L-MTSF: large-scale datasets for re-evaluating long-term multivariate time series forecasting

Multivariate time series (MTS) are prevalent and play a critical role in sensor networks across diverse domains like transportation, environmental monitoring, and energy. Accurate long-term forecasting of these time series is essential for many tasks, such as resource planning and risk management, attracting increasing attention. Numerous models have been proposed, from Transformer-based architectures to linear models. However, most evaluations rely on widely used benchmark datasets that can be considered small-scale. This constraint may hinder models from revealing their true capabilities, especially for advanced models requiring more training data to capture intricate dependencies. This scarcity of large-scale datasets prevents researchers from comprehensively evaluating models and analyzing performance across different training data sizes. Motivated by this, we curate a collection of large-scale benchmark datasets, including extensions of two popular existing ones, enabling more representative, fair, and scalable evaluations. This paper provides detailed descriptions of our data curation, comprehensive analyses of dataset characteristics, and extensive experiments uncovering insights not observable on small-scale datasets. We believe this work will encourage greater attention toward data-centric perspectives in long-term MTS forecasting and support developing more robust and generalizable models.