DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models

Authors: Jinpeng Li, Zhongyi Pei, Huaze Xue, Bojian Zheng, Chen Chia Wang, Jianmin Wang Date: 2026-02-25 Paper ID: openalex:2602.22066

Summary

DualWeaver is a novel framework designed to bridge the gap between successful univariate Time Series Foundation Models (Uni-TSFMs) and multivariate forecasting tasks. It achieves this by introducing a pair of learnable, structurally symmetric surrogate series, generated by a shared module that explicitly models cross-variable dependencies. These surrogates are adapted to be compatible with the input structure of the TSFM via the forecasting objective. A key innovation is the parameter-free reconstruction of the final multivariate predictions directly from the symmetric surrogates, avoiding complex, task-specific parametric decoders. Empirical results across various real-world datasets demonstrate that DualWeaver achieves superior accuracy and stability compared to existing state-of-the-art multivariate forecasters.

Key Contributions

• Proposal of DualWeaver, a framework that repurposes univariate Time Series Foundation Models (TSFMs) for multivariate forecasting via learnable surrogate series.
• Introduction of a shared auxiliary feature-fusion module that captures cross-variable dependencies to generate the surrogate series.
• Implementation of a parameter-free prediction reconstruction mechanism leveraging the symmetric structure of the surrogates.
• Inclusion of a theoretically grounded regularization term designed to enhance adaptation robustness and prevent collapse during the transfer from univariate to multivariate tasks.

Limitations

The paper focuses on adapting existing TSFMs; the performance is inherently dependent on the quality of the pre-trained univariate model. Further investigation into the theoretical limits of surrogate representation capacity might be beneficial.

Open Questions & Future Work

• extensibility-of-feature-fusion-module

Key Concepts

• DualWeaver: A novel framework that adapts univariate Time Series Foundation Models (Uni-TSFMs) for multivariate forecasting using structurally symmetric surrogate series.

Datasets

• real-world datasets

Limitations

The paper focuses on adapting existing TSFMs; the performance is inherently dependent on the quality of the pre-trained univariate model. Further investigation into the theoretical limits of surrogate representation capacity might be beneficial.

Links

• Abstract
• PDF