Before the 2026 World Cup begins, two authoritative systems have released their respective "win probabilities"—and their top-ranked teams differ.
Prediction markets (Polymarket, Kalshi price aggregation) rank France as the top favorite at approximately 17%. Opta’s supercomputer ranks Spain as the top favorite to win the European Championship at 16.1%.
Both numbers appear to be "probabilities," but they are generated in completely different ways—one is a price determined by market clearing with hundreds of millions of dollars in trading volume, and the other is a frequency derived from a supercomputer simulating the entire World Cup ten thousand times.
This article does not predict who will win or evaluate which set of data is more accurate; it only answers one question: when you see the number "France 17%," how was it derived, and how reliable is it?
This is the next level down from EP06—previously, we discussed how the market structures of prediction markets and traditional gambling differ; this article explains how the probability embedded in prices is calculated. Data as of May 31, 2026.
Act One · Probability in Price: How the Market Produces Probability
The mechanism of prediction markets is straightforward: the contract price for each outcome ranges from 0 to 100 cents, and the price directly reflects the implied probability. The France contract is priced at 17 cents, meaning the market assigns approximately a 17% chance of France winning—the correct guess earns $1 per contract, while incorrect guesses earn $0.
However, prices on a single platform can be noisy. Aggregators like DeFi Rate compile quotes from multiple venues—including Kalshi, Polymarket, Polymarket US, and Gemini—using volume-weighted average prices (VWAP) on an hourly basis to derive a cross-platform implied probability. As of May 30, 2026, the World Cup champion contract had accumulated approximately $523 million in trading volume, with a settlement date set for July 20, 2026—the day after the final on July 19.
This price did not appear out of nowhere. It is the result of market makers continuously quoting bid and ask prices, combined with traders constantly executing trades. Notably, the liquidity providers for prediction markets are exclusively crypto-native institutional trading firms: Wintermute, with annual trading volume exceeding $3.5 trillion and coverage across more than 70 exchanges, began providing bid-ask quotes for Polymarket and Kalshi in 2026; Jump Trading and Susquehanna are also actively market making.
Wintermute’s OTC trading head, Jake Ostrovskis, summed up the current state of the market in one sentence:
Prediction markets have the demand profile of a major asset class but the liquidity profile of an early-stage one.
Prediction markets have the demand scale of major asset classes but only the liquidity depth of an early-stage market. In other words—the credibility of the "probability" embedded in the price depends on how much real liquidity supports it. We’ll return to this in Act Three.
Act Two · Probability in Simulation: How Models Generate Probabilities
Opta’s supercomputer takes a different approach. It first uses team data—form, historical performance, world rankings, and recent international results—to estimate the probability of win, draw, or loss for each match using Power Rankings, an Elo-derived rating algorithm. It then simulates the entire World Cup 10,000 times and counts how many times each team wins the tournament; that frequency becomes its "probability of winning."
Results for 2026 (stated as facts, not predictions): Spain 16.1% (also the only team with a probability exceeding 50% to reach the quarterfinals at 52.1%), France 13.0%, England over 10%, defending champion Argentina in fourth place also over 10%, Portugal 7.0%, Brazil 6.6%.
Here’s an counterintuitive methodological detail worth noting: One of the inputs to the Opta model is the betting market odds. This means the comparison between "market vs. model" is not between two entirely independent systems—the model has already partially incorporated market information. When you compare market prices with Opta probabilities, the differences you observe are smaller than those between two independent sources.
Please note a timing issue: The authoritative FiveThirtyEight soccer model (SPI), remembered by many, ceased updates after its founder, Nate Silver, departed in 2023; the original website shut down in September 2023, and the entire 538 platform was discontinued by ABC in March 2025. This article treats it solely as a historical methodology and comparative reference for the 2018 and 2022 tournaments, not as an active forecasting source for 2026.
Act Three · Who’s More Accurate? An Honest Blank
Which is more accurate: the market or the model?
The honest answer is: no rigorous, peer-reviewed academic study has directly compared the Brier scores (a standard measure of prediction accuracy) of prediction markets with those of Opta/538 during the 2018 and 2022 World Cups. Figures such as the platform’s claimed “90% accuracy” typically come from the platform itself or non-peer-reviewed blogs and cannot be considered independent conclusions. This document explicitly acknowledges this gap rather than fabricating an answer.
But there’s a commonly misstated case worth correcting. Many say “Argentina’s 2022 championship was a huge upset”—but that’s inaccurate. Before the tournament, Argentina was the second or third favorite: Opta gave them a 13.1% chance (second), and bookmakers offered +500 odds (roughly 16.7%, also second). The real story isn’t “an underdog won,” but rather—nearly all major models and markets bet on Brazil, yet the second favorite, Argentina, won; the only outlier that pegged Argentina at around 8% was FiveThirtyEight. This is more precise and more telling: so-called “authoritative probabilities” can vary by a factor of two across different sources.
Price itself is not a perfect probability. A phenomenon verified repeatedly for nearly a century is called the longshot bias: in classic horse racing markets, bettors systematically overestimate longshots and underestimate favorites—actual win probabilities for longshots are lower than the odds suggest, leading to greater long-term losses from betting on longshots (Snowberg and Wolfers).
The truly counterintuitive aspect is that this bias has not disappeared in crypto prediction markets, which are touted as more rational and efficient. Multiple studies based on vast datasets from Polymarket and Kalshi have consistently identified the same directional bias—University College Dublin analyzed over 300,000 Kalshi contracts and found that low-priced contracts were settled at rates lower than their implied probabilities, while high-priced contracts were settled at rates higher than implied (i.e., longshots remain overvalued). A calibration study based on 292 million trades (arXiv preprint 2602.19520) also found that long-duration contract prices are systematically compressed toward 50%, underestimating the true advantage of favorites. A microstructure preprint (arXiv 2604.24366), analyzing 30 billion order book events over 52 days, quantified the cost of the longshot side: the bid-ask spread for the lowest-probability contracts reached 1,300 to 1,800 basis points—an order of magnitude higher than in traditional markets—due to market makers pricing in inventory risk characterized by "bounded upside and asymmetric downside."
In other words: a bias recorded at racetracks a century ago still holds true today in on-chain markets with billions of dollars in trading volume—the "probability" embedded in prices becomes less reliable the closer it is to the longshot end.
The ledger is public.
Here’s something traditional sports betting cannot do: Polymarket is built on Ethereum smart contracts, where every trade is recorded on-chain and publicly auditable. The two studies mentioned above were possible precisely because researchers could directly reconstruct the direction of every trade from on-chain transaction records—something impossible in traditional betting with closed ledgers. Settlement is also on-chain: trades are collateralized with USDC and automatically settled by smart contracts, eliminating the need to trust a centralized bookmaker to hold your funds.
But transparency does not equal immutability. A thin order book means that small markets can be easily moved by small amounts of capital. During the event period (June 11 to July 19), contract prices for each match will drift in real time according to the score—that will be the most vivid live example of "how prices are formed."
Act Four · Variables Beyond Price: Regulation
Price is also influenced by a non-market factor: regulatory uncertainty.
On May 18, 2026, the Governor of Minnesota signed SF4760 into law, becoming the first U.S. state to classify the operation and advertising of prediction markets as a felony (effective August 1, 2026). Within 24 hours, the CFTC (Commodity Futures Trading Commission) filed a lawsuit, and Kalshi was sued on May 28. CFTC Chair Michael Selig stated:
This Minnesota law turns lawful operators and participants in prediction markets into felons overnight.
This Minnesota law turned legal operators and participants of prediction markets into felons overnight.
Behind this lies an unresolved jurisdictional dispute: On April 7, the Third Circuit Court of Appeals ruled in favor of Kalshi (event contracts are derivatives and fall under the CFTC’s jurisdiction), while on April 16, the Ninth Circuit heard Nevada’s appeal and appeared to side with Nevada—this split between the two circuits may ultimately be resolved by the Supreme Court. To date, 17 states are challenging prediction market operators, and 14 states have related legislation; Spain has ordered ISPs to block Polymarket and Kalshi as of 2026.
It is crucial to distinguish between two things: prediction markets follow the federal regulatory path under CFTC event contracts, while sports betting follows state licensing pathways—the same World Cup contract has entirely different legal statuses across jurisdictions. Regulatory uncertainty itself is a variable underlying the price.
Conclusion · Back to Those Two Numbers
Back to the beginning—“France 17%” and “Spain 16.1%”.
Now you understand where these two numbers come from: one is the price cleared in the market based on hundreds of millions in trading volume, influenced by longshot bias and liquidity depth; the other is the frequency derived from a supercomputer simulating the entire World Cup ten thousand times, affected by model lag and partially incorporating market information.
Which is more accurate? No rigorous cross-election comparison can answer this question. Beibei will publish a post-event review after the World Cup ends and contracts settle on July 20, analyzing what the market and the model got right and what they got wrong.
Before that, whenever you see another "championship probability," it's worth asking: How was this number calculated?