![[2602.22066] DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.22066] DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models
Summary
The paper presents DualWeaver, a novel framework that enhances multivariate forecasting using univariate time series foundation models through learnable surrogate series.
Why It Matters
This research addresses the challenge of extending univariate forecasting success to multivariate scenarios, which is crucial for improving predictive accuracy in various applications such as finance, healthcare, and environmental monitoring. The proposed method demonstrates significant advancements over existing techniques.
Key Takeaways
- DualWeaver adapts univariate time series models for multivariate forecasting.
- It utilizes learnable surrogate series to capture cross-variable dependencies.
- The framework includes a regularization term to enhance robustness.
- Extensive experiments show DualWeaver outperforms state-of-the-art methods.
- Code for the framework is publicly available for further research.
Computer Science > Machine Learning arXiv:2602.22066 (cs) [Submitted on 25 Feb 2026] Title:DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models Authors:Jinpeng Li, Zhongyi Pei, Huaze Xue, Bojian Zheng, Chen Wang, Jianmin Wang View a PDF of the paper titled DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models, by Jinpeng Li and 5 other authors View PDF HTML (experimental) Abstract:Time-series foundation models (TSFMs) have achieved strong univariate forecasting through large-scale pre-training, yet effectively extending this success to multivariate forecasting remains challenging. To address this, we propose DualWeaver, a novel framework that adapts univariate TSFMs (Uni-TSFMs) for multivariate forecasting by using a pair of learnable, structurally symmetric surrogate series. Generated by a shared auxiliary feature-fusion module that captures cross-variable dependencies, these surrogates are mapped to TSFM-compatible series via the forecasting objective. The symmetric structure enables parameter-free reconstruction of final predictions directly from the surrogates, without additional parametric decoding. A theoretically grounded regularization term is further introduced to enhance robustness against adaptation collapse. Extensive experiments on diverse real-world datasets show that DualWeaver outperforms state-of-the-art multivari...
