with high accuracy and to infer statistically significant predictors of fire. The best predictions were achieved by accounting for observed and unobserved time-invariant predictors and also for spatial dependence. The most accurate model predicted the top 20% municipal fire counts with 76% success rate. It was over twice as accurate in identifying priority municipalities as the current fire brigade allocation procedure. Of the 47 potential predictors, deforestation, forest degradation, primary forest, GDP, indigenous and protected areas, climate and soil proved statistically significant. Conclusively, the current criteria for allocating fire brigades should be expanded to account for (i) socioeconomic and environmental predictors, (ii) time-invariant unobservables and (iii) spatial auto-correlation on fires.