TimesFM

Google Research's foundation model for time-series forecasting, now a family across three generations. The variant this wiki has hands-on documentation for is google/timesfm-2.5-200m-pytorch: a 200-million-parameter transformer, installed with pip install timesfm[torch] and loaded through TimesFM_2p5_200M_torch.from_pretrained (timesfm-2-5-forecasting-tutorial).

Two hundred million parameters is small enough to be a footnote in a language-model context and is the whole model here, which is the first thing worth noticing about this class: a foundation model for numeric sequences doesn't need to carry the world's text.

What it does at inference

• Zero-shot. Point it at a series it has never seen and ask for a forecast. No fine-tuning step, no target-dataset training.
• Probabilistic by default. Every forecast returns ten quantiles (q10–q90 plus the mean) rather than a single line, with fix_quantile_crossing=True keeping the intervals monotonic. The intervals are what make the anomaly-detection use possible at all.
• Bounded context and horizon. Compiled with max_context (tested to 1,024 steps, required to be a multiple of 32) and max_horizon (direct to 256; longer via recursive chunking, at a cost in accumulated error).
• Covariates through XReg. Dynamic numerical, dynamic categorical and static categorical regressors can be fused before or after the forecast (xreg_mode), which is how a univariate model gets told about a promotion.
• Compile-time configuration. normalize_inputs, per_core_batch_size, use_continuous_quantile_head, force_flip_invariance, infer_is_positive — the accuracy/cost trade-offs are set when the model is compiled, not per call.

Why a zero-shot forecaster matters to this spoke

fine-tuning carries the corpus's "don't reach for fine-tuning first" rule. TimesFM is the limit case of it: the intended path is no adaptation at all, on a task where the classical answer was to fit a fresh model per series — an ARIMA or ETS or gradient-boosted model estimated on that store's own history. A pretrained forecaster inverts that: one model, many series, nothing fitted.

That is the same move language models made, arriving late in a field that had decades of per-series statistics, and it makes forecasting the clearest non-text place to watch whether the pretrain-once-apply-everywhere pattern actually generalizes. Answered in part 2026-08-31: it generalizes against the cheap classical fits and not against a supervised model trained on the target data. On gift-eval's 97 configurations TimesFM places mid-table — average rank 8.55, behind three deep models and two other foundation families — while the benchmark's winner, PatchTST, is trained per dataset. See time-series-forecasting for the evaluation design, and note that the first source here measured the model on data it invented.

3.0 changes the model class (2026-08-31)

timesfm-3 is 330M parameters and natively multivariate: alternating causal temporal attention along time and full variate attention across series, so cross-series structure is pretrained rather than fused in at call time the way XReg does it. It decodes the whole horizon in one forward pass (Contiguous Patch Masking) instead of patch by patch, and takes past covariates, past-and-future covariates and several targets together, still with nine quantiles.

Two things about that release belong here rather than only on the source page.

The licence narrowed. The 2.5 card is apache-2.0; the 3.0 card is license: other / timesfm-non-commercial-license-v1.0 (both checked 2026-09-01). The weights are still downloadable and commercial use is not granted, which is a different artifact from the one described above even though the name did not change.

The accuracy claim is unnumbered. Google says 3.0 tops gift-eval, FEV-Bench and a TIME-leaderboard among pre-trained foundation models, in charts, with no scores. That comparison class excludes PatchTST, the per-dataset supervised model that actually won GIFT-Eval, so the mid-table rank recorded above stands unrefuted rather than superseded.

Snapshot caveat: version numbers, checkpoint names and configuration flags move. The inference behaviour described in the sections above the 3.0 heading is the 2.5 / 200M PyTorch build as documented on 2026-08-03, and has not been re-checked against 3.0. The model comes from ../llm-providers-wiki's google, which that spoke owns as an entity.

Related

gift-eval · timesfm-3 · timesfm-2-5-forecasting-tutorial · time-series-forecasting · fine-tuning · machine-learning · synthesis