Amazon Chronos repo released
Amazon published Chronos, a repository of pretrained language-model–based forecasting models aimed at time-series tasks, offering a starting point for production forecasting systems. Pretrained LM approaches to forecasting can simplify candidate generation and contextual patterning in sequential data, which is useful for product ML problems like personalization or demand prediction. Having a ready repo lowers the bar for prototyping model-backed forecasting in recsys and product pipelines. (x.com)
Most forecasting at big companies still starts with a long column of numbers: daily sales, hourly traffic, weekly returns. The hard part is turning that column into a guess about the next 7 or 30 points without hand-building a custom model for every dataset. (amazon.science) Chronos takes a shortcut from language models. It turns numbers into tokens, which are small code-like pieces, the same way a language model turns words into tokens before it predicts the next word. (amazon.science) To do that, Chronos first rescales a series so a tiny store and a giant warehouse can be described in the same numeric “units.” It then quantizes the values, which means it bins nearby numbers together the way a thermometer rounds temperature into marked lines. (arxiv.org) Once the series is converted into tokens, Chronos trains a transformer model to predict what token comes next. Amazon’s original paper says the first Chronos family used Text-to-Text Transfer Transformer models ranging from 20 million to 710 million parameters. (arxiv.org) That setup lets one pretrained model start forecasting on a new dataset without task-specific retraining. The paper describes this as zero-shot forecasting, and Amazon reported gains on unseen datasets after pretraining on public time-series collections plus synthetic data generated with Gaussian processes. (amazon.science) The news now is that Amazon’s Chronos code is easy to grab and run from a public repository instead of living only as a paper and model idea. The official repository, `amazon-science/chronos-forecasting`, is live on GitHub under an Apache-2.0 license and includes notebooks, scripts, and packaged code. (github.com) That repository is not just a research dump. The README says it supports the original Chronos models, the faster Chronos-Bolt release from November 27, 2024, and the newer Chronos-2 release from October 20, 2025. (github.com) Chronos-Bolt is Amazon’s speed play. Amazon says those models cut error by 5 percent and can run up to 250 times faster while using 20 times less memory than the original Chronos models of the same size. (github.com) Chronos-2 is the broader forecasting tool. Amazon says it handles one series, many series, and forecasts that include outside variables in one model, which is useful when demand depends on things like price, promotions, or weather. (github.com) For product teams, that changes the starting point. Instead of spending weeks building candidate models for every demand-planning or recommendation-adjacent forecasting problem, they can begin with a pretrained model, test it on their own history, and fine-tune only if the baseline is good enough. (aws.amazon.com) The bigger shift is that forecasting is starting to look like the rest of machine learning infrastructure. When Amazon ships a repo, PyPI package, Hugging Face checkpoints, and deployment guides around one forecasting family, it turns a research paper into something an engineer can plug into a pipeline this week. (pypi.org)
