RMISC: A Large-scale Real-world Multivariate Corpus for Time Series Foundation Models
Time Series Foundation Models (TSFMs) have become a powerful tool for forecasting, but they are currently limited by how they are trained. Most modern models rely on synthetic data—data generated by computer algorithms—because it is easy to scale. However, synthetic data often fails to capture the complex, messy, and interconnected relationships found in real-world data. This paper introduces the RMISC corpus, a massive, high-quality collection of real-world multivariate time series data, to determine if training models on authentic data leads to better performance than training them on synthetic alternatives.
The RMISC Corpus
The researchers developed the Real-world Multivariate tIme Series Corpus (RMISC) to provide a more accurate foundation for training TSFMs. The archive contains approximately 200 datasets and 142 billion time points across diverse fields, including energy, finance, environment, industry, and traffic. Unlike synthetic data, which relies on simplified mathematical assumptions, RMISC captures the actual temporal dynamics and cross-variable dependencies—such as how wind speed and rainfall collectively influence temperature—that are essential for accurate real-world forecasting.
How the Corpus Was Built
Creating the RMISC was a rigorous five-stage engineering process. It began with sourcing data from open and legal real-world archives. The team then processed the data to handle noise, missing values, and outliers. They unified the data into a consistent, hierarchical structure to make it easy for researchers to use. A key part of the process involved creating a standardized metadata system, which ensures that every dataset is traceable and includes information about its origin and licensing. Finally, the team refined the data to ensure it was reliable and balanced across different domains, providing both a full version and a smaller, domain-balanced version for easier experimentation.
Impact on Model Performance
To test the value of this new corpus, the researchers pretrained four advanced TSFMs—Chronos-2, GTT, Moirai-2.0, and TimesFM-2.5—using three different approaches: univariate data, synthetic multivariate data, and the real-world RMISC data. The experiments revealed that models trained with real-world multivariate data consistently outperformed those trained on synthetic or univariate data. Specifically, incorporating real-world data improved the models' ability to generalize to new, unseen scenarios (out-of-distribution performance). The study concludes that the best results come from a balanced recipe: combining real-world univariate data, synthetic multivariate data, and the new real-world multivariate RMISC data.
Key Takeaways
The findings suggest that while synthetic data is useful for scaling, it is not a complete substitute for the complexity of the real world. By providing an openly accessible, large-scale, and high-quality archive, the authors aim to help developers build more robust TSFMs that can handle the intricate relationships present in actual time series data. The RMISC corpus is now available to the research community to support further development and benchmarking of multivariate foundation models.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!