Homicide forecasting for the state of Guanajuato using LSTM and geospatial information

Abstract

In the last years, intentional homicides have increased significantly in Mexico. A proven strategy to confront the problem is applying predictive methods used to anticipate the resources and logistics of the security corps. This work tackles the forecasting of intentional homicides using three forecasting methods: ARIMA, LSTM, and NeuralProphet, applied to the 16 municipalities of Guanajuato state with the highest count. The approach is replicable to all Mexico's municipalities since the same data are reported. We conducted an exhaustive search of optimal hyper-parameters of the LSTM and an exhaustive search for the optimal lag for NeuralProphet. In the same regard, different combinations of neighboring municipalities were tested to include geospatial information. The methods are compared via MAE, MSE, and bootstrap hypothesis tests. LSTM improved with geospatial data, so the best LSTM model showed a superior performance to the ARIMA by 23.1% in the MAE and 35.6% in the MSE. On the other hand, NeuralProphet showed a similar performance to the ARIMA according to the bootstrap hypothesis test, showing no statistically significant difference between them. The results show that the phenomenon is related to the spatial context and encourage the use of geospatial information in forecasting models.