Check implementation of rescale_weights #760
Description
We recently added weights = time_diff.shape[0] * weights / weights.sum() to the team model if rescale_weights is True (the default for AIrsenal as we showed it to improve performance). The intent is to make it so that the sum of the weights equals the number of matches of data we have for a team. This should be computed separately for each team, but the expression above is on a flat list of results across all teams, so I don't think it's doing what was intended. Edit: I think it's fine now as per the comment below, it's normalised so the sum of the weights equals the total number of matches, so teams with fewer matches do have less total weight in their data.
I was thinking about this because the team parameters we're fitting for Sunderland are currently very unstable - they're varying massively between gameweeks - which maybe indicates that we're weighting the couple of matches of data we have for them much more heavily than intended (so the prior from the FIFA ratings is having next to no effect).
If we change it the best epsilon value might change, so we should re-run a sweep.
We also have this in the player model, but there it's in a loop across each player, so I think it's correct there.