World vs Model · FAQ
← BoardMethodology →Glossary →
Plain-English answers to the questions people actually ask about this project. For the maths, see the Methodology; for jargon, see the Glossary. Everything here is research & education — not financial advice, not a tipping service, and not gambling.
No. No real money is staked and no positions are ever held. Every figure on the site is a paper simulation of market structure — a way to keep score of a forecast, not a wager. There is nothing to win and nothing to lose.
No. Nothing here is advice, a solicitation, or a tip. The whole point is a public, falsifiable scorecard — testing whether transparent models can out-forecast a liquid market — not telling anyone what to back.
The site itself is just public research, so reading it is fine. But prediction-market platforms such as Polymarket are restricted or banned in several jurisdictions (for example, Singapore). The Polymarket links here are reference-only; know and follow your local laws.
No. This is an independent, solo research experiment. Market data comes from Polymarket's public APIs, but the project is not affiliated with, or endorsed by, Polymarket or FIFA.
Can a transparent, rules-based model beat the betting market? Two models take on the crowd across all ~240 Polymarket World Cup 2026 markets, and every disagreement is scored against the market as results come in.
The zero-knowledge model knows no football — it only re-shapes the market's own prices with a favorite–longshot correction. It is the honest baseline. The informed model is an independent Elo simulation of the bracket, not derived from the market, so it can genuinely disagree with the crowd. The board lets you switch between them.
A single source: Polymarket's public Gamma and CLOB APIs. Each round's prices are de-vigged (rescaled to sum to the real number of slots) so they are comparable, apples-to-apples, with a model.
Nobody knows yet — deciding that is the experiment, and they are built to be judged rather than assumed. There is an important subtlety, though. The zero-knowledge model only re-shapes the market's own prices, so it can never disagree with the market about who goes how far — its bracket is the market's bracket by construction. It earns its keep purely on calibration: are the probabilities the right size? The informed model is the only one that can genuinely disagree on the ordering, because it is built from Elo, independent of the market. So the two are scored on different strengths — out of sample, by Brier score and skill versus the market — and the live scorecard, not us, decides which (if either) beats the crowd.
The informed model plays the entire tournament 20,000 times. Each run:
Across all 20,000 runs we then tally how often each team reaches each round, wins the cup, and meets each opponent in the final — that is where the Outcome map's distributions come from. Two honest touches: the knockout rounds use a flatter set of ratings (a single match is far more of a coin-flip than a three-game group), and each run nudges every team's rating by a random amount to reflect that we don't know any team's true strength exactly — so the output is a genuine spread of outcomes, not false precision. The full formulas live on the Methodology page.
We don't invent them. The base numbers are real World Football Elo ratings (eloratings.net), captured at a fixed date, plus a modest home-advantage bonus for the three co-hosts. From a rating gap, the chance team A beats team B is the standard Elo logistic — a 400-point edge means about 10-to-1 odds. Once matches are played, each result updates both teams' ratings with a standard, margin-of-victory-aware Elo step (a bigger win, or an upset, moves the numbers more), so the forecast re-forecasts itself as the tournament unfolds. For the knockout phase we deliberately flatten the ratings toward the field average, because one-off matches are more random than a round-robin.
Plenty — and we would rather say so. The informed model is built from team-strength ratings, so it cannot see what the ratings don't already capture: injuries, suspensions, who actually starts, motivation, fatigue and travel, the weather, or a team peaking at just the right moment. It uses the official bracket, but the best-third assignments and the "no same-group rematch in the Round of 32" rule are handled as a close approximation, not the exact draw logic. And because the ratings are public, a model built on them is usually playing catch-up to a sharp market that already prices most of this in — beating it is meant to be hard. The zero-knowledge model is more limited still, by design: it adds no football knowledge at all, only a correction to the market's own prices. None of this is hidden — the Methodology page has a full "what's missing" section.
Yes. As results come in, the model re-forecasts: each played match updates both teams' Elo ratings (so later rounds reflect current form), completed group games are held fixed rather than re-simulated, and only the matches still to come are rolled forward. Crucially, the pre-tournament forecast stays frozen in an append-only ledger — that locked-in prediction is what gets scored, so the model can't quietly rewrite history to look good. The live board shows the updated view, and the as-it-unfolds strip near the top wakes up once the first games are played, highlighting what moved and which upsets landed.
Every forecast is timestamped and frozen before the tournament into an append-only ledger — a stamp is never edited. As results land, each call is scored out of sample by its Brier score and its skill versus the market. The receipts are the point, not any single pick.
Nobody knows yet — that is the experiment. Liquid markets are very hard to beat, and the honest answer is settled only by the live scorecard as matches resolve. The project is built to be falsifiable: if the models do not beat the market, the scorecard will say so.
A secondary, what-if view that puts a dollar number on the disagreements: a $1,000 paper bankroll, conviction-weighted and dollar-neutral, shown net of trading costs. No real money is involved — it is a way to size the edges, not a portfolio.
Use the top nav. From left to right: Scoreboard (the three contestants), Disagreements (where the zero-knowledge model fades the market most), 🔮 Outcome map (the model's full distribution over how far each team goes, with the single most-likely bracket at the bottom), 🏆 Bracket score (the round-weighted points race + per-round Brier), Board (the dense market-vs-model table), Books (the paper PnL), and Model (the informed-model summary). The Outcome map and Bracket score are top-level sections — no tabs to click through.
Each row is a team. The left block is the market's de-vigged price to clear each round — advance (last 32), reach the quarters, the semis, the final, and win it. The right block is the win-the-cup number three ways: the market, your chosen model, and the edge between them (model minus market). A green edge means the model rates the team higher than the market; red means lower. Use the toggle to switch the model column between zero-knowledge and informed, click any row to see the reasoning behind it, and click a column header to sort.
It is a distribution, not a prediction. Title race shows each team's chance to win the cup; how far each team goes shows where the simulation has each team bow out; the field narrows tracks who fills each round; and the bracket at the bottom is just the single most-likely path. Even the favourite usually wins only about one time in six. The Bracket score section immediately below it is the scoring layer — each side fills the bracket with its own probabilities, and as rounds resolve we tally a round-weighted points race (Informed model vs Market) alongside a per-round Brier.
The poll talks to a tiny external service. If that request is blocked — most often on a corporate network or by an ad / tracking blocker — the bubble can't load or record a vote. It generally works on a normal home or mobile connection.
The board is a static page, rebuilt from live data several times a day by a scheduled job, and again whenever played results are added. The timestamp at the top shows the last build.