PREDIÇÃO DE SINISTROS AGRÍCOLAS: UMA ABORDAGEM COMPARATIVA UTILIZANDO APRENDIZAGEM DE MÁQUINA

Abstract

Crop insurance has gained greater attention in Brazil since the beginning of the past decade, with the implementation of the Rural Insurance Premium Subvention Program. The present study tested the performance of Machine Learning algorithms for insurers to forecast the occurrence of a claim, using data from policies and climate databases between the years of 2006 and 2017. The Random Forest, Support Vector Machine and k-Nearest Neighbors algorithms were tested, and the second method showed a better predictive performance of claims when evaluated by the metrics Accuracy, Precision, Positive and Negative True Rates and Matthews Correlation. However, all methods presented a low predictive capacity for the occurrence of claims.