Summary
- Google Research has released TimesFM-3, a foundation model that takes in multiple time series at once and forecasts the future.
- With 330 million parameters, pretrained on more than a trillion time-series data points, the model ranked first across all three benchmarks — Gift-Eval, FEV-Bench, and Time.
- The model is available now on GitHub and Hugging Face, with BigQuery AI.FORECAST integration expected to follow within a few weeks.
Why multivariate forecasting was needed
Google Research has released TimesFM-3, a multivariate time-series forecasting foundation model, on GitHub and on Hugging Face's TimesFM-3 model page. Built to take in several data streams at once and produce future values in a single pass, the model targets the kind of real-world forecasting problems seen in retail, finance, and healthcare.
To put it plainly: until now, AI time-series forecasting models predicted the future by looking at just one data stream, like sales or visitor counts. But in an actual store, accurate forecasts require also feeding in information that's already known ahead of time — things like a promotion schedule or the sales volume of related products. TimesFM-3 is built to take in several of these inputs together and calculate how they influence one another.
The trouble is that real-world problems like sales or visitor forecasting are almost always shaped by several factors at once — related product sales, promotion calendars, weather forecasts — and univariate models simply had no way to account for those signals.
The gap, illustrated with ice cream sales
Google Research used ice cream sales forecasting as an example. A conventional univariate model, looking only at past sales records, just extends the day-of-week pattern into the future — it has no way of noticing that a promotion is scheduled for a specific date. TimesFM-3 takes the promotion schedule in as a "past-future covariate," a signal that's already known in advance, learns the relationship between promotions and sales spikes from historical data, and applies that relationship to future promotion dates. The result, according to Google Research, was a forecast of roughly a 20% sales increase on each promotion day.
330 million parameters, one forward pass
Like its predecessors, the model uses a decoder-only transformer architecture, but it groups 32 time steps into a single patch and runs those patches through a two-dimensional attention structure that moves across both the time axis and the axis linking different time series to one another.
For past-future covariates, the model applies a "look-ahead" strategy, designed so each token bundles together the current patch and future patches. The forecasting approach has also changed. Earlier versions generated patches one at a time in sequence, which built up latency and compounding error. TimesFM-3, building on research applying Contiguous Patch Masking, masks the entire future window and fills in every predicted value simultaneously in a single forward pass. At each time step, it also outputs nine quantiles spanning the 10th to 90th percentile, giving a view into the uncertainty range around each forecast.
First place across three benchmarks
Google Research evaluated TimesFM-3 on three benchmarks: Gift-Eval, FEV-Bench, and Time. Compared against multivariate-capable models like Chronos-2 and the Toto 2.0 family, as well as its own predecessor TimesFM-2.5, TimesFM-3 came out on top among pretrained foundation models across all three benchmarks — on both point-forecast and probabilistic-forecast metrics.
| Evaluation mode | Benchmarks | Result |
|---|---|---|
| Univariate mode (no covariates) | Gift-Eval, FEV-Bench, Time | Matches or beats other foundation models |
| Multivariate mode (using covariates) | Gift-Eval, FEV-Bench, Time | First place on both point and probabilistic metrics |
Even in univariate mode — evaluating each time series independently, with no covariate information — TimesFM-3 matched or outperformed competing models, according to Google Research. Switching to multivariate mode pushed its ranking up another notch.
How to try it
Where to start — TimesFM-3 checkpoints are available from the GitHub repository and Hugging Face's TimesFM-3 model page.
Step-by-step usage
- Download the TimesFM-3 PyTorch checkpoint from Hugging Face or GitHub.
- Prepare the target time series you want to forecast (say, sales), along with historical covariates like related product sales and future-known covariates like promotion schedules.
- Feed this data into the model in one shot, and it returns the full forecast window along with nine quantiles in a single forward pass, with no iteration required.
Who it's for — Loading the model directly through code is a path suited to teams with development chops. If you want to try time-series forecasting without ML expertise, you can already use BigQuery's AI.FORECAST command with the previous model, TimesFM-2.5, for univariate forecasting.
What you can try — A retail chain could feed in promotion schedules and inventory data together to get an early read on sales spikes during discount periods. A manufacturer could use equipment run schedules and raw-material delivery timing as covariates to forecast production volume swings. In observability, you could feed in multiple server metrics at once and examine how an anomaly in one metric ripples into the others.
Editor's take
The original pitch was a single pretrained model — much like an image or language model — that could forecast time series across multiple domains with no additional tuning. Two years later, Google has taken that model from univariate to multivariate, moving it much closer to the messy problems teams actually face. That trajectory is now converging with Salesforce's Gift-Eval benchmark, where TimesFM-3 competes head-to-head against other multivariate models like Chronos-2 and Toto 2.0.
Put a time-series model of this size into real practice, and the conclusion tends to be the same: univariate models handle recurring patterns like seasonality or day-of-week cycles just fine, but accuracy falls off a cliff the moment something breaks the pattern — a promotion, a holiday, any reason "this time is different." TimesFM-3's decision to build in covariate inputs reads as a direct response to exactly that weak spot.
What's worth watching for Korean retail and manufacturing firms is that the model is open source, so it can be validated against proprietary data without any contract. That said, BigQuery integration hasn't yet been extended to TimesFM-3, so teams hoping to run it with a single command from the cloud console would do better to experiment with TimesFM-2.5 for now while getting their data pipelines ready.
Once BigQuery's AI.FORECAST command starts supporting TimesFM-3 in the coming weeks, the barrier will drop another notch — letting data analysis teams that don't touch code directly start testing multivariate forecasting for themselves.
Comments