Comparação entre modelos de regressão estatísticos e modelos de aprendizado de maquina interpretáveis na modelagem de risco no seguro agropecuário

Abstract

Agricultural insurance is a relevant instrument for mitigating climatic and production risk in the Brazilian agribusiness sector, with claim severity modeling being a central element for insurers' pricing and risk management. This study aimed to compare regression models and interpretable machine learning models in predicting the severity of rural insurance claims, evaluating both predictive performance and the interpretability of the approaches. A database of 190,423 policies with claim occurrence was used, drawn from the Rural Insurance Premium Subsidy System (SISSER), covering the period from 2016 to 2024. Four models were estimated: Multiple Linear Regression (MLR), Generalized Linear Model (GLM) with Gamma distribution, Random Forest, and XGBoost, evaluated through cross-validation and compared using the MAE, RMSE, and sMAPE metrics. The interpretability of the machine learning models was assessed using the SHAP (SHapley Additive exPlanations) technique. Results showed that the machine learning models outperformed the statistical ones in predictive performance, with an approximate 12% reduction in RMSE compared to the Gamma GLM, albeit with greater relative overfitting between training and testing; among them, XGBoost achieved the lowest MAE and RMSE, and Random Forest the lowest sMAPE. The guarantee limit and the year 2021, marked by severe drought, stood out as the main determinants of severity across all models, a result that converged between the statistical and machine learning approaches. Segmenting the database between agricultural and livestock activities produced no performance gain, given the predominance of agricultural policies in the sample. It was concluded that combining machine learning models with explainability techniques constitutes a viable alternative for reconciling predictive accuracy and transparency in agricultural risk modeling.