Time series foundation models are making their way into production workflows, following the same playbook that transformed language and vision AI. Instead of building custom models for each forecasting task, practitioners can now load a pretrained model and generate predictions with minimal setup. AWS recently released Chronos-2 in October 2025, representing the latest advancement in this approach.

The workflow shift is dramatic. Traditional forecasting required data preparation, model selection among ARIMA or LSTM variants, then extensive training and tuning. Chronos-2 compresses this into a single inference call. Users provide historical data and forecast horizon, receiving not just point predictions but predictive quantiles for uncertainty quantification. This addresses cold-start scenarios where limited data previously killed projects.

Chronos-2 uses a "group ID" mechanism to handle different forecasting patterns. Assign unique IDs for univariate forecasting, shared IDs for multivariate targets, and group covariates with targets for informed predictions. The model learns shape patterns like cycles and trends rather than specific numeric values, enabling cross-domain generalization.

Chronos-2 currently leads in zero-shot accuracy across multiple benchmarks and outperforms classical statistical methods without task-specific tuning. However, domains too divergent from pretraining data still require careful validation, and inference costs rise compared to specialized models. The technology shifts forecasting accessibility from ML experts to domain specialists with basic Python skills.