A Tutorial On Chaotic Dynamics For Economists

Purpose: This paper examines whether stacking (Wolpert, 1992) improves short-term forecasts of global oil consumption and electricity demand during periods of major shocks, including COVID-19, geopolitical conflicts, and trade disruptions. Given diverging trends between OECD and non-OECD economies, we evaluate forecast performance separately across these regions.

Design/methodology/approach: We combine XGBoost, random forest, elastic net, multi-layer perceptron, and a unit fixed-effects model within a stacking framework using a non-negative OLS meta-learner. We implement both a top-down approach, which models aggregate demand directly, and a bottom-up approach, which models disaggregated components and aggregates predictions, to assess the value of decomposition.

Findings: We find that stacking provides a robust benchmark, consistently matching or outperforming the best individual learner across most specifications. In the top-down setting, stacking generally delivers the strongest performance, particularly for global and non-OECD oil demand and global electricity demand. Bottom-up stacking improves accuracy for most oil sub-aggregates and performs comparably to the best base learner for electricity demand by source. Comparing strategies, bottom-up stacking performs better for electricity demand, while both approaches perform similarly for oil demand. Results by demand size suggest that retransformation bias from the log scale (Duan, 1983) and cross-country heterogeneity are particularly important for large economies. Overall, while no single model dominates across all settings, stacking provides a simple and reliable approach to improving forecast accuracy.

Originality/value: Our contribution is threefold. First, we provide a systematic comparison of stacking methods for forecasting global energy demand across OECD and non-OECD economies. Second, we evaluate the relative performance of top-down and bottom-up strategies within a common framework. Third, we examine how forecast performance varies with the scale of energy demand, highlighting the role of retransformation bias (Duan, 1983) and cross-country heterogeneity.