ECMWF AI Weather Quest – Benchmarking AI Models for Sub-Seasonal Forecasting

Hi everyone,

I’d like to share a project we’ve been working on at the European Centre for Medium-Range Weather Forecasts (ECMWF) — the AI Weather Quest. It’s an open, global initiative to benchmark AI/ML models for sub-seasonal forecasting, particularly targeting week 3 and 4 lead times (Days 19–25 and 26–32).

Forecasting at these timescales is notoriously difficult. This competition is designed to create a transparent, structured benchmark for AI-based sub-seasonal prediction, and is built entirely around open-source tools and data.

:brain: What it’s about:

Participants submit quintile probability forecasts at 1.5° resolution for:

  • 2m temperature
  • Precipitation
  • Mean sea level pressure

Forecasts are evaluated using the Ranked Probability Skill Score (RPSS) and visualised on a public ECMWF-hosted portal.

We support:

  • Any AI/ML methodology
  • Any programming language
  • Any dataset (with open historical datasets provided)

Teams of up to 10 people submit forecasts weekly across four 13-week competitive periods.

:hammer_and_wrench: Open-Source Tools

We provide a fully open Python package to support participation:

  • :white_check_mark: Forecast submission formatting
  • :white_check_mark: ERA5-based training data retrieval
  • :white_check_mark: RPSS-based evaluation tools

:microscope: Currently: Testing JJA Period (May–August 2025)

We’re now entering a non-competitive testing period, where participants can:

  • Submit weekly forecasts
  • Run evaluations locally using the same scoring system
  • Prepare for the competition phase starting in August

We’re hosting a Testing Period Launch Webinar on 7 May:

Thanks for reading — and hope to see some of you in the Quest!