Weather Forecasts from Previous Days to Compare Run-To-Run Performance
We began collecting past model run data in January 2024, with most weather models offering data from this point onward. Some models were added later in 2024 and 2025, providing data from a later start date. For GFS, data has been integrated from April 2021 onward, but only for 2m temperature.
Weather models constantly churn out updates, each predicting the future at different lead times. Think of Day 0 as latest forecast close to measurements, Day 1 as a glimpse 24 hours back, and Day 2 as a 48-hour rewind. Each day further back forecasts longer into the future and, typically, increases volatility. Data jumps become wilder past Day 6 or 7, highlighting the inherent challenge of long-term forecasting.
This data serves multiple purposes, including answering questions such as "what did yesterday's forecast predict for today?" or by comparing past forecasts with real-time observations, we can assess a forecast's accuracy and volatility. When combined with machine learning techniques, models can be trained specifically to enhance forecasts for the next 2 or 3 days.
The frequency of model updates varies, ranging from hourly to every six hours. For local models with shorter prediction horizons (2-5 days), we naturally have access to a shorter "time machine" of past predictions (2-5 days).
Weather Models Sources: The Previous Runs API uses the same models as available in the general weather forecast API. Please refer to the Forecast API documentation for a list of all weather models and weather variables.
Data Availability: Data is generally available from January 2024 onwards. Exceptions are GFS temperature on 2 metre, which is available from March 2021 and JMA GSM + MSM models which are available from 2018. More data from previous runs can be reconstructed on request (depending on data availability from official sources).