H2O.ai has released TimeH2O-v1, a time-series foundation model for forecasting, and says it ranks first on the fev-bench leaderboard by win rate.

According to H2O.ai, TimeH2O-v1 placed first among 32 models in the benchmark's default view. It posted a 91.9% win rate and a skill score of 50.8 across 100 forecasting tasks, ahead of Timestral-209M, t0-gamma, TimesFM-3 and Chronos-2.

The release puts H2O.ai into direct competition with larger technology groups and specialist model developers in a part of the artificial intelligence market focused on operational prediction rather than text generation. Time-series forecasting is used in demand planning, inventory management, cash forecasting, energy load management and patient volume planning.

Unlike forecasting systems trained separately for each dataset, TimeH2O-v1 works in zero-shot mode, H2O.ai said. This means it reads the history of a time series and generates a forecast without training on that specific dataset at the point of use.

H2O.ai said TimeH2O-v1 was not trained on any fev-bench dataset, and that no training or fine-tuning takes place during forecasting, leaving the model's weights unchanged. The benchmark measures probabilistic accuracy on real-world time series across different domains and frequencies against a seasonal-naive baseline.

On the leaderboard view cited by H2O.ai, TimeH2O-v1 beat rival models in 91.9% of head-to-head comparisons. Timestral-209M followed at 89.6%, while t0-gamma recorded 88.1%.

H2O.ai said the model's skill score of 50.8 means forecast error was 50.8% lower than the seasonal-naive baseline when aggregated across all 100 tasks.

ai, and can forecast both univariate series and series that include covariates such as price, promotions or weather.

Source link