RMISC corpus: 142 billion points across ~200 time series

RMISC, a large-scale real-world multivariate corpus for time series foundation models, was submitted to arXiv on 7 Jul 2026 by Qian Sun, Yong-Ming Tian, Jia-Wei Huang, Cheng Feng and Shao-Qun Zhang (arXiv:2607.06504). The corpus contains around 200 datasets and 142 billion time points, and the authors use it to test whether pretraining on real-world multivariate data improves TSFM generalization.

What is RMISC and what does it contain?

RMISC is an openly accessible archive of real-world multivariate time series, and it contains around 200 datasets totaling 142 billion time points drawn from diverse domains. The paper presents RMISC as a considerably large-scale, high-quality, real-world, multivariate corpus intended for pretraining and evaluating time series foundation models.

The authors describe the corpus as covering diverse domains but do not enumerate every dataset in the abstract; they emphasise scale and real-world multivariate structure as the distinguishing features versus commonly used synthetic multivariate data. The arXiv record lists the title, authors, and the submission date of 7 Jul 2026 alongside the dataset claims.

How did the authors pretrain and evaluate time series foundation models?

The team pretrained four advanced time series foundation models (TSFMs) on three kinds of data: univariate data, synthetic multivariate data, and real-world multivariate data, then evaluated zero-shot generalization on standard in-distribution and out-of-distribution benchmarks. This direct comparison tests whether multivariate real-world pretraining yields stronger zero-shot performance than the synthetic alternatives.

Specifically, the paper reports that the experiments involve four pretrained TSFMs and distinct pretraining regimes (univariate, synthetic multivariate, real-world multivariate). Evaluation focused on zero-shot generalization across established benchmark splits labelled in the abstract as in-distribution and out-of-distribution. The authors conclude that incorporating real-world multivariate data predominantly improves generalization performance for both univariate and multivariate TSFMs.

Why it matters

RMISC gives researchers a single, large-scale real-world corpus to test whether the gains claimed for synthetic pretraining hold up when models see genuine multivariate temporal dynamics. If real-world multivariate pretraining consistently improves zero-shot generalization, then model builders will have clearer guidance on what data to prioritise when scaling TSFMs.

A corpus of around 200 datasets and 142 billion time points shifts the scaling discussion from synthetic volume toward empirical variety and cross-variable structure. That shift affects benchmark design, dataset curation, and the choice of pretraining feeds for teams aiming to deploy TSFMs across domains where zero-shot transfer matters.

What to watch

Watch for community uptake of RMISC and for follow-up papers that reproduce the abstract's core claim that real-world multivariate pretraining predominantly improves zero-shot generalization: the concrete signals will be independent replications of the authors' experiments, release of the full RMISC archive and associated code or pretrained checkpoints, and comparative studies that report numeric zero-shot gains across the same in-distribution and out-of-distribution benchmarks described in the paper.

The arXiv submission (arXiv:2607.06504) gives the dataset scale and the experimental outline; the next concrete milestones are public availability of the corpus and detailed benchmark numbers from the authors or from independent teams.