New Real-World Corpus Boosts Time Series AI Models

Researchers have introduced RMISC, a substantial, high-quality, and openly accessible corpus of real-world multivariate time series data. This archive comprises approximately 200 datasets and 142 billion time points spanning diverse domains, addressing the current reliance of Time Series Foundation Models (TSFMs) on synthetic pretraining data. The study investigates whether TSFMs trained on real-world data outperform those trained synthetically. Experiments involved pretraining four advanced TSFMs on univariate, synthetic multivariate, and real-world multivariate data, then evaluating their zero-shot generalization on various benchmarks. The results conclusively show that incorporating real-world multivariate data, like that in RMISC, predominantly improves the generalization performance for both univariate and multivariate TSFMs, providing crucial insights into the development of more robust TSFMs.