Overview
Most forecasting work still means building and tuning a new model for every series, product, and region. TimeH2O, a time-series foundation model, forecasts data it has never seen, with no training run on your data. TimeH2O-v1 ranks first of 32 models on fev-bench by win rate (as of Oct 23) and was never trained on any fev-bench dataset.
Sri Ambati and Kaggle Grandmasters Pascal Pfeiffer and Mark Landry explain how TimeH2O works, where it beats a tuned gradient-boosted model and where it doesn’t, and what to try first. Then they run it live, from raw data to forecast.
Who should attend
Data scientists, ML engineers, and analytics leads who own forecasting: demand, capacity, energy load, or financial signals.
- What a time-series foundation model is, and how it differs from LLM-based forecasting
- How TimeH2O forecasts a new series with no training on it first, with or without covariates
- Where it beats a tuned gradient-boosted model, and where a tuned model still wins
- How to run your first forecast on your own data


