Evaluating the Relationship Between Temporal Aggregation and Forecast Bias in Statistical Time Series Models

Posted: Nov 02, 2022

• Claire Cook
• Colton Phillips
• Connor Edwards

Abstract

The practice of temporal aggregation represents a fundamental preprocessing step in time series analysis, routinely employed to reduce noise, manage computational requirements, and align data with decision-making timeframes. Despite its widespread application across domains including economics, environmental science, and operations management, the systematic effects of aggregation choices on forecast accuracy remain inadequately characterized. Current literature predominantly focuses on aggregation's impact on forecast variance and computational efficiency, while largely overlooking its nuanced relationship with forecast bias. This research gap is particularly consequential given that biased forecasts can lead to suboptimal decisions with significant real-world implications, from inventory mismanagement in supply chains to inefficient resource allocation in energy systems. Our investigation addresses this gap by examining how different levels of temporal aggregation systematically influence forecast bias across multiple statistical modeling approaches. We challenge the conventional assumption that bias remains relatively constant across aggregation levels and instead propose that aggregation induces non-linear transformations in the data generating process that fundamentally alter forecast distributions. The novelty of our approach lies in developing a comprehensive framework for bias decomposition that separates aggregation effects from inherent model limitations, enabling practitioners to make more informed decisions about temporal resolution selection.