Google Research releases TimesFM-3, a zero-shot foundation model for multivariate forecasting
Forecasting is adopting the pretrain-once approach that reshaped language models, and future-known covariates are the feature that makes it useful.
Google Research published TimesFM-3 on 31 August 2026, describing it as a zero-shot foundation model for multivariate time series forecasting. The post, by research scientists Ayush Jain and Rajat Sen, says the model is designed to forecast several related series at once, in a single forward pass, without training on the user's own data.
Earlier versions, including TimesFM-2.5 released in September 2025, were restricted to univariate forecasting, meaning they could only use the history of one series. TimesFM-3 adds support for past covariates, which are features known only historically, and for dynamic covariates known into the future, such as scheduled promotions or weather forecasts. The model decodes the full horizon non-autoregressively and returns probabilistic forecasts across nine quantiles, from the 10th to the 90th percentile.
The model has 330 million parameters and was trained on more than one trillion time points drawn from real-world and synthetic datasets. It groups 32 contiguous time steps into patches and uses a decoder-only transformer with alternating causal temporal attention and full attention across variates. Google reports the best average rank on GIFT-Eval for both point and probabilistic forecasting, strong results on multivariate datasets in FEV-Bench, and top placement on the TIME leaderboard ahead of Chronos-2 and Toto 2.0. In univariate mode it is said to match or beat competing models.
Weights are available on Hugging Face and code on GitHub, with a BigQuery integration expected in the coming weeks. The post illustrates the multivariate case with ice cream sales, where related products, foot traffic, weather, promotions and holidays feed the forecast, allowing the model to anticipate a sales bump of around 20 percent on promotion days.
Why it matters
Forecasting has long been a per-dataset discipline built on tabular models; a pretrained transformer shipped as weights changes who can do it and how quickly. Covariate support is what makes such models usable for operational problems, where part of the future is already scheduled. Benchmarks such as GIFT-Eval are becoming the leaderboards for this field in the way language benchmarks did for LLMs. Forecasting is joining the foundation-model stack rather than sitting beside it.
Forecasting has long been a per-dataset discipline built on tabular models; a pretrained transformer shipped as weights changes who can do it and how quickly.
Input series and covariates
The target history plus related series, past-only features and future-known signals such as promotions.
Patching into 32-step tokens
Groups of 32 consecutive values become one embedding, shrinking the sequence the model must attend over.
Temporal and cross-variate attention
Alternating layers look along time within a series and across the related series at each step.
Whole horizon decoded in one pass
Every future step is produced together rather than one at a time, avoiding compounding error.
Nine quantiles per step
The output is a band from the 10th to the 90th percentile, not only a single point.
What you can learn from this
- Zero-shot forecasting means the model was never trained on your data. Like a language model predicting the next word, a time series foundation model learns general patterns of seasonality, trend and noise from an enormous corpus, then applies them to a series it has never seen. This works because many real-world series share structure, such as weekly cycles or sudden level shifts, even when the domains differ. The trade-off is that a model fitted to your own history can still win when your data is unusual. Benchmark rankings are averages across many datasets, not a guarantee for yours.
- Patching turns a long series into a short sequence of tokens. Instead of feeding one time step per token, the model groups 32 consecutive values into a single patch and embeds it as one vector. This shrinks the sequence length the attention mechanism must process, which reduces compute sharply because attention cost grows with the square of sequence length, and it lets the model see longer histories. It is the same idea as image patches in vision transformers.
- Covariates are how a forecast learns what is scheduled. Past covariates describe things that happened alongside the target, such as foot traffic, while dynamic covariates describe things known for the future, such as a promotion calendar. A model that only sees the target's history cannot know that next Tuesday has a sale; one that accepts future-known inputs can. That is why covariate support matters more for operations than a small accuracy gain on a benchmark. It is also where domain knowledge enters, because you decide which signals are genuinely known in advance.
- Non-autoregressive decoding predicts the whole horizon at once. Autoregressive forecasting predicts one step, appends it, and predicts again, so early errors compound and long horizons take many passes. Generating every step of the horizon in one forward pass avoids that drift and is faster at inference time. The cost is that the model must learn the shape of the whole horizon jointly, which needs more and more varied training data.
- Quantile forecasts express uncertainty, not just a point. Returning the 10th through 90th percentiles gives a band that says how wide the plausible outcomes are, which is what inventory, staffing and capacity decisions actually need. A model is trained for this with a pinball loss that penalises each quantile differently depending on whether the truth lands above or below it. Evaluating probabilistic accuracy separately from point accuracy is why benchmarks report both. A wide band on particular days is informative in itself, because it flags where the model has little confidence.
We teach this
How to use this in practice
- Run TimesFM-3 against a series you already have. Install the package from the GitHub repository, load the weights from Hugging Face, and feed it a daily series such as your website traffic or a personal metric with at least a few hundred points. Hold out the last 30 days and compare the model's median forecast with what actually happened. Keep the script, because you will rerun it in the next two steps. Done looks like: a chart with the actuals, the median and the 10 to 90 percent band, plus the mean absolute error written down.
- Beat it with a baseline, or fail trying. Compute a seasonal naive forecast, which repeats last week, and a simple exponential smoothing forecast for the same holdout, then compare errors. A foundation model is only worth its inference cost if it beats the trivial baseline on your data. If it loses, note by how much and on which days, because that pattern usually reveals structure the pretrained model has not seen. Done looks like: a three-row table of errors with a one-sentence verdict.
- Add one dynamic covariate and measure the change. Pick something you know in advance, such as a holiday flag or a planned campaign, add it as a future-known input, and rerun the forecast. Note whether the band narrows or the median shifts on the flagged days, and whether the error on those days improves. Then supply the same covariate for the past window only, to see how much of the gain came from future knowledge versus historical context. Done looks like: a before-and-after pair of charts and a note on which days changed.
- Diagram the model's data flow. Draw the raw series being cut into 32-step patches, the patches becoming embeddings, the alternating temporal and cross-variate attention layers, and the output head emitting nine quantiles per horizon step. Mark which parts are shared with a language model and which are specific to time series. Done looks like: a one-page sketch you can explain to someone else in two minutes without notes.
Sources
- TimesFM-3: A zero-shot foundation model for multivariate forecasting — Google Research
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report