The Football Simulator methodology
A transparent description of how HORIZON 6.2 and HORIZON GLOBAL 3.0 convert team strength, completed matches, sustainable form, confirmed evidence and uncertainty into fixture and season probabilities.
Match forecasting
Each club has attack and defence strength on a log-goal scale. Venue, sustainable-form effects, verified transfer effects, schedule context, and exponentially weighted result residuals alter the expected goals for a fixture. The Premier League engine uses an 8 by 8 Poisson score matrix with a Dixon-Coles correction. The wider competition engine uses a 13 by 13 score matrix with each competition's observed draw rate applied before probabilities are normalized to 100%.
Global competition calibration
The Championship, League One, League Two, National League, La Liga, Bundesliga, Serie A, Ligue 1 and UEFA competition pages use HORIZON GLOBAL 3.0. A reproducible pipeline ingests five completed seasons of free Football-Data.co.uk results across 19 domestic leagues and divisions. Venue-aware Elo, attack, defence, recent form, scoring environment and empirical draw rate are calculated from 36,079 completed matches. Each club input is shrunk toward a structured prior according to sample confidence, with prior-only teams labelled in the interface. Completed 2026/27 results are reconciled against the exact published fixture pair before they can be locked through the priority live-source hierarchy.
Sustainable form
The final 25 matches of 2025/26 are compared with each incumbent club's full-season xG and xGA rates. Goals minus xG and xGA minus goals are treated as finishing and shot-prevention variance rather than copied forward. Thirty per cent of the underlying rate change is carried into 2026/27, adjusted by a sustainability multiplier and capped at plus or minus 0.055 on each attack or defence axis. Promoted-club Championship form is not blended directly into Premier League xG without a league-strength conversion.
Parameter uncertainty
The Football Simulator creates 24 correlated posterior states for team strength and league scoring tempo. Promoted and lightly observed clubs receive wider uncertainty. One shared posterior state is used across an entire simulated season path, preserving correlation between a club's future fixtures. The interface reports central estimates and 10th to 90th percentile ranges.
Season simulation
The default production setting is 100,000 seeded Monte Carlo seasons. Completed results are fixed and every future fixture is resolved as a scoreline inside every path. Domestic tables apply points, goals and competition-specific goal-difference or head-to-head rules, with the club strength index used only as the final deterministic fallback. UEFA paths simulate the league phase, knockout play-offs, two-legged rounds and final. The model reports Monte Carlo error and does not label a run stable when its maximum probability error is too wide.
Transfer treatment
Confirmed incoming and outgoing movements are stored in one ledger. First-team effects depend on modeled ability, likely minutes, role need, tactical fit, age, transfer type, source reliability, data completeness, and attack or defence share. Academy and uncertain first-team movements can be ledger-only with zero model weight. Rumours do not enter team strength.
Player counterfactuals
Tracked players have separate attack, defence, expected-minutes and tactical-fit assumptions. A permanent departure removes the full baseline contribution. An arrival adds a fit- and minutes-adjusted contribution to the buyer and, where the selling club is in the league, removes the source contribution. Injury and suspension scenarios are weighted by league matches missed. A minutes scenario measures the reduction from the player's existing expected share. The baseline and counterfactual use the same seed so the displayed change is driven by the assumption rather than unrelated simulation noise.
Confirmed player availability
Official club or league confirmation is required before an injury or suspension changes the baseline. When a return date is not supplied, The Football Simulator publishes a conservative expected-matches assumption and a wider planning range. The expected absence is spread across the league season, while the full source wording remains visible. The assumption is revised as soon as a club publishes clearer timing.
Betting value
Bookmaker odds never enter the core forecast. Compatible outcome groups are converted to fair market probabilities with power-method margin removal. A pick must have a positive lower-bound edge after model uncertainty, sufficient data quality, current pricing, and acceptable liquidity. High lineup-sensitive markets are restricted before confirmed teams. Stakes use capped quarter-Kelly research units. No bookmaker price means no value prediction.
Top goalscorer and each-way value
The Golden Boot model runs 100,000 player outcomes alongside the published season forecast. Club scoring volume comes from the same 380-fixture paths, while player goal share, likely minutes, role and availability uncertainty remain independent inputs. Correlated club scoring and player-specific role variation are applied on every path. Win and top-four probabilities use dead-heat payout equity. Win expected value uses the best displayed outright price. Each-way expected value uses a separately verified eligible quote, splits the stake evenly between win and place legs, and applies the displayed place fraction. Ante-post place terms expire at kickoff unless a fresh in-season bookmaker capture confirms they remain available.
Market data and freshness
A dated opening-week best-price comparison is published separately from executable odds. It can be used to compare no-vig market probability with the model, but it cannot produce an actionable grade. The live adapter only unlocks execution research when a licensed provider supplies a current quote. Price age, bookmaker breadth, margin removal, lineup status and market dispersion all reduce confidence or fail the gate.
Transfer discovery and confirmation
Official league and club sources are the confirmation layer. Reputable news desks cross-check contract and timing details. Specialist journalists can create a fast discovery signal, but that signal remains visibly separate with zero model weight until an official confirmation is available. This prevents speed from being mistaken for certainty.
Evidence and limitations
Pre-kickoff forecasts are frozen with their model version and input hash. After verified results arrive, the system calculates multiclass Brier score, ranked probability score, log loss, outcome accuracy, and calibration error. Betting picks retain their quoted price and later closing price. A pre-season model has no 2026/27 out-of-sample accuracy sample, so the product displays that evidence as collecting rather than claiming a proven edge.