How to use multivariate time-series forecasting without fine-tuning: the capabilities of Times FM3

Times FM3 is a pretrained model for time-series forecasting. It processes several related metrics jointly and uses known future factors without separate fine-tuning for each task. This mode is called zero-shot forecasting: the model receives a history of observations, a description of the structure of the input series, and factor values that are already known over the forecast horizon.

This approach is useful where one metric depends on another. A store can forecast foot traffic, revenue, and demand for complementary products simultaneously. An industrial enterprise can combine raw-material consumption, equipment utilization, and output in one set. For financial planning, several prices, trading volumes, and external indicators are suitable if their dynamics are observed synchronously.

Which tasks are suitable for zero-shot forecasting

A zero-shot scenario is suitable for an initial forecast assessment, quickly launching a pilot, and situations where there is not enough accumulated labeled data to train a separate model. Times FM3 receives historical values and generates a continuation of the series using patterns learned during pretraining. The user does not need to first select an architecture, train weights on every new dataset, or create a separate model for each store.

Data preparation is still necessary. If the system receives inconsistent time series rather than a coherent set of values, the model does not know what each column means. The analyst determines the observation frequency, shared calendar, forecast horizon, and roles of the series in advance. One series serves as the forecast target, while others are used as related targets or external factors.

• Sales. You can forecast sales for several product groups, inventory levels, and orders if these metrics share the same time step.
• Foot traffic. The model receives store traffic, retail-location utilization, and related conversion metrics.
• Financial series. The approach can be applied to several prices, turnovers, rates, and other metrics if their dynamics are observed synchronously.
• Industrial demand. The model is used to jointly forecast output, material consumption, line utilization, and technical parameters.

How to prepare several related time series

Multivariate forecasting begins with a shared time scale. Each row should correspond to one observation moment, and each column or channel should have a consistent meaning. Missing values, duplicate timestamps, a mixture of hourly and daily data, and undocumented changes in measurement units distort the relationships between metrics.

Before launching a forecast, check four properties of the dataset:

1. all series use the same time zone and observation frequency;
2. historical values contain no data from the future relative to the forecast point;
3. missing values and outliers are marked or handled according to a predefined rule;
4. for each series, it is determined whether it should be forecast or used only as a factor.

Related metrics do not have to use the same scale. Foot traffic may be measured in thousands of visits, while complementary-product sales may be measured in units. Scaling and normalization should match the interface of the specific Times FM3 implementation. Manual normalization of some channels must not be mixed with the original values of other channels without checking the inverse transformation.

It is best to preserve the original dataset before applying any transformations. After obtaining a forecast, this makes it possible to restore the original measurement units, compare the results with actual values, and identify the stage at which an error occurred. A separate preparation log records deleted entries, missing-value imputation rules, the time zone, the frequency, and the date of the last observation.

History and known future factors

Input data is divided into historical values and factors known over the future horizon. A historical series contains observations up to the forecast point. A future factor contains values that have already been determined or are available from an external plan. This distinction is important: future sales are unknown, while a planned discount, holiday, or advertising campaign may be known in advance.

Weather requires a separate check. Observed temperatures for past days belong to the history, while a weather forecast for future days is an external forecast and carries its own error. If the model is given actual future temperatures that became known only after the period ended, data leakage occurs. The result will look better than it would in real-world operation.

Discounts, advertising, holidays, and schedule changes are encoded as time series with a clear period of effect. For a promotion, it is important to provide not only the start date but also the duration, discount intensity, or advertising budget if these parameters are genuinely available in advance. If a future factor is unknown, it cannot be considered a known-future covariate. In that case, use only the history, a separate forecast of the factor, or a scenario set of options.

What parameters and pretraining mean

The materials available for this overview confirm a Times FM3 size of approximately 330 million parameters. They do not provide primary confirmation that the model was trained on more than one trillion time points, so this corpus size cannot be presented as an established characteristic of Times FM3. The scale of pretraining alone does not guarantee accuracy for a particular enterprise, market, or product category.

A large corpus helps the model encounter different trends, seasonal patterns, noise, and relationships between channels. Transferability is limited by the data distribution. A series with rare accidents, abrupt legislative changes, or a new type of demand may differ from typical pretraining examples. The practical rule is as follows: use zero-shot forecasting as a way to quickly obtain a baseline forecast, but not as a substitute for evaluation on your own data.

How to evaluate a forecast on a time-based holdout period

A multivariate forecast is evaluated on a period the model has not seen. Randomly shuffling rows is unsuitable for time series: it destroys the order of events and allows future information to leak. Use several temporal cutoffs. At each cutoff, take the history up to the forecast point, provide the permissible future factors, and compare the result with the actual values.

1. Choose several forecast points in the past.
2. Fix the same horizon for Times FM3 and baseline methods.
3. Provide only the future factors that would have been known on the selected date.
4. Compare the forecasts with a naive method, seasonal repetition, and a suitable statistical model.
5. Calculate the error separately for each channel, horizon, and group of entities.
6. Check uncertainty intervals, not just the average point error.

The average error across all channels can conceal a failure on one important metric. In retail, analyze stores with different levels of foot traffic separately; in industry, distinguish normal operating modes from transitional modes; and for financial data, compare calm and volatile periods. The more costly an error in a particular channel, the less useful a single aggregate metric becomes.

What should you do if the zero-shot forecast is worse than a simple seasonal baseline? Do not immediately increase the number of factors. First check the calendar, missing values, measurement units, cutoff-point shift, and the correctness of future features. If there are no technical errors, compare several horizons and segments. The model may be useful for some series and unsuitable for others.

Why a claim of benchmark superiority cannot be considered an independent comparison

The claim that Times FM3 outperforms other models on key benchmarks comes from Google. The source materials contain no independent confirmation of this result, so it is correct to call it a developer claim rather than an established universal fact. A benchmark result depends on the composition of the datasets, rules for excluding series, run parameters, baseline methods, and which data was used during pretraining.

For model selection in a company, your own time-based evaluation is more important. It should replicate the future usage scenario: the same horizon, frequency, constraints on factor availability, and data-update order. A public ranking helps create a shortlist of candidates, but it does not replace backtesting on data from the specific business.

Recommended implementation sequence

• Describe the target series, related channels, and external factors.
• Bring the data to a shared time scale and check the quality of the observations.
• Separate past values and future factors according to their availability at the forecast point.
• Run a zero-shot forecast on several historical cutoffs.
• Compare Times FM3 with simple and statistical baseline methods.
• Check errors by channel, horizon, segment, and operating mode.
• Only then decide whether to move to fine-tuning, an ensemble, or separate models.

Frequently asked questions about Times FM3

Does Times FM3 need to be fine-tuned for each store?

No. Zero-shot mode allows new time series to be provided without separate fine-tuning for each store. Evaluation on a holdout period is still necessary, since the absence of fine-tuning does not guarantee accuracy.

Can holidays and promotions be taken into account?

Yes, if their future values are known when the forecast is generated and are provided as separate time-based factors. Actual future sales or weather observations that became known later must not be provided.

How does a multivariate forecast differ from a set of independent forecasts?

Multivariate mode allows the model to account for relationships between several channels. Independent forecasts treat each series separately and do not use joint dynamics unless they have been specifically added as features.

Can Times FM3 be considered the best model?

No, this conclusion is insufficient for a production decision. Google claims strong benchmark results, but the source materials contain no independent confirmation of superiority. The model should be compared with baseline methods on your own historical cutoffs.