Choosing a time series model
Select a checkpoint using the task contract, evaluation evidence, license, and runtime constraints.
Specify the task before the family
Record target shape, cadence, history length, horizon, and which covariates are actually available at issue time. Decide whether you need a point estimate, quantiles, joint multivariate predictions, anomaly scores, or representations. Those tasks are not interchangeable, even within a foundation-model family.
Compare against useful baselines
Use seasonal-naive and relevant classical methods alongside TSFMs. Evaluate the forecast horizon and error metric that drive your decision. Segment results by frequency, domain, missingness, and difficult series rather than relying only on a pooled leaderboard score. Keep a held-out evaluation that was not used to select settings.
Validate the deployable checkpoint
Check revision, preprocessing, tensor shapes, uncertainty decoding, memory, and licensing. TimesFM 3.0 downloaded weights have different usage restrictions from 2.5; the source-code license alone is not sufficient. Choose an actual supported browser or provider contract after measuring model quality.