Software:Chronos (pretrained model)

This template is being used in the wrong namespace. To nominate this talk page for deletion, go to Miscellany for deletion.

Chronos is a framework for pretrained probabilistic time series models developed in 2024 by Amazon Web Services researchers. By tokenizing time series values through scaling and quantization into a fixed vocabulary, Chronos utilizes existing transformer-based language model architectures, specifically training them via cross-entropy loss. The model variations within the Chronos family are based on the T5 family, with sizes ranging from 20 million to 710 million parameters.

Development

Chronos was pretrained on a broad array of publicly available datasets, alongside a synthetic dataset created through Gaussian processes to enhance its ability to generalize across different tasks. This approach allowed Chronos to undergo comprehensive pretraining, preparing it for a wide range of forecasting applications.

Performance

In an extensive benchmark covering 42 datasets—which included both traditional local models and modern deep learning approaches—Chronos demonstrated notable achievements:

Training Corpus Performance: On datasets included in its training corpus, Chronos models significantly outperformed competing methods.

Zero-Shot Performance: When tested on new datasets, without specific training, Chronos displayed comparable or occasionally superior performance relative to models that were trained on those specific datasets.

Impact

The results from various benchmarks highlight Chronos's capability to leverage time series data from diverse domains. This enables it to improve zero-shot accuracy on unseen forecasting tasks markedly. The introduction and success of Chronos models mark a significant step forward, suggesting that pretrained models can serve as effective and simplified solutions in forecasting pipelines across a range of fields.