TimeH2O
A family of time series foundation models: send one series or thousands and get point and quantile forecasts back, with no training run and no model per series.
- TimeH2O is a family of time series foundation models: send one series or thousands, with or without covariates, and get point and quantile forecasts back with no training run.
- TimeH2O-v1 has 108M parameters and is built for accuracy; TimeH2O-v1 Tiny has 5.83M parameters and is built for speed and a small footprint.
- Every TimeH2O forecast returns a point estimate and nine quantile levels, from 0.1 to 0.9.
- Both TimeH2O models are open weights on Hugging Face under the H2O Model Non-Commercial License; production use needs a license from H2O.
- TimeH2O does not produce per-forecast reason codes, so regulated forecasts go to H2O Driverless AI, our AutoML.
Forecasts for every series, without a model for each.
Classical forecasting builds, tunes and monitors one model per series. TimeH2O is one model that reads the history you send and forecasts what comes next, so the long tail of series finally gets a forecast.
- Zero-shot forecastsForecasts a series it has never seen from its trend, seasonality and calendar effects, so a new store or SKU gets a forecast on day one.
- Thousands of series, one callOne endpoint covers thousands of series, so there is no model to build, tune or monitor for each store, SKU or sensor you track.
- Point and quantile outputEvery forecast returns a point estimate and nine quantile levels, from 0.1 to 0.9, so planning can use service levels instead of one number.
- Covariates when you have themAdd outside drivers to a forecast through the covariate adapter when you have them. Forecasts made without covariates are preserved exactly.
- Gap filling for real dataTrained to fill missing stretches such as outages and maintenance windows, so imperfect real-world history still gives a usable forecast.
- Two sizes, one familyTimeH2O-v1 is built for accuracy at 108M parameters. TimeH2O-v1 Tiny, at 5.83M parameters, is built for speed and a small footprint.
Series in, forecasts out.
- 01Supply your seriesSend one series or thousands, with timestamps and, if you have them, covariates. No model is built per series.
- 02Load the weightsInstall the h2ofm package, then call TimeH2O.from_pretrained with the Hugging Face model ID, in your own environment.
- 03Read the forecastGet a point forecast and nine quantiles for the horizon you ask for, ready for planning or for a downstream check.
- 04Nothing to retrainNo model per series means no per-series retraining. New data is simply a new call to the same pretrained model.
The fast path for time series, beside tabH2O and Driverless AI.
TimeH2O is the time series foundation model in the H2O Predictive AI Platform, beside tabH2O, the tabular foundation model. Many series, little history and no regulator in the loop: TimeH2O. A regulator who needs reasons: H2O Driverless AI, our AutoML.
What ships today
Questions people ask before they call us.
TimeH2O is the time series foundation model and the first choice for forecasting. tabH2O can also forecast, and fits when the forecast is really a table-shaped regression with many row-level drivers and you supply the lag features yourself.
Try it on your own series.
Download the weights and test them on your history, or bring a forecasting problem to H2O and we will run it with you.