I Tested Google's TimesFM-3: Forecasting Without Training on My Data

I have spent enough time around forecasting systems to be suspicious of any model that claims it can predict a business it has never seen. Traditional forecasting usually starts with a long process: collect historical data, engineer features, train several models, validate them, and keep retraining as the business changes.

TimesFM-3 asks me to consider a different starting point. Google trained one general model on more than a trillion real and synthetic time points, then designed it to forecast a new dataset without task-specific training. I can give it sales, traffic, inventory, weather, promotions, or other time series and ask what is likely to happen next.

That does not make it an oracle. It does make it a very interesting baseline.

What TimesFM-3 is

TimesFM-3 is a foundation model for time-series forecasting. The comparison with a language model is useful, with one important difference: instead of learning patterns in sentences and predicting the next token, it learns patterns across charts and predicts future values.

The original TimesFM, released in 2024, was trained on 100 billion time points and focused on univariate forecasting. It could examine the history of one signal, such as daily revenue, and extend that pattern into the future.

Version 3 changes the scope. It has 330 million parameters, was pretrained on a corpus of more than one trillion time points, and can work natively with several related signals. Google describes three inputs that matter:

• Multiple targets, such as sales for several related products
• Past covariates, such as historical foot traffic
• Past-future covariates, such as scheduled promotions, holidays, or a weather forecast that is already known for the prediction period That last category is the part I find most useful. A sales forecast that ignores next week's promotion calendar may be statistically tidy and operationally useless.

How I tested the idea

I used a simple ice cream sales scenario with 156 days of data. The model received 128 days of history and produced a 28-day forecast. The dataset included ice cream sales, cone packs, syrups, promotions, temperature, and foot traffic.

I first ran the forecast without the extra business signals. The result tracked the general pattern surprisingly well, but promotional spikes were hard to anticipate because the model had no information about future campaigns.

When I added the promotion calendar, the forecast began responding to the planned events. Weather and foot traffic changed the shape again. This is the practical difference between extending one line and modeling a business situation. The target series tells the model what happened. Covariates help explain why it happened and what is already scheduled to change.

In this small exercise, TimesFM-3 reached a MAPE of 4.73. The seasonal naive baseline scored 6.72, and the other traditional machine-learning baseline scored 8.89. I would not treat one synthetic retail scenario as a universal benchmark, but it was enough to show me that zero-shot forecasting can be useful before a data science team builds a custom pipeline.

The latency in the demonstration was about 203 milliseconds for the 28-day run. That number depends on hardware and implementation, so I see it as an observation from this test, not a product guarantee.

What happens under the hood

TimesFM-3 does not process every individual time point in isolation. It groups contiguous values into patches of 32 time steps. Patching reduces the number of tokens and helps the model focus on broader patterns instead of reacting to every noisy day.

The transformer then treats the data as a two-dimensional grid. One attention step moves across time within each series and can only look backward, which prevents future target values from leaking into the forecast. The next step moves across variables at the same time position, allowing sales to relate to promotions, weather, or another product.

Known future covariates remain visible. Unknown future values, including the target, are masked. Through a method Google calls Contiguous Patch Masking, the model predicts the full horizon in a single forward pass rather than generating one future patch at a time.

It also produces nine quantiles, from the 10th to the 90th percentile. For planning, that is more useful than a single confident-looking line because inventory and staffing decisions depend on uncertainty.

What I like

The biggest gain is speed to a credible baseline. A team can test whether a forecasting problem has useful signal before spending weeks on feature engineering and model selection. A smaller business without a forecasting team can get a reasonable first model instead of relying entirely on spreadsheets.

The multivariate interface also matches how operators think. Demand planners already look at historical sales, promotions, weather, product relationships, and store traffic together. TimesFM-3 gives the model the same context without requiring a separate training run for every dataset.

Google reports top average performance among pretrained foundation models on GIFT-Eval, FEV-Bench, and TIME, for both point and probabilistic forecasting. That is stronger evidence than my small test, although public benchmark leadership still does not prove that the model will win on a particular company's data.

Where I would be careful

Zero-shot does not mean data-free. The model still needs clean history, sensible timestamps, aligned variables, and future covariates that are genuinely known. A bad promotion calendar or a broken traffic series can still produce a bad forecast.

The model also does not know every business constraint. It cannot infer supply limits, product cannibalization, a pricing rule, a warehouse closure, or an unusual market event unless those effects appear in the inputs. For a large retailer, even a small accuracy improvement can be worth millions, so a custom model and human planning process may still justify the cost.

There is also a licensing limit. The TimesFM source code is under Apache 2.0, but the TimesFM-3 pretrained weights are currently distributed under a separate non-commercial license. They can be used for research and evaluation, but not for commercial production with the default weights. Google has announced a BigQuery integration as the intended managed path, with availability described as coming after the model release.

My verdict

TimesFM-3 does not kill traditional forecasting. It changes what I expect from the first week of a forecasting project.

I would use it immediately as a zero-shot baseline, a feasibility check, and a way to give smaller teams a forecast they did not have before. I would compare it with seasonal baselines and a well-tuned model, then inspect where each one fails. If the business stakes are high, I would keep the custom work, domain constraints, and human review.

The real improvement is not that one model replaces every forecasting stack. It is that a useful multivariate forecast can now exist before the stack is built.

Sources

• Google Research: TimesFM-3, a zero-shot foundation model for multivariate forecasting
• Google Research TimesFM repository