Why Soccer Is Harder to Model Than Most Sports
Because scoring events are rare. A low-scoring match is a small sample, so outcomes carry more noise relative to skill than in higher-scoring sports, draws add a third outcome that must be modeled explicitly, and results validate slowly. Event-level metrics such as expected goals exist mainly to recover signal that scorelines lose.
Low scoring is the root cause
Most of what makes soccer difficult follows from one property: goals are rare.
A basketball game produces a large number of scoring events, so the final margin aggregates many independent-ish contributions and the better team wins reasonably often. A soccer match produces very few, so a single deflection, offside call, or save moves the result in a way that no amount of underlying superiority reliably overcomes.
Statistically, each match is a small sample. That has three consequences that compound.
Outcomes are noisy relative to team quality, so results are a weak signal about which side was better.
Models converge slowly, because the number of informative events per match is small and the number of matches in a season is limited compared with sports playing far more games.
And apparent edges are hard to distinguish from variance, because the noise floor is high relative to any plausible difference in skill or model quality.
None of that makes soccer unmodelable. It means the honest error bars are wider than in higher-scoring sports, and a method that treats them as narrow will look excellent in-sample and disappoint afterwards.
The draw, and why structure matters
A meaningful share of matches end level, which makes the draw a real outcome rather than an edge case.
That changes model structure rather than just adding a class. Approaches that work in two-outcome sports need reworking, and the common designs handle it differently.
Goal-based models. Model each side's scoring rate and derive the outcome distribution, which produces the draw naturally as the probability of equal scores. This tends to respect the underlying process and requires assumptions about how the two sides' scoring relates rather than being independent.
Ordered models. Treat the outcome as ordered from away win through draw to home win, which captures the ordering and loses the scoreline detail that carries useful information.
Direct multiclass. Predict three probabilities directly, which is flexible and tends to be worse calibrated on the draw specifically, since the draw is the outcome least separable by team-strength features.
The draw is genuinely the hardest of the three to predict well, because it does not correspond to one side being better. It corresponds to the difference being small, which is a statement about the distribution rather than about either team.
Why expected goals exists
Expected goals assigns each attempt a probability of scoring based on its characteristics, then sums them. Its purpose is not to be a better description of what happened, since the scoreline is what happened. It is to estimate the underlying process with less noise than a count of a rare event, which is exactly the problem low scoring creates. It is an estimate with its own error, and treating it as ground truth substitutes one noisy measure for another with more decimal places.
Heterogeneity across competitions
Soccer is played under one set of rules across a very large number of competitions with different characteristics, and that breaks the assumption that a model transfers.
Scoring rates differ by league. Home advantage differs and has shifted over time. Squad rotation patterns differ where clubs play in multiple competitions. Promotion and relegation change the composition of a league every season in a way that fixed-membership competitions do not experience. And the quality gap between the strongest and weakest teams in a division varies enormously.
The practical consequences are specific.
A model fitted on one competition should be validated on another before being trusted there, and frequently will not survive that test.
Pooling leagues to increase sample size introduces bias unless the differences are modeled explicitly, which is a real tradeoff rather than a technicality: you are trading variance for bias and should know which one is hurting you.
And team strength is not stable across a season, since squads change in transfer windows and fixture congestion affects rotation. A rating system that assumes slow drift will misprice teams immediately after a window.
What this means for evaluation
The single most useful discipline in soccer modeling is being strict about validation, precisely because the noise makes it easy to fool yourself.
Use proper scoring rules rather than accuracy. Log loss or Brier score evaluate the probabilities you actually produced, while accuracy discards the distinction between a confident correct call and a marginal one, which is the entire content of a probabilistic model.
Compare against the market baseline after removing margin, not against a naive prior. A model that beats a coin flip is not informative. A model that improves on devigged market prices is making a real claim, and most do not.
Respect time. Validate on data after the training period, never on random splits, since random splits leak future information through in-season effects.
And size the evaluation honestly. With few matches and noisy outcomes, distinguishing a genuinely better model from a lucky one takes far more data than intuition suggests, and reporting a result from a short window is the standard way apparent edges get published and then evaporate.
Parlay Ledger works at the data layer underneath this: ingesting odds across sports, normalizing them across sources, and supporting analytics and modeling on top. Getting prices, timestamps, and outcome identifiers consistent is what makes an evaluation like the one above possible at all.
Frequently asked questions
- Why is soccer harder to model than basketball?
- Scoring events are rare, so each match is effectively a small sample and the result is noisy relative to underlying quality. A basketball final margin aggregates many scoring events, while a single deflection or officiating decision can determine a soccer match regardless of which side played better.
- How should draws be handled in a model?
- Explicitly, as a genuine third outcome. Goal-based models produce it naturally as the probability of equal scores and tend to respect the underlying process. Ordered and direct multiclass approaches also work, though the draw is the hardest outcome to calibrate because it reflects a small difference rather than one side being stronger.
- What is expected goals actually for?
- Estimating the underlying scoring process with less noise than counting a rare event. It is not a better description of what happened, since the scoreline is what happened. It is an estimate with its own error, so treating it as ground truth replaces one noisy measure with another that has more decimal places.
- Can a model trained on one league be used on another?
- Usually not without validation, and often not at all. Scoring rates, home advantage, rotation patterns, and the strength spread within a division all differ by competition, and promotion and relegation change league composition each season. Pooling leagues to gain sample size trades variance for bias unless those differences are modeled.