The way I suspect you are doing it is
a. Generate the deterministic cross-validated hindcasts
b. Calculate the error variance of those hindcasts (Remi had asked me to output this variance a few years ago)
c. Use that error variance to calculate probabilistic hindcasts (the forecast error variance has to be calculated for each hindcast separately because the forecast error variance is a function of the hindcast error variance and of the values of the predictors, which obviously change from year to year)
So this process cheats a little bit in that you have to calculate all the deterministic hindcasts before you can calculate the probabilities.
In an email 2024-03-15 Simon wrote: