How does the same historical forecast score under different deepscale versions? Each line shows skill on a fixed (country, season, init) across rebench runs — x-axis is the rebench date.