World Cup 2026 prediction retrospective: can AI predict football, or is it just guessing?

How this test stays comparable

Every model received the same structured match packet frozen before kickoff, rather than being allowed to search the web. It included schedule and venue, public ranking snapshots, recent form, head-to-head and World Cup history, and public squad and injury summaries. Private odds and web browsing were excluded. The test therefore measures how models interpret the same public evidence, not which model is better at finding information.

For the knockout stage, settled results rebuild the real subsequent fixture before a new prediction is generated. A prediction can therefore use earlier results that were public at the time, but cannot follow a model's previously incorrect imagined bracket.

The Arena's per-match score has three parts:

In the current implementation, “probability calibration” means the probability the model assigned to the actual home/draw/away result. It is not a complete statistical calibration test; that distinction matters below.

Final leaderboard: two Chinese models take the top two places

The gap is small but not zero. DeepSeek V4 Pro's average score is 0.028, roughly 5%, above GPT-5.5. The top two models are both supplied by Chinese providers.

The total score is not the entire story. Claude Opus 4.8 and GPT-5.5 hit many more exact scores, suggesting that their scoreline distributions were not necessarily worse. The total score places more weight on outcome direction and probability on the actual result, where DeepSeek and Qwen performed better in combination.

“Always pick the home side” is a sanity baseline, not a market baseline

Actual outcomes across the 104 matches were 46 home wins (44.2%), 29 draws (27.9%), and 29 away wins (27.9%). A rule that ignores all team information and always picks the home side would achieve 44.2% outcome accuracy.

The five models achieved between 61.5% and 66.3%:

• Even Gemini 3.5 Flash and GPT-5.5 beat the always-home rule by 17.3 percentage points.
• DeepSeek V4 Pro beat it by 22.1 percentage points.
• None of the models collapsed into a near-random or simple home-bias rule.

This shows that the models are not merely ignoring teams and mechanically backing the home side. It does not show that they predict football better than a market. Teams have genuine differences in strength, and any credible pre-match baseline should use them.

The appropriate benchmark depends on what is being tested:

So the instinct to compare against a handicap is useful, but it is not the direct yardstick for this article's home/draw/away outputs. The direct market benchmark is 1X2; handicap markets test estimated goal difference. The most persuasive question is whether a model can add information beyond a strict pre-match market consensus without reading odds.

The Arena's match packets excluded private or premium odds, and its live archive did not retain historical pre-kickoff 1X2, handicap, or totals snapshots from traditional bookmakers. Its score therefore cannot be interpreted as “beating the line.” A later retrospective can, however, construct a prediction-market baseline from Polymarket's public price history.

Supplemental backtest: models versus Polymarket before kickoff

We found 104 standard home-win / draw / away Polymarket events and matched 101 of the Arena's 104 fixtures; three Curaçao fixtures had no corresponding market. For each binary contract, we took the latest historical price at least five minutes before kickoff. We retained only fixtures where all three contracts had a price no more than two hours old. That leaves 70 comparable matches. The median observation was about 55 minutes before kickoff, so this is best read as a market consensus about an hour before kickoff, not a final closing line.

The raw three-contract prices summed to 1.003 on average, close to a complete probability distribution. They were normalized by their sum before comparison, and every metric uses only pre-kickoff prices.

This gives a more meaningful and more restrained result than the always-home baseline. DeepSeek chose three more correct directions in these 70 matches, but did not beat the market on probability quality: Polymarket had the best Brier score and log loss. More importantly, the paired comparison contains only six model-only wins and three market-only wins, far too little evidence to declare a stable model lead.

The backtest therefore does not support the claim that AI beat a prediction market. The fairer statement is that the best model can be directionally competitive with the market and occasionally edge it, while the market's full probability distribution still represents uncertainty better. Historical traditional 1X2, handicap, and totals lines remain the stronger baseline that must be added next.

Limits of the capability: direction is possible; scores and rare events remain hard

Exact-score hit rates range from only 7.7% to 19.2%. That is not a surprising failure: a single shot, red card, goalkeeper mistake, penalty, or extra-time event can change a football score. Models can infer relative strength from public material without predicting these low-frequency, high-impact events.

Knockout matches make the point. Across 32 knockout fixtures, DeepSeek's outcome accuracy was 65.6% and GPT-5.5's was 62.5%, while Qwen3.7 Plus fell to 53.1% and Claude and Gemini were both 56.3%. Small samples are volatile; one World Cup knockout stage is not a final verdict on model stability.

The final is equally illustrative. The Arena settled Spain 0–0 Argentina: DeepSeek and Gemini called a draw, while the other three models did not. All five predicted 1–1, so none hit the exact score. They captured the outline of an even matchup, but not a dependable concrete result. Champion selection also requires a separate check against the eventual extra-time or penalty winner; it should not be replaced by a 90-minute score.

Chinese and US models: what this tournament does and does not show

For a simple comparison, DeepSeek V4 Pro and Qwen3.7 Plus are grouped as Chinese models; Claude Opus 4.8, Gemini 3.5 Flash, and GPT-5.5 as US models. Aggregating model-match observations gives:

Three limited observations follow.

First, the Chinese group was better at the home/draw/away call in this common task and assigned more probability to what eventually happened; that explains its total-score lead.

Second, the US group recorded more exact-score hits, driven mainly by Claude Opus 4.8 and GPT-5.5. Their score outputs appear sharper, but that sharpness did not turn into an aggregate scoring advantage.

Third, this is not a conclusion that Chinese models are generally better than US models. There is only one World Cup, unequal group sizes, different versions and provider interfaces, and score prediction is sensitive to prompts, temperature, output constraints, and the pre-match data packet. The evidence only says that, under this packet and these 104 matches, the two selected Chinese models averaged better results; it should not be generalized to countries or entire model ecosystems.

How current AI should be evaluated for sports prediction

The most appropriate label is neither “oracle” nor “dice roller,” but a match-information synthesizer constrained by data quality.

Its strength is compressing rankings, form, head-to-head records, squad news, and injury information into consistent home/draw/away probabilities and explainable judgments. In this sample, that synthesis was better than a simple home-bias rule. Whether it outperforms market consensus cannot be answered from the original Arena data alone.

Its limits are just as clear. It has no live match information, no dependable private injury intelligence, and no causal model for rare events; general-purpose language models are not trained specifically for probabilistic forecasting. The more specific the score, the closer the teams, and the later the round, the more visible those limits become.

How the next edition should be measured more rigorously

The Arena has shown that a common frozen packet and per-match archive are feasible. The next edition should add five things:

1. Frozen traditional baselines: Elo, a ranking/form logistic regression, and de-margined pre-kickoff 1X2 implied probabilities.
2. Separate archived snapshots of pre-kickoff 1X2, Asian handicap, and totals prices with source timestamps. Odds must remain outside model input and be used only for offline comparison.
3. An immutable pre-match timestamp, input-packet hash, and model parameters for every prediction, with automatic proof that it predates kickoff.
4. Brier score, log loss, reliability plots, handicap win rate, and CLV alongside the current score, to test calibration and market increment properly.
5. Repeated experiments across tournaments, leagues, and seasons before making claims about model, regional, or architectural differences.

The restrained conclusion from this World Cup is that AI can turn public football information into directional predictions better than a simple home-bias baseline. Without a pre-match market baseline, it cannot claim to be “more accurate,” much less to have beaten the line. It cannot replace specialist models, markets, or sustained human data work. Using it as a decision aid and source of hypotheses is reasonable; treating it as a certain answer or betting advice is not.

Data notes

• Live results come from SportMonks-backed Arena settlement data; this retrospective uses the platform's 104 settled match results.
• Predictions use fixed, auditable pre-match packets, with no web browsing or private odds.
• Traditional bookmaker 1X2, handicap, and totals histories were not archived by the Arena. This article does not claim to beat traditional lines. The Polymarket supplement covers 70 matches and represents prices about an hour before kickoff.
• All percentages are calculated from model-by-settled-match records; rounding can create small terminal differences.
• This is an evaluation of Arena results, not betting advice.