prediction_timedelta: is time dimension the time of initialisation or the time of forecast?

[oliverangelil]

I've selected a 3-day prediction_timedelta for GraphCast, PanguWeather, & NeuralGCM for Z500 geopotential, using the available forecast datasets on WB2

I've plotted 100 timesteps including ERA5 for 1 grid-cell:

It's clear that I've mis-aligned my timestamps, and that the time dimension that comes from prediction_timedelta is the time of initialisation and not the time of forecast

I.e. I cannot just compare the forecasts to the truth out of the box. Rather I need to roll-back ERA5 by 72 hours or roll-forward the AIWP models by 72 hours in order to make a comparison.

I just want others to be aware of this because I feel it might be a source of error to when comparing "truth" to forecasts!

I assume this roll-back is done in the weatherbench2.evaluation functions?

[oliverangelil]

yup - it's in the docs :)
https://weatherbench2.readthedocs.io/en/latest/init-vs-valid-time.html#init-time-convention