Beyond Holistic Models: Systematic Component-level Benchmarking of Deep Multivariate Time-Series Forecasting

Researchers introduced TSCOMP, a large-scale benchmark for deep multivariate time-series forecasting.

It decomposes models into preprocessing, encoding, architecture, and optimization components, then evaluates 20,000+ model-dataset combinations.

The results show that carefully choosing simpler components can outperform manually designed complex models.

TSCOMP also creates a large performance corpus to support automated model selection and zero-shot model construction.