Abstract:

The reliable prediction of school enrollment demand is of paramount importance for infrastructure planning and resource allocation within the public education system. Currently, Machine Learning models applied to this domain rely on historical school flow series and demographic variables, assuming stationarity in social patterns. However, the increasing frequency of extreme climate events in Brazil imposes external shocks that may render such models obsolete or biased in vulnerable regions. This work proposes a methodology to validate the viability of Gradient Boosting algorithms (LightGBM and XGBoost) in the face of climate variability. By cross-referencing open microdata from the School Census (INEP) with public calamity records from the Integrated Disaster Information System (S2iD) and socioeconomic indicators from IBGE, the study aims to quantify the discrepancy in predictive accuracy, based on root mean square error (RMSE) and mean absolute error (MAE) metrics, between municipalities with climate stability and those affected by recurrent disasters. The central hypothesis is that traditional models suffer significant performance degradation in 'Risk Zones,' failing to capture enrollment volatility caused by forced migrations or academic disruptions. The research seeks to propose the integration of climate variables as a requirement for the new generation of educational management tools.

Keywords:

School Enrollment Prediction, Machine Learning, Climate Events, Educational Data Mining, Public Policy.