Linear vs. Machine learning approaches for cross-temporal forecast reconciliation with an application to Italian energy load data
Accurate electricity load forecasting is crucial for efficient power grid management. Forecast reconciliation has emerged as a key technique to improve forecast coherence and accuracy, particularly within hierarchical and constrained time series frameworks. This paper compares linear and machine learning (ML) approaches for crosstemporal forecast reconciliation, focusing on Italian energy load data. The study evaluates the performance of both methods, emphasizing the role of covariance matrix estimation in forecast accuracy. Our results, measured using the Mean Absolute Scaled Error (MASE), indicate that linear reconciliation methods, when incorporating validation errors for covariance estimation, outperform machine learning-based alternatives, highlighting the importance of enhanced error estimation in forecasting. The findings provide insights into the trade-offs between model complexity and predictive accuracy, offering valuable implications for energy demand forecasting and broader time series reconciliation applications.