2026 World Cup · group stage · v3
Group stage, v3.
v2 with one more change: a format-aware draw lift. The 48-team best-thirds format rewards drawing, so even matches draw more than history alone implies. Winner picks unchanged; the draw carries more weight where it should.
What v3 is, and why. The 48-team format lets a team advance on three draws (cf. Slovenia at Euro 2024, P3 W0 D3), so even matches draw more than v2's historical calibration captures. v3 adds a closeness-conditional lift to v2's zero-inflated Skellam: ω = 0.04 + 0.10·(1 − |p1 − p2|), extra draw mass in even games, almost none in mismatches (a 93% favourite stays ~93%). The lift was validated out-of-sample: it improves W/D/L log-loss on the expanded-format Euros (the optimum, 1.026 → 1.017) while being net-harmful on the old 32-team World Cups, i.e. genuinely format-specific, not a blanket draw bump. Like v2 it is pre-committed forward in its own ledger and graded head-to-head against v2 as the remaining group games resolve. v3 launched mid-tournament, so it covers the games still to be played.
v2 vs v3, head-to-head
Log-loss and Brier on the games both versions pre-committed, as they resolve. Lower is better. The clean live test of whether the format draw lift helps.