World Cup Winner Markets vs Elo Ratings: Which Forecasts Better?

Compare winner-market odds vs Elo simulations for World Cup champions using log loss/Brier, plus a checklist for liquidity and bias.

Define “forecasts better”

For “fifa world cup who will win,” “better” can’t mean “picked the champion once.” A World Cup is a one-off draw from a huge space of possible paths, and in 2026 it’s a 48-team, 104-match tournament. If you grade only a single yes/no winner pick, you’re throwing away almost all the information a forecast can express.

What we want to compare is: given the same point in time, which approach gives the most accurate probability distribution over all teams winning the tournament (and, ideally, other outcomes like finalists). That’s a forecast you can score even when the eventual champion wasn’t the top-ranked team.

Even a very strong favorite is still unlikely to win an entire tournament. So a “hit rate” of 0% or 100% from one World Cup mostly measures luck, not forecasting quality. You need methods that reward being “less wrong” when the upset happens.

Probability, not rank

A ranking (including an Elo table) answers “who’s stronger right now,” not “what’s each team’s chance to lift the trophy.” Those are different objects.

A winner forecast should allocate probability mass across all contenders. That forces you to be explicit about uncertainty: whether your #1 and #2 are close, or whether the field is wide open. It also makes forecasts comparable. Two systems can agree on the ordering but disagree sharply on the gap between teams—and the gap is where the money is.

That’s why a single headline pick is a dead end for evaluation.

Score and diagnose

To say one forecast “does better,” you need a scoring rule that grades probabilities, not just picks.

• Calibration vs discrimination: Calibration asks whether events assigned 20% happen about 20% of the time; discrimination asks whether higher probabilities are consistently assigned to outcomes that occur.
• Proper scoring rule: a way to score probabilistic forecasts that, in expectation, rewards truthful probabilities rather than just naming the most likely outcome.
• Log loss (logarithmic / ignorance score): a proper scoring rule that penalizes overconfidence by taking the negative log probability assigned to what actually happened.
• Brier score: a (strictly) proper scoring rule based on squared error between forecast probabilities and outcomes.

Read the scores alongside a diagnosis: sharp forecasts with bad calibration are just overconfident. And a well-calibrated forecast that’s never willing to separate teams isn’t very useful.

One concrete tell that you’re in the right evaluation frame: forecasts can be meaningfully “good” even when they assign, say, only about a 20.5% chance to the final matchup that actually happens.

Make them comparable

To compare “fifa world cup who will win” forecasts fairly, you have to force both sources into the same object: a timestamped probability distribution over teams winning the tournament, with probabilities that sum to 1 so they can be scored.

That also fixes a common category error: markets naturally quote tournament-winner chances, while Elo naturally outputs match-level strength. If you don’t standardize the output, you end up “grading” measurement choices (like last trade vs midpoint) instead of forecasts.

Winner market probabilities

1. Start with the contract you’re actually reading. In many venues this is a binary event contract: one share settles to $1 if you’re right and $0 if you’re wrong.

2. Convert price/odds into an implied probability. Implied probability means a probability inferred from a contract price or odds quote (for example, a $0.30 “YES” share is read as ~0.30).

3. Pick a price convention and stick to it. “The market” doesn’t give you one number—there’s last trade, best bid, best ask, and midpoint. If you grade last trade you can overweight a stale print; if you grade midpoint you can overweight a wide spread. This choice can hand one approach an unearned advantage.

4. Normalize across teams and account for frictions. In a “winner” menu, the raw implied probabilities often don’t sum to 1 because of margin/fees or because the set of outcomes isn’t perfectly exhaustive. The comparable object is the normalized vector (divide each team’s implied probability by the sum across listed teams), and you should keep the pre-normalization sum as a diagnostic for how much “vig” you’re washing out.

Elo to winner odds

1. Use Elo for what it natively provides: match win expectancy. An Elo table (for example, World Football Elo Ratings at eloratings.net) gives you a probability-like expectation for team A vs team B, not “win the World Cup.”

2. Update ratings consistently. If you’re updating ratings yourself, document the update rule you’re using and don’t switch midstream. FIFA’s men’s ranking procedure uses a model called SUM, which took effect in August 2018 after approval by the FIFA Council (see FIFA’s men’s ranking procedure).

3. Turn match probabilities into tournament-winner probabilities via simulation. Tournament simulation (Monte Carlo) means running many simulated versions of the full bracket using match-level win probabilities, then counting how often each team wins the tournament.

4. Simulate the same “state of the world” the market saw. Elo-to-winner conversion is where bracket mechanics enter: group assignments, knockout paths, and any draw shocks. If your simulation uses a different draw or ignores the actual bracket, you aren’t comparing it to a market that has already priced that structure.

When you later score forecasts, the outcome you’re scoring against is concrete (a champion, and optionally a specific final matchup). The whole point of this normalization step is that your methods had to assign probabilities to every other team (and every other finalist pairing) too.

Probability hygiene: a step-by-step apples-to-apples workflow

If you want to answer “fifa world cup who will win” with a fair market-vs-Elo comparison, lock down three choices up front: the timestamp, the normalization rule, and the exact market price you’re grading. Here’s a reproducible workflow.

1. Pick one forecasting moment (timestamp) and reuse it for both.
   Decide what “as of” means for your test and don’t blur it later. Common choices are: pre-tournament (before the opening match), immediately after the draw (once group/bracket structure is known), or after a specific matchday. Whatever you choose, record it and make sure your market snapshot and Elo snapshot are taken from the same moment.

2. Freeze the market snapshot at that timestamp.

• Select a single price convention (last trade, midpoint, best bid, or best ask) and apply it to every team.
• If your data source provides an explicit “Last updated” time, keep it with the snapshot so you can tell “fresh” prices from stale ones.

3. Convert each team’s chosen market price into an implied probability.
   For binary event shares that pay $1 if the team wins and $0 otherwise, the simplest read is: price 3 implied probability for that outcome.

4. Normalize the market menu into a proper distribution.

• Sum the raw implied probabilities across all listed teams.
• Divide each team’s raw implied probability by that sum so the probabilities add to 1.
• Keep the pre-normalization sum as a diagnostic. If it’s far from 1, you’re washing out a lot of margin/fees and you should expect the normalization choice to matter.

5. Freeze the Elo snapshot at the same timestamp (same information set).
   Use the Elo ratings as they stood at that moment—don’t “peek” at later matches when building the simulation inputs.

6. Generate tournament-winner probabilities from Elo using the same tournament state the market saw.
   Run a tournament simulation that matches the real structure available at that timestamp (groups/knockout paths once known). The output you keep is the tournament-winner probability for each team.

7. Align the outcome set (teams) between the two forecasts.
   Markets sometimes list “field/other” outcomes or omit longshots; Elo simulations can include every qualified team. To stay apples-to-apples, ensure both probability vectors refer to the same set of teams. If one side includes an “Other” bucket, you either (a) include an equivalent bucket on the other side, or (b) drop/merge outcomes consistently and renormalize.

8. Only now, grade the forecasts.
   When you score market vs Elo (log loss, Brier, calibration plots), you’re grading two timestamped, normalized winner-probability distributions—not a mixture of timestamp differences, spread conventions, or missing outcomes.

What evidence suggests

Across football forecasting studies that actually score probabilities, odds- and market-derived signals are usually a tough baseline to beat—but not a perfect one.

For a trading/pricing prior, that points to a simple default: treat a liquid winner-price snapshot as the baseline and make Elo clear it on proper scoring—not on a one-tournament “picked the champ” anecdote. The table’s match-forecast evidence is the same story in numbers (e.g., 1.380 vs 1.403 informational loss, and near-identical Bundesliga hit rates with 0.844 correlation): markets/odds tend to be hard to beat, and differences can be small enough that market design and liquidity decide which one you should trust on a given day (see the EURO 2008 comparison in Forecasting sports tournaments).

When markets fail

Winner markets are a strong default for “fifa world cup who will win” only when you can treat the quoted prices as a clean, widely-shared aggregation of belief. When that assumption breaks in observable ways, an Elo-only tournament simulation becomes the better fallback baseline—not because Elo is magic, but because “the market” you’re looking at isn’t doing the job you think it is.

Two common failure modes are structural: participation (not enough real money/attention at the moment you’re sampling) and contract mechanics (a design that makes the probabilities hard to interpret even if trading is active). A third is timing: you’re sampling right when information has changed but trading hasn’t caught up.

Bias, access, timing

Use this checklist as switching conditions—if you hit one, default to Elo simulation unless you have a clear reason not to.

• Thin participation (wide or jumpy prices): If prices move in chunks or sit stale, you’re not seeing a stable consensus. Wolfers & Zitzewitz frame “prices equal mean beliefs” as conditional, and they explicitly flag cases where efficacy can be undermined—like extreme prices near $0/$1 and/or constrained trading volume.

• Access constraints (your venue isn’t the venue): If you’re a U.S. reader and the page you’re referencing won’t let you trade, you’re not necessarily seeing the most informative pool of participants. Polymarket’s World Cup page shows an interstitial stating trading is blocked in the United States on polymarket.com and redirects to polymarket.us.

• Bad winner-contract structure: If each team is a separate binary contract, the menu can embed inconsistencies (the raw implied probabilities need not add to 1), and your “normalization” step starts to drive the ranking. A single multinomial-style winner contract (one book where outcomes are mutually exclusive) removes that degree of freedom.

• Favourite-longshot bias (or its reverse): Favourite-longshot bias—a systematic distortion where longshots’ implied probabilities are too high (or, in reverse cases, too low) relative to realized frequencies—is a sign the pricing is being pushed by taste, storytelling, or constraints, not just information. The direction isn’t guaranteed: a World Cup prediction-market experiment reported a reversed favourite-longshot bias while also describing bookmakers as perfectly calibrated in their analysis.

• Extreme prices aren’t “certainty”: On U.S.-regulated Polymarket, order prices are bounded (no orders below $0.001 or above $0.999). Treat prices pinned near the rails as a warning label: they’re exactly where “price-as-probability” assumptions are most fragile.

If you can’t clear this checklist, don’t argue with the market—stop pretending you’re reading one, and use Elo simulation as your conservative default.

Bias, access, timing

Use this checklist as switching conditions—if you hit one, default to Elo simulation unless you have a clear reason not to.

• Thin participation (wide or jumpy prices): If prices move in chunks or sit stale, you’re not seeing a stable consensus. Wolfers & Zitzewitz frame “prices equal mean beliefs” as conditional, and they explicitly flag cases where efficacy can be undermined—like extreme prices near $0/$1 and/or constrained trading volume (see Interpreting Prediction Market Prices as Probabilities).

• Access constraints (your venue isn’t the venue): If you’re a U.S. reader and the page you’re referencing won’t let you trade, you’re not necessarily seeing the most informative pool of participants. Polymarket’s World Cup page shows an interstitial stating trading is blocked in the United States on polymarket.com and redirects to polymarket.us.

• Bad winner-contract structure: If each team is a separate binary contract, the menu can embed inconsistencies (the raw implied probabilities need not add to 1), and your “normalization” step starts to drive the ranking. A single multinomial-style winner contract (one book where outcomes are mutually exclusive) removes that degree of freedom.

• Favourite-longshot bias (or its reverse): Favourite-longshot bias—a systematic distortion where longshots’ implied probabilities are too high (or, in reverse cases, too low) relative to realized frequencies—is a sign the pricing is being pushed by taste, storytelling, or constraints, not just information. The direction isn’t guaranteed: a World Cup prediction-market experiment reported a reversed favourite-longshot bias while also describing bookmakers as perfectly calibrated in their analysis.

• Extreme prices aren’t “certainty”: On U.S.-regulated Polymarket, order prices are bounded (no orders below $0.001 or above $0.999). Treat prices pinned near the rails as a warning label: they’re exactly where “price-as-probability” assumptions are most fragile.

Decision rubric

For “fifa world cup who will win,” the only defensible “predicted winner” you can state without guessing is the team sitting at the top of a timestamped, normalized winner-probability distribution from the source you chose (market or model). “Most chances to win” is the same thing: highest probability mass, not #1 on a ratings table.

Because the FIFA World Cup 2026 format is 48 teams and 104 matches, treat data freshness as part of the forecast—not a footnote: the timestamp you sampled is inseparable from the probability you’re quoting (see FIFA’s Host Cities video).

If you can’t write down the timestamp, the probability source, and how you’d score it later, you’re not holding a forecast—you’re holding a screenshot.

Use markets first, switch fast

If you want a World Cup winner probability you can actually trust, start from a liquid, timestamped winner-market snapshot and treat it as the baseline—then judge it on proper probability scores, not whether it named the champion. The moment the “market” fails the observable checks (thin/stale prices, access constraints, messy per-team contracts that force you to normalize inconsistencies, or favourite-longshot distortions), stop treating the quote as wisdom and fall back to an Elo-based tournament simulation of the same draw state. When you’re not sure you’re looking at the most informative venue, do a cross-platform scan for the top team and the size of the gap so you don’t anchor on a single thin or inaccessible price. Your first step is simple: write down the timestamp, the full normalized probability vector, and the scoring rule you’ll use later—if you can’t, you’re not holding a forecast.