chronos-forecasting

Why we included this project

Forecasting usually means retraining a model per dataset, but Chronos skips that: you load one of its pretrained checkpoints and get probabilistic multi-step forecasts from historical values in a few lines of Python. The family spans 9M to 710M parameters, so there's a variant for fast, memory-constrained inference as well as heavier models, and the newer Chronos-2 line handles multivariate series and exogenous covariates zero-shot. Checkpoints ship through Hugging Face and the repo includes AWS deployment guides, making it practical for both quick experiments and production endpoints. If you want to sanity-check it before committing, the linked fev tool benchmarks Chronos against other forecasters on the same datasets.