Hierarchical Mixed-Effects and Stacked Machine Learning Ensembles with Data Augmentation for Leakage-Safe E-Waste Forecasting

Consumer electronics, with 62 million tons of electronic waste (e-waste) generated in 2022 and e-waste expected to grow to 82 million tons annually by 2030, pose critical challenges when it comes to national infrastructure and circular economy policies. This paper compares forecasting approaches using sparse panel data for 32 European countries (2005–2018, Eurostat/Waste Electrical and Electronic Equipment (WEEE) Directive), focusing on leakage-safe prospective validation to guarantee true predictive performance. We make one-step-ahead predictions with conservative features (primarily lagged values) to account for temporal autocorrelation but with reduced multicollinearity (Variance Inflation Factor (VIF) ≈ 1.0). Cross-paradigm comparisons such as time-series baselines Autoregressive Integrated Moving Average (ARIMA), Seasonal ARIMA (SARIMA), Long Short-Term Memory (LSTM), hierarchical mixed-effects models, pooled machine learning (9 methods), and block-bootstrap-augmented stacking ensembles demonstrate stacking’s effectiveness, with a weighted validation R² of 0.992 for held-out 2017–2018 data. Time-series approaches demonstrate negligible predictive power (mean R² = −9683) given non-stationarity and limited samples, while the hierarchical approach provides virtually no benefit (Intraclass Correlation Coefficient (ICC) 0.011) amidst computational instability. Bootstrapping improves high-variance tonnage forecasts (Root Mean Squared Error (RMSE) reductions of 18.6%) while being detrimental to stable units, thus reinforcing parsimony. Feature ablation validates that only a few lags are necessary, preventing leakage from rolling means or calendar trends. Our method enables conservative year-ahead forecasts with quantified uncertainty, conservative estimates of e-waste management, allowing for buffer planning and policies even when data is scarce. By using strictly out-of-sample tests rather than biased ones, this work characterizes achievable year-ahead performance under sparse annual panels.