Education as an equalizer? AI analysis exposes the enduring grip of socioeconomic status on academic achievement across schooling levels

Socioeconomic status (SES) shapes academic trajectories from the earliest years of schooling. This study examines whether its predictive power on grade retention varies across the three cycles of basic education in Portugal and how the relative influence of SES components shifts across levels. Data from 31,777 students were analysed using Logistic Regression and CatBoost, trained with stratified fivefold cross-validation over 10 seeds to predict grade retention from five Graffar indicators. SES was most predictive in the 1st Cycle (CatBoost AUC = 0.831; Logistic Regression AUC = 0.823) and declined slightly across cycles. Parental education dominated in the 1st and 3rd Cycles, while parental occupation emerged as the primary predictor in the 2nd Cycle. Yet both models retained meaningful discriminative power in the 3rd Cycle (CatBoost AUC = 0.701, recall = 71.1%), suggesting that compensatory policies have not substantially disrupted the structural link between family background and academic outcomes.