Forecasting K–12 Enrollment in the UAE: Classical Time-Series vs. Machine Learning Approaches

Accurate student enrollment projections are essential for education planning and teacher workforce allocation. We develop a comprehensive framework for forecasting K–12 enrollments in the United Arab Emirates (UAE), comparing classical time-series models (ARIMA, Prophet) with machine learning models (XGBoost, LSTM). Using national data from 2020–2025 disaggregated by region, gender, and school type, we report split-conformal prediction intervals (PIs) and empirical coverage alongside point forecasts. Because each regional series contains only five annual observations, we mitigate data scarcity by training a global XGBoost on pooled regional panels while keeping ARIMA/Prophet local and the LSTM intentionally small with strong regularization. In a one-year holdout (2024→2025), global XGBoost achieves the lowest error (MAPE ≈ 1.2%), outperforming ARIMA and Prophet (MAPE ∼2.8–3.0%). A rolling-origin 1-step cross-validation yields similar errors across methods (∼3% MAPE), underscoring small-sample difficulty and the value of pooling. Forecasts highlight stark regional differences—notably rapid growth in Ajman and Sharjah versus stagnation in Dubai—and we link projections to student–teacher ratios to assess capacity pressure. Without accelerated hiring, high-growth regions risk ratios exceeding 20:1 within five years. Contributions include: (i) an empirical comparison of classical and ML techniques on a national education dataset; (ii) a data-pooling strategy enabling ML gains under scarcity; and (iii) integration of staffing implications with enrollment forecasts to support proactive policy decisions.