In their research paper, authors Abdul Fatir Ansari and his colleagues present Chronos, a novel framework for pretrained probabilistic time series models. By tokenizing time series values and training transformer-based language models on them, they demonstrate significant outperformance on training datasets and comparable, sometimes superior, zero-shot performance on new datasets. This approach showcases the potential of pretrained models in simplifying forecasting pipelines and improving accuracy on unseen tasks. The use of a synthetic dataset and diverse real-world data sets make their results valuable for a wide range of applications.
https://arxiv.org/abs/2403.07815