On the best-performed time window size for homicide count forecasting

In the last two decades, violence and homicides have been consistently increasing in Mexico; the official data shows relations to other crimes and time-dependent territorial patterns. Even though this phenomenon has been studied from different perspectives, from social sciences to geostatistical studies, to the best of our knowledge, machine learning and computational intelligence methods have not been applied to the Mexican case. This work presents a two-fold contribution; first, we compare the Bayesian Ridge Regression, Linear Regression, Multilayer Perceptron, Ridge Regression, Linear Regression with Stochastic Gradient Descent, and Support Vector Machine, for forecasting the number of homicides for the future three months, in the 15 most violent municipalities of Mexico using as predictors a database of 50 crimes that are monthly released by the Mexican Government. The methods' performance varies according to the predictors selected or extracted from the database and the time window size. Hence, second, we use feature engineering techniques for improving their performance by exhaustively testing combinations of techniques. The resulting statistical measurements are useful to find directions about the best-performed method, the best sliding or rolling time window, and the adequate functions for feature extraction.