Simulation · Google Research
TimesFM 2.5
A decoder-only foundation model for zero-shot numerical time-series forecasting without task-specific training.
What TimesFM 2.5 is
TimesFM 2.5 is a 200-million-parameter time-series foundation model developed by Google Research. It generates point forecasts and probabilistic intervals directly from raw numerical sequences, operating in a zero-shot manner without requiring model retraining, fine-tuning, or per-series calibration.
Traditional forecasting methods often require researchers to specify seasonal patterns, configure series frequencies, or fit bespoke statistical models for each observation channel. TimesFM 2.5 simplifies this workflow by taking unannotated historical data and projecting future trajectories, removing the need for manual frequency indicators while supporting covariate inputs through external regression.
The model produces both point forecasts and continuous quantile bands from the 10th to the 90th percentiles. Because it evaluates purely numerical series, unrecorded external events—such as abrupt policy shifts or assay protocol revisions—remain unseen by the architecture. Its median forecast reflects trajectory continuation from the observed past, while its prediction intervals capture model uncertainty.
How it works
TimesFM 2.5 uses a decoder-only transformer architecture with fused query, key, and value (QKV) attention projections designed to optimise inference speed.
The model was pretrained on diverse sequence datasets, comprising the GiftEvalPretrain collection, Wikimedia Pageviews (with data up to November 2023), Google Trends queries (with an end-of-year 2022 cutoff), as well as synthetic and augmented series.
Key facts
- Architecture
- Decoder-only transformer with fused QKV matrices
- Base parameter count
- 200 million
- Quantile head parameter count
- 30 million
- Pretraining datasets
- GiftEvalPretrain, Wikimedia Pageviews, Google Trends, and synthetic data
- Required GPU memory
- 8 GB
What you can use it for
- Zero-shot sequence forecasting
- Project future steps for newly encountered time series without fitting a specialised statistical model or training on local data.
- Prediction interval estimation
- Derive continuous quantile forecasts from the 10th to the 90th percentile to assess uncertainty in future trajectories.
- Covariate-assisted projection
- Integrate auxiliary dynamic features alongside primary target series using covariate regression (XReg).
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
TimesFM 2.5
Strengths and limits
Good at
- Performs zero-shot forecasts on unseen series without per-task tuning or parameter fitting.
- Eliminates the requirement to manually supply input frequency annotations.
Watch out for
- Cannot incorporate external domain factors such as experimental protocol changes or unrecorded interventions unless they are captured directly in the input values.
- Historical pretraining data excludes Google Trends past late 2022 and Wikimedia pageviews beyond November 2023.
- Requires 8 GB of GPU memory for standard execution.
Inputs and outputs
You provide
Observed series
At least 8 numbers, oldest first. A CSV column, a JSON array, a spreadsheet row or a space-separated list all work. Paste a table and the last column is read as the value, the rest as labels — the result says which. Only the most recent 1,024 points are used.
Steps aheadoptional
How far to forecast, in the same units as the input's spacing.
You get back
History, forecast and the 10th–90th percentile band on one axis.
Point forecast and the 10th, 50th and 90th percentiles per step.
Observations used, horizon and the range of the forecast.
Licence
Commercial use permitted · Apache-2.0
Apache-2.0 for both the code and the published weights. Google notes it is not an officially supported Google product.
Questions
Can I use TimesFM 2.5 commercially?
Commercial use permitted. Apache-2.0 for both the code and the published weights. Google notes it is not an officially supported Google product.
How do I run TimesFM 2.5 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 16GB, typically in 6.0s, and a run costs about 14 credits. Visitors without an account get 1,000 credits a month.
Does TimesFM 2.5 require frequency or seasonality metadata?
No. The model operates on raw sequences of numbers and does not require explicit frequency indicators or predefined seasonality parameters.
What forecast outputs does the model produce?
It provides point forecasts as well as continuous quantile distributions, including the 10th, 50th, and 90th percentiles for each step in the prediction horizon.
Can TimesFM 2.5 be fine-tuned on proprietary data?
Yes. Fine-tuning is supported through parameter-efficient methods such as LoRA via Hugging Face Transformers and PEFT.
Does TimesFM 2.5 support exogenous covariates?
Yes. Covariate support is available for TimesFM 2.5 through XReg.
How to cite it
A decoder-only foundation model for time-series forecasting, ICML 2024. arXiv:2310.10688.
Written from TimesFM 2.5's own documentation (model card, code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.