chronos2_forecaster

Chronos-2 foundation-model forecaster.

Chronos2Forecaster adapts the generic InferenceBackend to the OpenSTEF Forecaster contract. It owns the Chronos-2 specific pre- and post-processing while the backend stays model-agnostic:

• Preprocessing builds the context, attention_mask, group_ids, future_covariates and future_covariates_mask tensors. Every non-target feature column is treated as a known covariate: its history feeds an extra context row and its horizon values feed future_covariates. Chronos-2 shares attention within a group, so the target series and its covariates share one group_id. Chronos-2 normalises each series internally, so raw values are fed unscaled.

• Postprocessing picks each series’ target row out of the batched output, slices the model’s frozen horizon to the requested length and resamples the model-native quantile grid onto the requested quantiles.

The model is zero-shot: there is nothing to train, so fit() is a no-op and is_fitted is always True once a backend is attached.

Classes

Chronos2Forecaster(**data)	Zero-shot probabilistic forecaster backed by a Chronos-2 checkpoint.
Chronos2HyperParams(**data)	Hyperparameters for Chronos2Forecaster.