Forty8 runs a single, transparent forecasting pipeline from raw match data to projected bracket, recomputed from open sources after every matchday. The inputs, parameters and choices are documented on this page.
Every probability on this site comes from one model, not an ensemble of three. Here is each stage in the order it runs.
Team strengths start with Elo ratings whose constants are fitted on out-of-sample predictive likelihood, not the commonly-reposted folk defaults. Alongside Elo, Pi-ratings provide a second signal, and a rating-uncertainty term discounts teams whose historical data is sparse (qualifying-path newcomers, recently restructured programs).
Rating differences feed a time-decayed attack/defence Poisson model with the Dixon–Coles low-score correction (correlating 0-0 and 1-0 outcomes) and shrinkage toward confederation means. The output is a full scoreline distribution, not just a win probability, because group-stage tiebreakers require goal differences and goals scored, not just results.
When bookmaker outright odds are available they are de-vigged with Shin's method, averaged on the logit scale across books, and inverted through tournament simulation into per-team implied strengths, folding in anything the market knows that raw match history misses (current form, injury news, squad depth). When squad data is loaded, a weighted squad-quality signal supplements the ratings before simulation begins.
The calibrated model runs the full 48-team 2026 format 100,000 times: group tiebreakers (head-to-head sub-table → goal difference → goals scored → fair play → seeded randomness), Annex-C ranking of third-placed teams to fill the Round-of-32 bracket, knock-out draws including extra time and penalty shootouts (modelled as near-equal coin-flips; evidence for a skill effect is weak). Championship odds are reported with 90% simulation intervals.
The displayed bracket is the modal projected path: the most-frequent outcome in each match slot across the 100,000 runs. It is not a single simulation and it is not guaranteed to be self-consistent (a team might appear as modal winner in two slots simultaneously in a theoretical sense). Treat it as the single most-likely story, not as a prediction that every game goes a specific way.
Every input is public-domain or CC0. Every file's SHA-256 hash goes into a manifest, so any published probability can be traced back to the exact bytes that generated it.
Community-maintained results dataset covering international matches from 1872 to present, released under CC0.
Match schedule from the openfootball project. Public domain, community-maintained.
Bookmaker outright odds de-vigged with Shin's method. Used as an optional signal; the model runs without them when odds are unavailable.
Projected line-ups and squad quality scores loaded at team-sheet announcement. Invisible to the model between updates; acknowledged in limitations.
Calibration is verified with reliability tables, not simply claimed. Backtests are strictly walk-forward (no data leakage): the model is trained only on matches that preceded the test window. Expected three-way accuracy sits near 55%, which is roughly where the best published models and leading bookmakers operate. Anyone claiming substantially more without rigorous out-of-sample evidence is cherry-picking.
What the model cannot do, and where it can mislead.
Every forecast set is committed to a public, append-only history before kickoff and never edited afterwards.
Predictions are timestamped and committed before matches begin. The full archive (every forecast we have ever published) is available in the repository. If a forecast looks good in retrospect, you can verify it was made in advance; if it looks bad, we cannot hide it. Judge the model on the archive, not on highlighted successes.