Google Research releases new Google Search. TimesFM-3The model is a time series foundation with over 330 millions parameters that can forecast multiple series at once. Each TimesFM checkpoint up to 2.5 was univariate, a single series with its own historical data, and nothing more. TimesFM-3 has been pretrained for multivariate prediction on over 1 trillion timepoints. It accepts multiple target, past covariates and past-future comorvariates without task-specific tweaking. The model is ranked first among all pretrained models in GIFT Evaluation, fev Bench, and TIME Leaderboard on point and probability metrics.
Is it deployable?
PartiallyThe TimesFM repository code is Apache-2.0But the TimesFM 3.0 weights ship under timesfm-non-commercial-license-v1.0. These are only for non-commercial and non-production purposes. It is available for benchmarking today. You can’t ship behind a forecast API.
What’s changed
Each TimesFM release up to 2.5 had a single-variable. This model predicted one series based on its history. The majority of real forecasting issues aren’t shaped this way. Google’s example Ice cream sales are affected by foot traffic, the weather, promotional campaigns, holidays, etc.
TimesFM-3 was pretrained to multivariate forecasting. It has 330,000,000 parameters and is pretrained with more than 1 trillion times points from real and synthetic series. There are three input types that work with zero-shot and no fine-tuning for specific tasks:
- Multi-targets Forecasting jointly with quantile and point outputs.
- Past covariatesOnly known from the past, for example, foot traffic.
- Covariates of the past and futureThe future value of the object is known. For example, in a calendar for promotions.
Two types of architectural attention: first, patches.
Backbone is a transformer that only decodes. The points that are adjacent to each other are then grouped together into 32-step patches, and normalized by series. This prevents the use of scales with wildly different values. All tokens, including the target and past-covariates, are generated from one patch. Tokens for past and future covariates are generated using the lookahead technique: The current patch is combined with patches in the future, allowing the model to see events scheduled before they happen.
Tokens enter into a 2D Grid and then pass through two alternate attention mechanisms.
- Attention temporale causale The horizontal direction of the run is a straight line. The leakage is, therefore, strictly causal.
- Attention Full The vertical direction is used. Tokens are used to read every series of a certain timestep, and learn correlations between the different series.
The one forward pass is better than many
In earlier TimesFM versions, patches were decoded individually. This adds to latency and compute costs, as well as compounding errors. TimesFM-3 utilizes Contiguous Patch MaskingTiRex introduced the strategy of masking training times. The entire horizon is masked with placeholder tokens. There, targets and previous covariates will be masked. Future signals are still able to reach the model, as past covariates remain visible. Every masked patch of horizon is simultaneously filled with the alternating layers. Every horizon step, 9 quantiles are applied to each target, representing the 10th through 90th per centile.
Benchmarks
Google evaluates GIFT-Eval, fev-benchThe. TIME leaderboardTimesFM-2.5, Chronos-2 and the Toto 2.0 Family. TimesFM-3, among the pretrained models of foundations, is ranked highest on average for all three metrics, including point values and probabilistic measures. The package release notes Ranking #1 in TIME over 50 datasets with 98 evaluation Tasks.
Interactive explainer
‘;
}
h+=’

