Premier League — Validation

Because accuracy and RPS measure different things, on different scales — they're not in competition.

• Accuracy only asks: was our single top pick right? It's blind to the probabilities, and a model almost never makes draw its top pick — so a draw-heavy round drags it down even when the forecast was good.
• RPS grades the whole probability spread (e.g. 55 / 25 / 20) against a uniform “coin-flip” (33 / 33 / 33). The coin-flip never commits, so it's never punished hard — which makes it a surprisingly tough baseline, and means a few % of skill is a genuine edge.

Our committed forecast cuts both ways (lower RPS = better):

When a favourite wins we crush the baseline; on a draw, the hedging coin-flip actually beats us. Matchday 1 had 9 draws in 24 games, so the wins and losses partly cancel — the net edge is the “skill” figure above. A few % over uniform is roughly what bookmakers manage, so it's a real, healthy edge.