XGBoost and MAPA's Budgetary Practice: Influence on HDI and Agribusiness GDP

This study applies the XGBoost algorithm, with hyperparameter optimization via GridSearchCV, to model two target variables — percentage variation of Agribusiness GDP (%GDP-AGRO) and the Human Development Index (HDI) of Brazilian states — from MAPA's open budgetary data (2018–2022), integrated with variables from the 2017 Agricultural Census (IBGE) and the CEPEA/USP historical series. The dataset comprises 465 observations across federal budget units (MAPA, EMBRAPA, INCRA, CONAB, SFB). Feature selection employed Recursive Feature Elimination (RFE) and multicollinearity analysis.

The %GDP-AGRO model achieved R² = 0.8764 (RMSE = 0.00956); the HDI model achieved R² = 0.9632 (RMSE = 0.00712). In both targets, the percentage of rural properties receiving technical guidance emerged as the most important feature (28.18% and 52.11%, respectively), followed by investments in Projects/CAPEX (21.59%) in the GDP model and geographic location — Southeast region (31.89%) — in the HDI model. The results provide empirical evidence for data-driven public policy formulation in the agricultural sector.

This record includes two versions:
- English version: artigo_xgboost_mapa_EN.pdf
- Portuguese version: artigo_xgboost_mapa_PT.pdf

Source code and data: https://github.com/Jotta-se/MBA_USP
Acknowledgements: MBA in Data Science & Analytics, USP/ESALQ, 2024.