Dear all,
As many of you know, the final version of ERA5 is released with a three-month latency. To keep things running smoothly, I had set up an automated suite to upload the latest historical quintile bounds and potential ERA5 training data (i.e. daily values of weekly precipitation accumulations).
Unfortunately, thanks to the Microsoft team ( Discrepancy between computed and downloaded quintiles ), I recently discovered an issue in that workflow. The automated computation of historical quintiles was drawing on ERA5 data that omitted values from January to July 2024. As a result, the historical quintiles for January to July 2025 were incorrect, and the same issue propagated into the January to July 2026 quintiles. The mistake went unnoticed because I was using nearest when selecting the corresponding historical date, which allowed the script to silently substitute the wrong timestamps. (See the referenced GitHub page for the fix made to the AI Weather Quest package).
I’ve now fully recomputed the historical quintiles and updated the workflow to ensure the correct date is always selected by removing the nearest option.
I apologise for any inconvenience this error may have caused. Since all data for September to November were complete, this issue does not affect evaluation scores or leaderboard rankings.
Thanks in particular to the Microsoft team for spotting the error!
Thanks all,
Josh