2026 World Cup · group stage · v2
Group stage, v2.
The same model as the original board, with one change: draws are calibrated. Same winner picks, better-weighted draw probabilities.
What v2 is, and why. The original board (v1) builds win/draw/loss from a plain Skellam goal model, which is known to under-predict draws. v2 swaps in the zero-inflated Skellam (Karlis & Ntzoufras), with one parameter fit on historical data and frozen, lifting the draw probability to its empirical level (the group-stage draw rate moves about 19% to 22%, and the close games rise from ~25% to ~28%). It is a pre-committed fork run forward in parallel with v1, it never edits v1, and it is graded head-to-head as games resolve. The winner pick is unchanged; what changes is how much weight the draw carries, which is exactly what calibration and CLV reward. v2 launched after the opener, so it covers the remaining group games.
v1 vs v2, head-to-head
Log-loss and Brier on the games both versions pre-committed, as they resolve. Lower is better. This is the real test of whether the draw fix helps.