A machine learning proposal to predict poverty

Due to the high rate of inclusion and exclusion errors of traditional methods (Proxy Mean Test) used for the identification of households in poverty condition and selection of the social assistance programs beneficiaries, this research analyzed different perspectives to predict households in poverty condition, using a machine learning model based on XGBoost. The models proposed were compared with baseline methods. The data used were taken from the 2019 household survey of Costa Rica. The results showed that at least one of our approaches using XGBoost gave the best balance between inclusion and exclusion errors. The best model to predict poverty and extreme poverty was build using an XGBoost with a classification approach