A dataset and benchmark for forecasting real-world time series with cause-driven irregularities and multimodal observations.

A Fine-grained Capability Assessment Framework for Time Series Forecasting Models Using Synthetic Data

We propose a prune-then-finetune paradigm to specialize time series foundation models in both levels of architecture and parameter weights, effectively improving downstream forecasting performance.

We pioneer the idea of implicit decomposition and propose a powerful decomposition-based enhancement framework based on it, which can consistently enhance the forecasting performance of various state-of-the-art time series models.

The paper designs a training objective to suppress the learning of redundant features in time series models.

We formalize the over-squashing phenomenon in spatiotemporal graph neural networks and analyze how it affects information propagation across the spatial and temporal dimensions.