A pretrained forecasting model produces a plausible curve without task-specific training. That can shorten a first experiment. It does not define the target, validate data availability, or establish that the forecast improves a planner's decision.

“Zero-shot” refers to applying a pretrained model to a target task without fitting it on that task in the stated evaluation. It does not mean the model has never been trained, has never seen similar data, or is immune to benchmark overlap.

Read current releases without flattening them

Google's 31 August 2026 TimesFM-3 announcement describes native multivariate forecasting and covariate support. Chronos-2, reported in October 2025, also addresses univariate, multivariate, and covariate-informed forecasting.

These are distinct model families and implementations. Provider benchmark claims concern their datasets, versions, and evaluation choices. Do not combine reported scores from different protocols into a universal ranking. Pin the checkpoint and library before running a local comparison.

Write the temporal contract first

For a synthetic weekly demand task, specify item or region, forecast issue time, horizon, frequency, missing-period meaning, and revision policy. A known future calendar feature is different from a future price or promotion that has not been committed.

Covariates must be available at the prediction cutoff or supplied under an explicit scenario. Using the actual future values in a backtest can turn the task into an easier one than deployment. Retain data vintages where records are revised after publication.

Compare with credible local baselines

Use seasonal-naive, an established local model, and the foundation-model candidate on the same rolling forecast origins. Fit preprocessing only within each training window and preserve a clean evaluation period. The time-series cross-validation guidance explains why evaluation must respect ordering.

Report errors by horizon and operational segment, not only an aggregate mean. MAPE is problematic near zero; select losses suited to the target and decision. A global improvement can conceal important route-level errors that cancel in the aggregate.

Check uncertainty and disruption

For published intervals or quantiles, evaluate coverage, width, and relevant loss on held-out origins. A very wide interval can achieve coverage without guiding the decision. A model trained across many series is not guaranteed to anticipate a structural break in this one.

Include missing recent observations, a late covariate, a new series, and an explicit fallback. Distinguish the forecast's statistical uncertainty from an unavailable input or unsupported operating condition.

Read the weights licence separately

The current TimesFM repository distinguishes its Apache-2.0 source code and older weights from the separately licensed TimesFM-3 downloaded weights. The retrieved TimesFM-3 licence restricts those weights to non-commercial, non-production use. Authorized Google Cloud service use follows a different contractual route described by the project.

This is a release-specific licence observation checked on 4 October 2026, not legal advice or an assumption about later versions. Check the exact artifact and hosted-service terms before deployment. An open code repository is not a blanket production licence for every checkpoint.

Make the first deployment claim narrow

Record task coverage, model/version, source cutoff, baselines, error and interval results, latency, cost, and licence route. Review whether the planner can act on the forecast and recover when it fails. A zero-shot result can be a strong baseline or a viable component, but it must still earn the operating promise.

Foundation models change how the model is obtained. They do not remove the temporal, evaluation, and product responsibilities that make forecasting useful.

Sources and further reading