Abstract
In recent years, multivariate modeling using time series foundation models (TSFMs) has gained traction, achieving advanced zero-shot generalization. However, modern multivariate TSFMs are primarily pretrained on synthetic data, which, while easier to scale, may fail to capture the complex temporal dynamics and cross-variable relationships inherent in real-world time series. This raises a crucial question: to what extent do leading TSFMs trained with a real-world corpus outperform those trained with synthetic data?
To address this, we establish the RMISC corpus, a considerably large-scale, high-quality, openly accessible, real-world, multivariate time series archive containing around 200 datasets and 142 billion time points across diverse domains. Furthermore, we pretrain four advanced TSFMs on univariate, synthetic multivariate, and real-world multivariate data, evaluating their zero-shot generalization capabilities on standard in-distribution and out-of-distribution benchmarks. Experimental results demonstrate that incorporating real-world multivariate data significantly enhances the generalization performance for both univariate and multivariate TSFMs. These findings provide deeper insights into how real-world multivariate data contributes to the development of stronger TSFMs.
Blogger's Review: This paper constructs the RMISC corpus and explores the importance of real data versus synthetic data in time series modeling, offering valuable empirical support for research in multivariate time series models. The results highlight the critical role of real data in improving model generalization, suggesting that future research should focus more on effectively leveraging real-world data.