- Group
- H
- Date
- Sunday 21 June 2026
- Kick-off
- 16:00 UTC
- Venue
- Atlanta Stadium, Atlanta
§ 01
The scoreline grid
The ensemble's full distribution over exact scorelines — Spain goals down the side, Saudi Arabia goals across the top. This is the same 11×11 grid the tournament simulator samples.
Fig. V7 Ensemble · Group H
Spain v Saudi Arabia — scoreline probabilities
Cells show P(exact scoreline); the right column and bottom row are the marginal totals P(Spain scores k) and P(Saudi Arabia scores k). Grid runs 0–7 goals per side; the 8–10-goal tail holds 0.76% of the mass and is omitted from the cells (not from the totals).
The grid makes Spain favourites at 85.7%, with a 11.2% draw. The single most-likely scoreline is 2–0 (15.8%), but no exact score clears 16% — the distribution is broad, as it should be.
§ 02
Result, goals & totals
The grid marginalised into the outcome split, expected goals each side, and the goal-total markets — every figure read straight off the same distribution.
Win · draw · loss
Rounded values sum to exactly 100%.
Expected goals (λ)
Poisson means feeding the grid; combined expected goals 3.20.
§ 03
Venue context
The conditions at Atlanta Stadium, Atlanta — the same environment layer the dossiers' climate exposure draws on.
Source · Open-Meteo & venue records. Travel and time-zone exposure are per-team — see each side's dossier.
§ 04
Why the model leans this way
The inputs behind the grid — each side's strength signals laid side by side. The forecast is the ensemble's, not a hand-weighting of these; this panel is the descriptive read.
Spain carry the Elo edge (586 points). On the decoupling axis, Spain is the side whose squad is valued higher relative to its record.
How a single-match forecast is built
The pairing is scored by the ensemble — Dixon-Coles bivariate-Poisson, the Bayesian hierarchical model and the global LightGBM-Poisson, log-pooled — yielding the 11×11 scoreline grid above. Win/draw/loss, expected goals (λ), over-2.5 and both-teams-to-score are all marginals of that one grid, so they are mutually consistent by construction. The strength inputs shown here feed the models; the forecast is their pooled output, not a manual weighting of these rows. The model matches the market out-of-sample (RPS 0.1891 vs 0.1905); it does not significantly beat it at n = 3. The ensemble, in full →