Chronos2

class sdm.models.Chronos2(pretrained: bool = True, device: device | str | None = None)

Bases: ICLModel

The time series forecasting foundation model from the “Chronos-2: From Univariate to Universal Forecasting” paper.

Chronos2 is a time-series foundation model for multivariate forecasting. It extends Chronos with native support for jointly forecasting multiple coevolving target series, incorporating historical covariates, and using dynamic covariates that are known across both the past and future forecast horizon.

Architecturally, it combines patch-based time-series tokenization with alternating temporal attention and group attention, allowing forecasts to use both within-series history and information shared across targets and covariates. Each patch encodes scaled values, time indices, and observation masks. It returns 21 quantile forecasts: the 1st and 99th percentiles and every fifth percentile from the 5th to the 95th, for each target series and query time step. Forecasts of up to 1,024 time steps are decoded in a single forward pass; longer horizons use autoregressive rollout.

Within the ICLModel protocol, Chronos2 treats rows as ordered time steps and uses up to 8,192 time steps as context. In a default forward pass, x_context contains historical covariates, y_context contains one or more past target series, and x_query contains future-known covariates (which must also be present in x_context). By default, examples in a batch are forecasted independently; cross_learning=True allows group attention to share information across examples in the batch.

Parameters:

• pretrained (bool) – Whether to load the pretrained checkpoint.

• device (torch.device | str | None) – The device.

Additional Forward Parameters:

cross_learning (bool) – Whether examples in a batch will share information through variate attention. (default: False)

Capabilities

Supported Input Semantic Types	numerical
Supported Target Semantic Types	numerical
Supported Prediction Tasks	regression
Multi-Target Support	✅
Context-only Column Support	✅
Related Table Support	❌

Default Recipe

Recipe(
  features=Sequential(
    Standardize(),
    Asinh(),
  ),
  target=Sequential(
    Standardize(),
    Asinh(),
  ),
  output=AverageEstimators(),
)