![[2509.20928] Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2509.20928] Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting
Summary
The paper introduces Conditionally Whitened Generative Models (CW-Gen) for probabilistic time series forecasting, addressing challenges like non-stationarity and inter-variable dependencies through innovative modeling techniques.
Why It Matters
This research is significant as it enhances predictive performance in time series forecasting, a critical area in machine learning, by incorporating informative priors. The proposed models can better handle distribution shifts and complex inter-variable relationships, making them valuable for real-world applications.
Key Takeaways
- CW-Gen improves time series forecasting by integrating conditional means and covariances.
- The Joint Mean-Covariance Estimator (JMCE) enhances predictive quality by learning both mean and covariance simultaneously.
- CW-Diff and CW-Flow extend the CW-Gen framework, demonstrating superior performance on various datasets.
- The models effectively mitigate distribution shifts, improving reliability in dynamic environments.
- Empirical results show CW-Gen outperforms traditional generative models in capturing complex data dynamics.
Statistics > Machine Learning arXiv:2509.20928 (stat) [Submitted on 25 Sep 2025 (v1), last revised 16 Feb 2026 (this version, v2)] Title:Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting Authors:Yanfeng Yang, Siwei Chen, Pingping Hu, Zhaotong Shen, Yingjie Zhang, Zhuoran Sun, Shuai Li, Ziqi Chen, Kenji Fukumizu View a PDF of the paper titled Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting, by Yanfeng Yang and 8 other authors View PDF HTML (experimental) Abstract:Probabilistic forecasting of multivariate time series is challenging due to non-stationarity, inter-variable dependencies, and distribution shifts. While recent diffusion and flow matching models have shown promise, they often ignore informative priors such as conditional means and covariances. In this work, we propose Conditionally Whitened Generative Models (CW-Gen), a framework that incorporates prior information through conditional whitening. Theoretically, we establish sufficient conditions under which replacing the traditional terminal distribution of diffusion models, namely the standard multivariate normal, with a multivariate normal distribution parameterized by estimators of the conditional mean and covariance improves sample quality. Guided by this analysis, we design a novel Joint Mean-Covariance Estimator (JMCE) that simultaneously learns the conditional mean and sliding-window covariance. Building on JMCE, we introduce Conditionally W...
