Network of Patterns: Time Series Forecasting with Pattern Passing

ICLR 2026 Conference Submission7201 Authors

Keywords: Time Series Forecasting, Decomposition, Multi-scale

Abstract:

Time series contain diverse pattern information, and many studies have leveraged these patterns to enhance representations for more accurate forecasting. A key challenge lies in how to organize multi-scale patterns for effective information aggregation. Previous studies typically partition sequences into multi-scale pattern segments and organize them into chain or tree structures, employing neural networks to aggregate features and improve predictive performance. However, information transmission in chain structures is strictly linear and accumulative, while tree structures can aggregate multiple patterns but remain constrained by hierarchical limitations. Moreover, segments at the same or neighboring scales do not necessarily exhibit strong dependencies.

To overcome these limitations, we propose the Network of Patterns (NoP), which flexibly connects all relevant pattern segments to enable interactions between any nodes. We further introduce a Pattern Passing strategy to efficiently propagate and aggregate pattern information across this network, achieving more comprehensive integration. Experimental results demonstrate that NoP not only effectively encapsulates informative pattern signals but also establishes new state-of-the-art performance on multiple time series forecasting benchmarks, surpassing chain- and tree-based methods.

Primary Area: learning on time series and dynamical systems

Submission Number: 7201