Description
TimeH2O-v1 is a sophisticated time series foundation model developed by H2O.ai, featuring 108 million parameters. It is specifically engineered for zero-shot probabilistic forecasting, meaning it can make predictions on unseen data without requiring prior fine-tuning on that specific dataset. A key capability of TimeH2O-v1 is its support for covariates, allowing it to incorporate additional relevant variables alongside the primary time series data to enhance prediction accuracy.
Read More
The model has demonstrated state-of-the-art performance across multiple benchmarks. On GIFT-Eval, it achieved a CRPS (Continuous Ranked Probability Score) of 0.441 and a MASE (Mean Absolute Scaled Error) of 0.654, outperforming models like Timestral-209M, t0-gamma, TimesFM-3, and Chronos-2. Similarly, on fev-bench, TimeH2O-v1 recorded an SQL skill score of 50.83 and a MASE skill score of 39.95, again surpassing its competitors.
TimeH2O-v1 is part of the broader TimeH2O family, which also includes TimeH2O-v1-tiny, a smaller 5.8M-parameter model optimized for maximum throughput rather than peak accuracy. The v1 model is best suited for scenarios where maximizing prediction accuracy is the primary goal. It can process up to 16,384 time steps of context and offers an unlimited prediction horizon, outputting 9 quantiles (from 0.1 to 0.9) for probabilistic forecasts. The model accepts target series with optional covariates as inputs.
Access to TimeH2O-v1 on Hugging Face requires users to agree to share their Hugging Face information and accept specific conditions, as it is licensed for non-commercial use. The model was trained on a diverse set of datasets, with careful exclusion of any data overlapping with fev-bench, GIFT-Eval, and TIME to ensure unbiased benchmark results. Training data sources include autogluon/chronos_datasets, Salesforce/GiftEvalPretrain, Salesforce/lotsa_data, hq-bench/quito-corpus, Dingdong-Inc/FreshRetailNet-LT, Azure Functions 2019 trace, Iowa Liquor Sales, and BLS Average Prices, each with their respective licenses.
- CategoryAI Models & LLMs
- DeploymentCloud
- PlatformsHugging Face
TimeH2O-v1 Highlights
108M parameters for robust forecasting
Zero-shot probabilistic forecasting capabilities
Support for covariates to enhance prediction accuracy
State-of-the-art benchmark results on GIFT-Eval and fev-bench
Context window of up to 16,384 time steps
Unlimited prediction horizon
Outputs 9 quantiles (0.1 to 0.9) for probabilistic forecasts
Optimized for maximum accuracy in time series predictions
Non-commercial license for usage
Exclusion of benchmark-overlapping data from training
Part of the TimeH2O model family
Publicly accessible on Hugging Face with conditions
Getting Started with TimeH2O-v1
Access Model: Navigate to the TimeH2O-v1 page on Hugging Face.
Accept Conditions: Review and accept the terms and conditions to access the model's files and content.
Log In/Sign Up: Log in to your Hugging Face account or sign up if you don't have one.
Download Files: Download the model files and associated resources.
Set Up Environment: Configure your development environment with necessary dependencies for using Hugging Face models.
Integrate Model: Integrate TimeH2O-v1 into your application or workflow for time series forecasting.
Provide Inputs: Supply target series data along with optional covariates to the model.
Generate Forecasts: Utilize the model to generate zero-shot probabilistic forecasts with 9 quantiles.
TimeH2O-v1's Use Cases
- Zero-Shot Forecasting
- Probabilistic Forecasting
- Covariate-Enhanced Prediction
- High-Accuracy Time Series Analysis
- Long-Horizon Forecasting
- Benchmark Performance Evaluation
- Research and Development







