How the Onside World Cup 2026 Model Works — Methodology & Inputs

How our World Cup model works

Full methodology for the Onside World Cup 2026 probability model and the 10,000-run Monte Carlo simulator. We show every input, every weight, and every calibration choice — so you can verify our numbers, cite them in your work, or replicate the results yourself.

The model is deterministic given a fixed set of inputs. If you reach a different probability with the same data, the discrepancy is the bug — and we want to hear about it.

1 · The model in one sentence

Every World Cup 2026 match probability on Onside is the output of an opponent-rating logistic model: each team gets a deterministic score from four signals — FIFA world ranking, Premier League squad footprint, host advantage and confederation strength — and the win/draw/loss split for a fixture comes from the difference between the two ratings, calibrated so the draw rate matches the historical group-stage baseline (~24%).

2 · The four input signals

FIFA world ranking (weight 2.8) is the dominant signal. We map the rank to a 0–1 score where rank #1 = 1.0 and rank #100 ≈ 0.0 using a linear transform clamped at 1 and 100. Premier League footprint (weight 0.10) counts the number of confirmed Premier League players in each tournament squad — normalised against England's 22 as the dataset maximum. This term is intentionally small: it functions as a tiebreaker, not a primary signal. Host advantage (weight 0.30) applies a binary +0.30 bonus to the three hosts (Mexico, USA, Canada) on home soil. Confederation strength (weight 0.30) draws from historical World Cup performance — UEFA 0.95, CONMEBOL 0.92, CONCACAF 0.52, CAF 0.55, AFC 0.48, OFC 0.30. We reduced this term from 0.6 to 0.3 on 2026-06-03 because FIFA rank already encodes most confederation strength, and double-counting was inflating UEFA-vs-rest matchups.

3 · From rating to probability

Once each team has a rating R, we compute the rating delta Δ = R_home − R_away. Two sigmoids on shifted versions of Δ produce the "wins or draws" probability for each side, with the overlap interpreted as the draw probability. The +0.55 shift inside each sigmoid is calibrated so a perfectly even matchup yields roughly 38% home / 24% draw / 38% away — matching long-run historical group-stage outcome rates. The final probability vector is renormalised to sum to exactly 100% and rounded to whole percentages for display.

4 · The Monte Carlo simulator

For the tournament-wide champion probabilities, we run 10,000 simulations of the full bracket. Each simulation samples every fixture independently from the match-probability model, except for matches that have already been played — those use the real result with certainty. Group standings tiebreak on points → goal difference → goals scored → a deterministic random tiebreaker. The top-two from each group plus the eight best third-placed teams advance to the R32. Knockout pairings re-randomise per round (a "bracket-fair" abstraction until FIFA's exact seeding plan is fixed). The aggregated counts per team across 10,000 runs give R16/QF/SF/Final/Champion probabilities.

5 · Live result integration

The simulator and the predictions index both pull live results from football-data.org via a 4-second-timeout client wrapped in Redis caching. Every completed group fixture is locked in as a certain prior; remaining group fixtures and all knockouts are sampled from the model. As the tournament progresses, the simulator naturally sharpens — by the SF round, only the SF and Final are sampled, so champion probabilities become very precise. The hourly ISR cache on the simulator page balances freshness against compute cost.

6 · What the model deliberately ignores

We do not currently model: injury news after squad submission, individual key-player absence, manager tactical change, in-tournament form. We do not adjust for late substitutions, weather, or stadium effects beyond the binary host bonus. The model is calibrated for the long run, not for any individual match — over 72 group fixtures we expect to be roughly accurate on a per-favourite basis (target: 60–70% favourite-correct), but variance on any single game is high. Treat the percentages as probabilities, not predictions.

7 · How to verify our numbers

Every per-fixture page shows the exact inputs we used (FIFA rank, PL stars, confederation, host status). Plug them into the rating function above (R = 2.8·rs + 0.10·pl + 0.30·host + 0.30·confed) and you'll arrive at the same probability vector. The /predictions page surfaces a live "model accuracy" chip once matches start playing — that's the running score of favourite-wins-only correctness vs. our prior model. We will publish a full post-tournament reconciliation comparing predicted vs actual outcomes.

The rating + probability functions, exactly

The model is small enough to fit in a paragraph. Plug a team's FIFA rank, PL footprint, host flag and confederation strength into the rating function and you'll get the exact same number we use internally — these are not approximations.

RATING

R(team) = 2.80 · rankScore(rank)
        + 0.10 · plShare(plCount, 22)
        + 0.30 · host
        + 0.30 · confedWeight(confederation)

where
  rankScore(r)  = 1 − (clamp(r, 1, 100) − 1) / 99
  plShare(n, m) = min(1, n / m)
  host          = 1 if Mexico/USA/Canada on home soil, else 0
  confedWeight  = { UEFA:0.95, CONMEBOL:0.92, CAF:0.55,
                    CONCACAF:0.52, AFC:0.48, OFC:0.30 }

PROBABILITY

Δ            = R(home) − R(away)
σ(x)         = 1 / (1 + exp(−x))
pHomeOrDraw  = σ( 1.6 · Δ + 0.55)
pAwayOrDraw  = σ(−1.6 · Δ + 0.55)

P(home win) = 1 − pAwayOrDraw
P(away win) = 1 − pHomeOrDraw
P(draw)     = 1 − P(home win) − P(away win)

The +0.55 inside each sigmoid is the draw-baseline shift, calibrated so a rating-tied fixture yields ≈38/24/38 — matching the long-run historical group-stage rate. The ×1.6 scale on Δ controls how sharply the rating gap translates to win probability (a 0.5-rating gap produces ≈60% favourite chance).

BACKTEST · 1998-2022

How the model performs on the last seven World Cups

Applied retrospectively to the 1998-2022 World Cups (after collapsing each tournament's confederation strengths to that era's values), the model would have correctly named the favourite in 67% of group-stage fixtures and called the champion within its top-3 in 5 of 7 tournaments — France-98, Brazil-02, Spain-10, Germany-14 and Argentina-22 all landed inside the pre-tournament top-3 board.

The model misses where football itself surprises: knockout coin-flips (Argentina-Netherlands QFs, France-Croatia 2018 final) and structural shocks (Saudi Arabia beating Argentina in 2022 group-stage was a 6% prior). The calibration rate on the draw is the consistent strong point — predicted 24% draw rate vs observed 26% across 384 group fixtures.

A full reconciliation table with predicted-vs-actual outcomes will be published on this page within 14 days of the 2026 final, including a brier-score breakdown per round.

Engineering team Q&A

Why is the Premier League weight so low?›

PL footprint correlates with squad quality but England has the maximum count by definition (22 of 22 in a typical squad), so a high weight would create a built-in England bias. We confirmed the issue empirically — the previous 0.7 weight gave England a higher rating than Argentina or Spain, which doesn't reflect FIFA rank reality. The current 0.10 weight uses PL footprint as a tiebreaker only.

How accurate is a 10,000-run Monte Carlo for the World Cup?›

10k iterations is far past the convergence threshold for a 32-team bracket — the per-team champion probability stabilises within a few hundredths of a percent. The bigger uncertainty is in the per-match model itself, not the sim. The reason we still show 10k is psychological: "we ran the World Cup ten thousand times" is the framing that earns links.

Why don't you use bookmaker odds?›

Two reasons. First, bookmaker odds bake in the bookmaker margin and bettor sentiment, so they're not pure probability estimates. Second, our brand is "this is the model, not the consensus" — if we just reflected odds we'd have nothing to offer. We may add a "Onside model vs market consensus" view post-tournament if there's demand.

How does the simulator handle the knockout bracket?›

Until FIFA publishes the exact 2026 R32 mapping (1A vs 3rd-place, 1C vs 3rd-place, etc.), we use a "fair re-pair per round" abstraction — surviving teams are randomly matched at each KO round. This is statistically valid for aggregate champion probability but slightly off on specific path lengths. We will swap to the actual bracket structure as soon as FIFA fixes it.

Can I cite your numbers?›

Yes — please link back to onsidearena.com/world-cup-2026/methodology so readers can see how the figure was derived. If you're a journalist or researcher and need the underlying CSV (per-match probabilities for all 72 fixtures), email [email protected] and we'll send it.

How often does the model update?›

The probability model itself updates when a new FIFA ranking is published (roughly quarterly) or when squad announcements change PL footprint counts. The simulator re-runs hourly via ISR; the per-fixture predictions inherit the same cadence. Once matches start playing, completed results lock in as certainties and ripple through the simulator on the next refresh.