TimeOmni-o1 combines text reasoning and time series forecasting in one architecture

Researchers report that TimeOmni-o1 combines a language model with a pretrained time series model to interpret text and numerical sequences together and generate both text and forecasts. Their preprint reports better forecasting results than dedicated time series models on three open benchmarks. The time series model compresses numerical sequences for the language model, which reasons over them alongside accompanying text. That reasoning then feeds back into the time series model to guide its forecast. Training uses MMTR, a dataset covering 12 domains with 500k+ samples and annotated reasoning traces. On MMTR, the combined architecture outperforms its individual component models, with gains across most domains rather than every domain.