Correlation vs Causation in Football Analytics

Correlation and causation are not the same. Correlation means that two football variables are associated: when one changes, the other tends to change as well. Causation is a stronger claim. It means that changing one variable produces a change in the other.

A team may record more possession in the matches it wins, for example, but that does not establish that possession caused those victories. Team quality, opposition strength, tactics and game state may influence both possession and results. Good football analysis therefore treats correlation as evidence of a relationship—not automatic proof of its cause.

This distinction matters whenever analysts evaluate tactics, players, recruitment decisions or predictive models. Misreading an association can turn a useful observation into an unreliable conclusion.

What Is Correlation in Football Analytics?

Correlation describes the extent to which two variables vary together. The US National Library of Medicine defines it as a statistical measure of how two variables relate to one another.

A positive correlation means that higher values of one variable tend to occur alongside higher values of another. A negative correlation means that higher values of one tend to occur alongside lower values of the other.

Possible football examples include:

• teams with higher wage bills tending to finish higher in the league;
• players taking more shots tending to score more goals;
• teams completing more final-third passes tending to create more chances; and
• teams allowing fewer shots tending to concede fewer goals.

Each relationship may contain useful information. None, by itself, proves that one variable causes the other.

How a correlation coefficient works

Pearson's correlation coefficient, commonly represented by r, measures the direction and strength of a linear relationship. It ranges from −1 to +1:

• +1: a perfect positive linear correlation;
• 0: no linear correlation; and
• −1: a perfect negative linear correlation.

A value closer to either end indicates a stronger linear association. However, the coefficient does not explain why the variables move together. It can also miss non-linear relationships and be distorted by outliers, restricted samples or poor-quality data.

What Is Causation?

Causation means that a change in one factor produces a change in another, with other relevant conditions held equal.

Suppose a club wants to know whether increasing the number of high turnovers it forces will improve its chance creation. A correlation between high turnovers and expected goals would be a useful starting point, but a causal claim would require stronger evidence that forcing those turnovers generates additional chances.

The analyst would need to investigate questions such as:

• Do the turnovers occur before the chances?
• Is there a plausible tactical mechanism connecting them?
• Does the relationship survive adjustment for team and opposition quality?
• Does it persist across competitions, seasons and match situations?
• Do tactical changes that increase high turnovers subsequently affect chance creation?

Causal inference is therefore not simply a search for a large correlation. It is an attempt to understand what would have happened under a credible alternative scenario.

Why Correlation Does Not Prove Causation

Two football variables can move together for several different reasons. Direct causation is only one possibility.

A third variable may influence both

A confounding variable is an outside factor that affects both the supposed cause and the outcome.

Consider the relationship between possession and winning. Teams recording more possession may win more frequently, but underlying team strength could help produce both outcomes. Strong teams may have better players, regain the ball more effectively, progress possession more reliably and create higher-quality chances.

The observed relationship could therefore be:

Team quality → more possession and a higher probability of winning.

This does not mean possession has no tactical value. It means the raw association cannot isolate possession's independent effect. As explained in Why Possession Percentage Can Be Misleading, analysts also need to examine territory, progression, chance creation and the circumstances in which possession occurred.

The direction of causality may be reversed

Sometimes the apparent outcome influences the apparent cause.

An analyst might notice that teams taking more shots tend to lose certain matches and conclude that shooting more often is harmful. A more plausible explanation may be that teams which fall behind attack more aggressively, take lower-quality shots and accept greater defensive exposure.

The score has changed their behaviour. Shot volume has not necessarily caused the poor result.

This is one reason game state is essential to football analysis. Teams leading, drawing and trailing do not behave in equivalent ways.

The variables may be measuring the same underlying quality

Two metrics can be strongly correlated because they capture overlapping aspects of performance.

Final-third possession and field tilt, for example, both contain information about territorial control. A relationship between them does not necessarily reveal a new causal mechanism. It may partly reflect how the variables are defined.

This can create a problem in modelling. Adding several highly related metrics may make a model appear more sophisticated without adding much independent information.

The relationship may be coincidental

Football datasets contain thousands of possible combinations of players, teams, events and performance indicators. If enough relationships are tested, some will appear important purely by chance.

Small samples increase this risk. A pattern found across five matches may disappear when the sample expands, particularly when finishing, injuries, red cards or opposition quality have shaped the original results.

This connects correlation directly to signal versus noise in football data. Before asking what caused a relationship, an analyst should first establish whether the relationship is stable enough to require an explanation.

Football Examples of Correlation Being Misread

Possession and winning

More possession can help a team control territory, move opponents and construct attacks. However, a league-level correlation between possession share and points does not prove that every team would improve by trying to dominate the ball.

Possession may reflect player quality, coaching style and financial strength. It can also be affected by the score: a leading side may voluntarily defend deeper while its opponent accumulates relatively unproductive possession.

The practical question is not simply whether possession correlates with winning. It is whether a particular team can use additional possession to create more valuable attacking situations without introducing larger defensive costs.

Pressing intensity and defensive performance

A lower passes-per-defensive-action figure may be associated with stronger teams or fewer goals conceded. It would still be unsafe to conclude that every club should press more aggressively.

Successful pressing depends on coordination, athletic capacity, defensive spacing and the consequences of the press being broken. Strong teams may be able to press intensely because they possess the necessary players and structure. Copying the visible behaviour without the underlying conditions may not reproduce the outcome.

Transfer spending and league position

Clubs that spend more on transfers often finish higher, but several explanations can operate simultaneously:

• additional spending may improve squad quality;
• successful clubs may generate more revenue and therefore spend more;
• wages, coaching and existing player quality may affect both variables; and
• recruitment efficiency may determine how much performance each pound buys.

A correlation between spending and results does not reveal the value of an individual transfer or show that spending alone created success.

Player availability and team results

A team may collect more points when a particular player starts. That does not automatically measure the player's causal contribution.

The player may be selected more often against suitable opponents, start when fully fit or appear alongside the strongest version of the team. Managers may also rest important players in matches where they expect rotation to be less costly.

Analysts must separate the player's effect from the circumstances governing selection.

Common Sources of False Causal Conclusions

How Football Analysts Investigate Causation

Observational football data rarely delivers the clean conditions of a controlled scientific experiment. Analysts cannot randomly assign Premier League teams to different wage budgets, coaches or tactical systems. They can nevertheless improve causal reasoning through careful research design.

Begin with a causal question

“Are pressing and winning correlated?” is descriptive. “Would increasing this team's pressing intensity improve its probability of winning?” is causal.

The second question identifies an intervention, a target outcome and a counterfactual: what would probably happen if the team changed its behaviour while other relevant factors remained comparable?

Identify a plausible mechanism

A causal explanation should describe the process connecting the variables.

For example:

Coordinated press → regain closer to goal → attack before the defence resets → improved chance quality.

The mechanism can then be tested at each stage. Did the tactical change increase advanced regains? Did those regains create more dangerous possessions? Did the change introduce additional vulnerability when the press was beaten?

Mechanisms do not prove causation, but they make the claim testable and help distinguish explanation from storytelling.

Control for relevant context

Analysts can adjust comparisons for factors such as:

• team and opposition strength;
• home advantage;
• score and time;
• red cards;
• competition level;
• player availability;
• fixture congestion; and
• managerial or tactical changes.

Statistical adjustment is useful only when the important variables are measured adequately. An apparently precise model can still be biased by missing information or badly defined inputs.

Compare similar situations

Analysts should seek comparisons in which the main difference is the factor being investigated.

Examples could include a tactical change within the same team, a player entering or leaving an otherwise stable line-up, or a rule change that affects some matches differently from others. These are not automatically valid natural experiments, but they may produce stronger evidence than a simple cross-sectional correlation.

Check timing

The proposed cause must precede the proposed effect. This sounds obvious, but event ordering is frequently lost when match data is aggregated into season totals.

If an attacking metric rises only after teams fall behind, a season-level association may conceal the actual sequence. Analysing possessions and match states chronologically can reveal whether the claimed mechanism is plausible.

Test whether the finding generalises

A relationship discovered in one dataset should be tested on new matches, seasons and competitions. Out-of-sample testing reduces the risk that the analyst has fitted an explanation to historical noise.

The relationship may not be universal. Tactical effects can depend on personnel, opposition and competition. A result that changes across environments is not necessarily useless, but its conditions need to be understood.

Prediction Does Not Always Require Causation

A variable can improve a forecast without being the direct cause of the outcome.

Market prices, for example, may predict match results because they aggregate information about team strength, injuries and participant expectations. The price itself does not necessarily cause the result.

Likewise, a metric may act as a proxy for a hidden characteristic that is difficult to measure directly. If the relationship remains stable, that proxy can still be useful in prediction.

The distinction depends on the decision:

• Prediction asks: does this variable help forecast what happens next?
• Causal analysis asks: what would happen if we deliberately changed this variable?

A predictive football model can perform well while offering a poor explanation of cause and effect. Conversely, a genuine causal relationship may add little predictive value if its effect is small or already reflected in other information.

Why This Matters for Recruitment and Coaching

Causal mistakes become particularly expensive when an observation is converted into an intervention.

A recruitment department may find that successful full-backs complete many progressive passes. Signing a player because he records the same volume assumes that the metric reflects a transferable ability. It may instead depend on his team's possession structure, teammates, league strength or tactical role.

A coaching staff can make the same error by imitating a successful team's visible behaviours. The relevant question is not whether the behaviour accompanies success elsewhere, but whether introducing it into this team is likely to improve performance.

This is why football clubs must combine data with context, domain knowledge and structured review. Evidence can narrow uncertainty, but it does not remove the need to understand how a decision is expected to work.

The GoalIQAI Framework for Evaluating a Causal Claim

When an article, model or analyst claims that one football factor causes another, ask:

1. Is the relationship real? Check sample size, measurement quality and repeatability.
2. Which variable came first? Establish the chronological order.
3. Could a third factor explain both? Consider team quality, game state, tactics and selection.
4. Could causality run in the opposite direction? Examine whether the outcome changes the apparent cause.
5. Is there a plausible mechanism? Explain the sequence connecting intervention and outcome.
6. Does the result survive contextual adjustment? Compare relevant situations rather than raw totals.
7. Does it generalise? Test new samples and different environments.
8. What evidence would challenge the claim? Define failure conditions before acting.

This framework does not guarantee causal certainty. It produces a more disciplined assessment of how much confidence the available evidence deserves.

Key Takeaways

• Correlation means that two variables move together; causation means that changing one produces a change in the other.
• A strong or statistically significant correlation does not, by itself, establish cause and effect.
• Team quality, tactics, opponent strength, selection and game state can create misleading football correlations.
• Reverse causality is common because match outcomes change how teams behave.
• Causal analysis requires a credible mechanism, correct event ordering, contextual adjustment and evidence that survives new samples.
• A non-causal variable can still be useful for prediction, but it may be dangerous as the basis for a tactical or recruitment intervention.
• The correct conclusion is not that correlations are worthless. It is that they are the beginning of an investigation rather than its end.

Related Guides

• Signal vs Noise in Football Data
• Variance in Football Betting Explained
• Regression to the Mean in Football Explained
• Game State in Football Analytics Explained
• Why Possession Percentage Can Be Misleading
• Football Betting & Analytics Knowledge Base

Stay Ahead of the Market

GoalIQAI explains the models, metrics and decision-making principles behind evidence-led football analysis. Subscribe for new guides, match analysis and research delivered directly to your inbox.