After Spain’s World Cup: Inside the Models Trying to Forecast Football’s Next Winner
Last Updated on 10 September 2026
Spain left the 2026 World Cup with a second star on its shirt and a useful warning for anyone who treats a forecast as a verdict. Ferran Torres’s goal in the 106th minute settled a 1–0 final against Argentina, but the scoreline concealed a much more one-sided contest: Spain had 20 attempts to Argentina’s two, according to the official match statistics cited by Reuters. Across eight matches, Spain conceded only once.
That is the kind of ending that makes a prediction look simple after the event. Spain had been widely regarded as a leading candidate, and the Opta supercomputer had placed Luis de la Fuente’s team first in its pre-tournament projections. Yet it gave them a 16.1% chance of winning across 25,000 simulations, leaving room for France, Argentina, England or a surprise from outside the traditional elite.
The distinction matters as attention turns towards the next World Cup. Forecasting the next winner is not finding a machine that knows the future. It is translating incomplete information into probabilities, then checking whether they remain honest when football produces a result no model can force.
The favorite was right — and still not certain
Opta’s 2026 forecast offers a neat lesson in how tournament probabilities should be read. Spain was the leading pre-tournament choice, but 16.1% also meant that the model expected Spain not to win in roughly five out of six simulated tournaments. That is not a contradiction. It is precisely what a probability is supposed to express in a sport with few goals, a short competition and a knockout bracket.
The projection moved as matches removed teams from the field. By the quarter-finals, Opta gave Spain a 21.3% chance of lifting the trophy, while France stood at 27.3%, England at 16.5% and Argentina at 17.3%. Spain’s probability had risen because fewer opponents remained, even though the team had not suddenly become 34% stronger than it was before the opening match. The tournament had simply become more legible.
Spain’s victory also shows why a single correct champion is a weak test. A model assigning 16% to Spain and 12% to France may be more informative than one giving Spain 40% and France 5%, even if both name Spain. The first represents the field. The second may only have guessed loudly.
What a World Cup forecast is actually predicting
A World Cup model normally performs two related tasks. First, it estimates the probabilities of individual matches: home win, draw and away win, or a complete distribution of possible scores. Second, it feeds those match probabilities through the tournament structure to estimate group positions, knockout qualification, the chance of reaching each round and the chance of becoming champion.
The second task is where casual predictions often become misleading. A team can be strong enough to beat most opponents and still receive a modest title probability if its likely route includes several elite sides. Another team with a similar underlying rating may draw a softer group or avoid the strongest bracket until the final. The model is not only asking, “How good is this team?” It is also asking, “What sequence of opponents is it likely to face?”
The 2026 format—48 teams and 104 matches—made that calculation more demanding, creating more possible paths and more occasions for injuries, suspensions, tired legs or one deflection to alter the bracket.
The first layer: ratings, form and the strength of opposition
Most serious forecasting systems begin with a rating of team strength. Elo-style ratings are popular because they update after every match and reward a team more for beating a strong opponent than a weak one. A rating can be adjusted for the importance of the match, the margin of victory, the venue and the time that has passed since the result. The underlying idea is modest: results contain information, but recent and meaningful results should generally carry more weight than an old friendly.
FIFA’s men’s ranking also uses an Elo-based approach, although it is a ranking system rather than a complete prediction engine. Independent models may use FIFA points, World Football Elo, their own rating scale or an ensemble of several ratings. What they share is the attempt to estimate latent strength instead of treating every win as identical.
The challenge is that national teams play relatively few matches. A club can offer a season of evidence; a national side may have only a handful of competitive games before a major tournament. One impressive result can therefore move a rating too far, especially when the opponent is depleted or the match is a friendly. Ratings are also slow to understand a new coach, a returning midfielder or an emerging generation.
From ratings to goals: the Poisson engine and expected goals
A result-based model can estimate the chance of a win, draw or defeat directly. A score-based model takes another route: it estimates how many goals each side is likely to score and derives the match result from that distribution. Poisson regression has long been used for this task because football goals are count data and are relatively rare. More advanced versions can allow the teams’ scores to be related, give extra weight to 0–0 and other low-scoring outcomes, or adjust for the state of the match.
The advantage of a score model is that it can estimate 0–0, 1–0, 1–1 or 2–1, preserve the goal difference that matters in a group table and calculate each team’s chance of advancing. A published World Cup model built around a nested zero-inflated generalized Poisson regression used Elo ratings, attack and defensive strength, match location, match importance and time depreciation before running Monte Carlo simulations. The architecture is not magic, but it makes the assumptions visible.
Expected goals, or xG, adds a process-based layer. Instead of asking only whether a shot went in, an xG model assigns a probability based on features such as location, angle, body part, assist type and, depending on the provider, the defensive context. A team that repeatedly creates high-quality chances may look healthier than its recent scoreline. A team winning through a series of low-probability finishes may be due for regression.
Yet xG is not a universal measurement. Providers use different definitions, event feeds and training data, so two xG totals are not automatically comparable. International football brings another problem: a national team may not produce enough comparable shots under the same coach and with the same players to support a stable estimate. Club data must also be adjusted for teammates, opponents, tactical instructions and competition level.
Why a low score can mislead
The 2026 final illustrates the point. Spain’s 1–0 victory was not a narrow contest in the way the score might suggest: it controlled territory and generated the first 20 attempts, while Argentina had none during the 90 minutes. But the goal arrived in extra time, after a substitution and a moment of space. A dominant process raises the probability of winning; it does not dictate the scorer.
The players who turn a team rating into a real lineup
The next generation of models is increasingly concerned with the difference between a squad and the eleven players actually available. Forecasting a national team four years before a tournament involves uncertainty about injuries, retirement, development, club minutes, tactical roles and the quality of replacements. The name on a roster is not the same as the contribution available on match day.
This is why line-up news can move a forecast more than a recent result. A model that knows the likely starters can update attack and defence separately, estimate the impact of a missing goalkeeper or holding midfielder, and widen the uncertainty when the replacement is untested. It can also model substitutions, although that requires reliable information about bench quality and likely game states.
Spain’s final offered a clean example. Torres came off the bench and scored in the 106th minute, while Nico Williams helped change the attacking rhythm. A model may capture depth in broad terms; it cannot know which substitute will create the decisive sequence. It estimates player impact, not a script for the match.
The bracket is part of the probability
Tournament simulations are often described as a giant coin toss repeated thousands of times. Each run begins with assumptions about ratings, lineups, goals, venue and rules; it draws scores, updates the table and advances teams through the knockout rounds. The title percentage comes from repeating that process.
The draw and the route remain crucial. A team that is projected to finish first in its group does not receive the same title chance as a team projected to finish second, because the next opponent and the later bracket may change. Penalty shoot-outs need their own treatment. So do extra time, red cards, travel, climate and the possibility that a team protects a lead rather than continuing to attack. Even a small change in a match probability can have a large effect once it is multiplied through several knockout rounds.
That will be especially relevant to the 2030 World Cup, scheduled mainly across Spain, Portugal and Morocco, with centenary fixtures in Argentina, Paraguay and Uruguay. A forecast must decide how much weight to give familiar conditions, travel patterns and home support, while remembering that the competition is still years away. Venue advantage is real enough to model, but it is not a guarantee of a deep run.
At the public-facing end of the process, a service such as NerdyTips translates a probability distribution into a readable football tip. That is useful only when the underlying probability is treated as an estimate rather than a promise. The same principle applies to a World Cup champion forecast: clarity about uncertainty is more valuable than a confident label.
The market is another model — and a formidable one
Bookmaker prices are a comparison point because they aggregate information a public model may not see immediately: training reports, local knowledge, likely lineups, professional opinion and money committed to a position. The price includes a margin, but it can be converted into a market estimate.
The market is not infallible—famous teams can be inflated and long-range prices can reflect commercial exposure—but any model should be compared with it. Beating an informed consensus is much harder than beating a weak prediction.
The strongest approach may be an ensemble: ratings for stable team strength, a goal model for scoring processes, a player layer for lineups and injuries, and the market as a broad information set. Combining them can reduce the chance that one noisy input dominates, but a complicated forecast that nobody can audit is not automatically better than a simpler model with honest limits.
Why “10,000 simulations” proves almost nothing by itself
The number of simulations is one of the easiest details to market and one of the least important to understand. Increasing a run from 10,000 to 100,000 reduces the random error created by the simulation itself. It does not repair a rating that overvalues old results, an xG feed that misses context or an injury assumption that is wrong.
This is the difference between simulation noise and model uncertainty. If a model produces a 12.0% title probability in 10,000 runs and 12.2% in 100,000, the estimate is more stable, not 10 times more accurate. False precision begins when a number such as 12.17% is presented without showing its sensitivity to lineups, ratings, venue or penalties.
A responsible forecast publishes scenarios or intervals: what if the starting striker is unavailable, the team draws the hardest group or recent friendlies are removed? Those questions reveal more than another zero added to the simulation count.
What Spain’s victory should teach the forecasters
The correct post-tournament exercise is not to ask which model named Spain. It is to ask whether the models were calibrated. If a system repeatedly assigns 20% probabilities to a class of events, that class should occur roughly one time in five over a large sample. Calibration can be checked with log loss, Brier score or the ranked probability score.
Tournament forecasting needs a long memory. One World Cup is too small a sample to prove that a method works, especially when the format changes. Analysts can back-test on earlier tournaments, use only information available at the time and compare against simple baselines. The test should include all teams and measure group predictions, match probabilities and knockout progression separately.
Spain’s 2026 campaign should update the models, but not through a crude trophy bonus. The title is evidence that the team combined strength, depth and resilience at the right moment. It is not evidence that every Spanish performance will remain at the same level, that every opponent will play the same way or that the next generation will inherit the exact same balance. Ratings should move; assumptions should be reviewed; the uncertainty should remain.
Can Spain win again in 2030?
The four-year gap
It is too early for an honest model to offer a precise 2030 percentage that readers should take literally. The tournament is close enough for its host geography to be known, but far enough away for the most important football variables to change. Four years can remove a goalkeeper, create a midfielder, reshape a defence and turn a promising teenager into the central player of a generation.
Spain will enter the conversation because it has recent tournament success, a coherent identity and the advantage of playing much of 2030 at home. France, Argentina, Brazil, England, Portugal and others will have plausible paths, but a ranking of names is not a forecast. Qualifiers, Nations League matches, club seasons and injuries still have to be absorbed.
The better question is not “Who will the model pick?” It is “What does the model know, what does it assume and how often has it been right at this confidence level?” Spain’s World Cup has made that question more interesting. The champions were forecast as a leading possibility, then survived the margins separating a favorite from an eliminated team. Football’s next winner will emerge from the same tension between measurable strength and unmeasurable moments.
Models can organize the uncertainty. They can expose a difficult bracket, price the absence of a key player and distinguish a good process from a lucky scoreline. They cannot remove the uncertainty. The most credible forecast for the next World Cup will be the one that remembers this final lesson from Spain: being the most likely winner is valuable information, but it is never the same as being destined to win.