Scoreline picks that maximise contest points — not headlines. [ market 60% · form 25% · xG 15% · ranking 15% · last-match v2 ]
FIFA № 3 (~1871 pts) — vs — FIFA № 4 (~1828 pts)
The consolation nobody wants and everybody scores in: sixteen of the last twenty third-place games went over 2.5, and the market prices this one accordingly (Over at 1.40, the heaviest of the tournament). France bring the sharper motivation — Mbappé sits level with Messi on 8 for the Golden Boot, and Messi plays tomorrow — while England may rotate after four brutal knockout rounds. Both semifinal losers conceded late; both defences will be remixed. 2–1 France is the play in an open game — 47/25/28, over 65%. FT 4–6 — open doesn’t begin to describe it: the highest-scoring game in World Cup history. England 3–0 up inside 45 (Rice 3′, Konsa 18′, Saka 37′), Saka completing a hat-trick (87′), Mbappé scoring twice (48′, 66′) to pass Messi as the tournament’s all-time scorer on 22 and all but claim the Boot on 10 — and stoppage goals from Dembélé and Bellingham closing a 6–4. Our Over read was right and our winner wasn’t: 0 pts.
Grid covers 94.0% of outcomes · over 2.5 goals: 65.3% · 90-minute scoreline
| Score | P(exact) | E[points] |
|---|---|---|
| 2–1◂ PLAY | 9.2% | 7.42 |
| 2–0 | 6.3% | 6.56 |
| 3–1 | 5.8% | 6.42 |
| 1–0 | 5.3% | 6.28 |
| 3–2 | 4.2% | 5.95 |
| 3–0 | 4.0% | 5.87 |
Start from the de-vigged closing market (60%), then blend twelve-month form & style (25%), an in-tournament xG model (15%) rating each team on the expected goals it created & conceded across its four games so far, and the FIFA ranking's Elo expectancy (15%). The xG signal is shrunk to the league mean so it nudges rather than dictates.
Each team's xG is nudged by how its most recent game beat/missed expectation; the attack term (finishing) is halved because it regresses. The read uses each side's final group game — now a week old for today's slate, so treat it as a light thumb on the scale.
A Poisson model with the Dixon-Coles low-score correction ρ gives a probability for every scoreline. Each is ranked by 30·P(exact) + 10·P(1X2) — the contest's scoring, which stack (an exact hit is worth 40). Knockouts are scored on the 90-minute result, so the model still picks within 90 minutes.
The bronze final — graded: 0, in the highest-scoring World Cup game ever played. France 4–6 England: a ten-goal madhouse — Saka’s hat-trick, Mbappé’s brace to become the World Cup’s all-time leading scorer (22) and Golden Boot leader (10), Bellingham at 90+8′. Our 65%-over read was emphatically right; the 2–1 France pick was not. Tournament 930/4080. One game left. Scoring: 30 exact · 10 winner (max 40).
Across 24 graded knockout picks: the model said 25% of games end level after ninety — 25% did (6 of 24). It projected 2.62 goals a game — 2.65 arrived. It claimed ~12.9% on its exact picks and landed 8.3% (2 of 24) — inside one standard error at this sample size. Nothing in the data says the probabilities are wrong.
Knockout haul: 210 of 960 (21.9% of max) against the model's own pre-match expectation of ~21%. Results hit 15/24 (win-picks 14/20 = 70%); two exacts ★ (Canada 0–1, Colombia 1–0). The actual scoreline landed in our published top-three 55% of the time against a modelled 33% — the grid's shape has been better than advertised; the top line has been unlucky within it.
Variance dominates at n=24. Two behavioural rules survive the audit: coin-flip draw picks stay (the draw share was priced exactly right), and no exact-chasing — the second-ranked scoreline has cashed five times to the top line's one, a gap comfortably inside luck; abandoning the E[pts] argmax would trade points for stories. Eight picks remain (four quarter-finals, two semis, the bronze final and the final): expected haul ≈ +70, projected finish ≈ 950/4120.