// the find
amazon-science/chronos-forecasting
Chronos: Pretrained Models for Time Series Forecasting
Chronos is Amazon's family of pretrained transformer models for time series forecasting, treating the problem like language modeling — tokenize the series, predict next tokens, sample trajectories for probabilistic outputs. Chronos-2 extends this to multivariate and covariate-informed forecasting. It's for ML engineers who want a strong zero-shot baseline without training their own model from scratch.
The Bolt variants are a genuine engineering win — patch-based encoding with direct multi-step decoding gets you 250x faster inference and 20x less memory than the original T5-based models at slightly better accuracy, which matters a lot when you're running inference at scale. The model size ladder (8M to 710M parameters) lets you pick your accuracy/latency tradeoff rather than being stuck with one option. Chronos-2's covariate support is the headline feature — most zero-shot forecasting models are univariate-only, so being able to pass known future features (holidays, promotions) is a real differentiator for demand forecasting use cases. The evaluation infrastructure is unusually honest: they publish raw per-dataset CSVs for every model variant, not just cherry-picked aggregate numbers.
The deployment story is AWS-first to the point of being a soft lock — the production recommendations all point to SageMaker or AutoGluon-Cloud, and while you can run it anywhere PyTorch runs, the docs treat self-hosting as an afterthought. Chronos-2's multivariate support is new and the architecture is more complex; the 120M parameter cap means you're not getting GPT-4-scale capacity here, and on datasets with many correlated series it may underperform models trained specifically on that domain. The original Chronos T5 models are essentially legacy at this point but still take up most of the README surface area, which creates confusion about what you should actually use. There's no streaming or incremental update path — you re-run inference from scratch each time, which is fine for batch weekly forecasting but painful for anything that needs low-latency updates as new data arrives.