H2O.ai’s New Forecasting Model Tops Benchmark Leaderboard
KUALA LUMPUR, Oct 27 (Bernama) -- H2O.ai, the trusted Sovereign AI Platform, has released TimeH2O-v1, a time-series foundation model for forecasting that ranks first of 32 models on the fev-bench leaderboard by win rate.
The model posted a 91.9 per cent win rate and a skill score of 50.8 across the benchmark’s 100 forecasting tasks, ahead of Timestral-209M at 89.6 per cent and t0-gamma at 88.1 per cent. It also outranked TimesFM-3 and Chronos-2, according to the leaderboard’s default view as of Oct 23, 2026.
“Building your business with accurate forecasting and prediction is the superpower of AI,” said H2O.ai Founder and Chief Executive Officer Sri Ambati in a statement.
He said the model could be fine-tuned into a domain-specific foundation model for each customer’s needs.
The company said most organisations still build and maintain a separate model for each forecast, such as demand, inventory, cash, energy load or patient volume, and retrain it when the data shifts. A forecasting foundation model aims to remove that step by working on new data straight away.
fev-bench tests models on real time-series data from many domains without any prior training on that data and measures their accuracy against a basic seasonal forecasting baseline. A skill score of 50.8 means the model’s forecast error was about half that of the baseline.
H2O.ai said TimeH2O-v1 was never trained on any fev-bench dataset. At forecast time, the model reads a series’ history and predicts from it, without any further training.
TimeH2O-v1 has 108 million parameters and can forecast a single series on its own or alongside outside factors, such as price, promotions or weather. The company said it performs well on data from finance, healthcare, supply chain and energy, and that organisations can fine-tune it on their own data and infrastructure for greater accuracy.
The model is available on Hugging Face under a non-commercial licence that allows internal benchmarking and testing.
-- BERNAMA