Previsão de homicídios no Brasil: Proposta de variável antecedente

In this article we investigate the capacity of antecedent variables, among them admissions for aggression, in the prediction of the number of homicides in Brazil. The main objective is to eliminate a gap regarding the large time lag in the dissemination of information about these deaths in the country, allowing updated conjuncture analyzes. For this, through a 'rolling window' scheme and 'Model Confidence Set' approach, we investigate whether multivariate models with leading variables show forecast performance superior to a set of univariate models. In applying the MCS approach, considering different evaluation statistics, loss functions and estimation windows, we find strong evidence of the ability of the leading variables used to provide additional information content in the prediction of the Brazilian criminal dynamics, with models of leading variables systematically surpassing univariate models, especially in extended periods of forecasting. In general, improvements related to the benchmark Random walk model, in terms of RMSE, MAE and MAD, are of the order of 60%.