Football Analytics: How Data Is Used to Analyse and Predict Football

What football analytics actually means

Football analytics is the use of data, statistics and mathematical models to understand what happens on the pitch and estimate what might happen next.

It has developed from relatively simple measures such as goals, results and league tables into a much broader discipline covering expected goals, team strength, probability modelling, player and team performance, simulation and statistical evaluation.

The important distinction is between describing football and modelling uncertainty in football.

A match ending 2-1 tells us what happened. Analytics can ask why it happened, whether the result reflected the balance of chances, how strong the teams were beforehand, and what the same information might imply about their next match.

That does not make football predictable in a deterministic sense. It makes uncertainty measurable.

Football Data Lab uses this approach to produce pre-match probabilities and then evaluates those probabilities against what actually happens. The purpose is not to produce tipster picks or claim certainty. It is to make a statistical estimate and then test how well that estimate performs.

At its simplest, football analytics is a process:

Football data → measurements → models → probabilities → evaluation

The starting point is data from matches. Depending on the question, that can include results, goals, shots, expected goals, team ratings, home advantage and other information.

That raw information is then transformed into measures that can be compared.

For example, a team that has scored 20 goals in ten matches has a different attacking record from a team that has scored eight. But goals alone do not tell us everything about how those results were produced.

The first team might have created a large number of high-quality chances. It might instead have scored repeatedly from unusually difficult shots. The second team might have created considerably more chances than its goal total suggests.

Analytics attempts to separate these possibilities.

The same principle applies to prediction. A model does not need to say that Team A will win. It can instead say that Team A has a 55% probability of winning, the draw has a 25% probability and Team B has a 20% probability.

That is a fundamentally different way of representing the problem.

The difference between football statistics and football analytics

Football statistics are the underlying measurements. Goals scored, shots, possession, points, expected goals and results are all statistics.

Football analytics goes a step further by using those measurements to answer a question. For example:

• How good was a team's attacking performance?
• How much of its recent success was driven by chance?
• How strong is one team relative to another?
• How likely is each possible match outcome?
• Are a model's probabilities accurate?
• Is a team consistently creating better chances than its opponents?
• Does a prediction model become more accurate as its confidence increases?

The distinction is important because a large collection of statistics is not necessarily an analysis.

Good football analytics starts with a question and uses appropriate data to answer it.

From match results to underlying performance

The final score is the most obvious piece of information from a football match, but it is also a relatively noisy one.

A team can lose despite creating better chances. A team can win despite being outplayed for long periods. A single deflection, penalty, red card or goalkeeper error can materially change a result.

That does not make the result unimportant. The result is what determines league points. It means that analysts often need to examine the process behind the result as well.

Expected goals, commonly abbreviated to xG, is one way of doing this. Rather than treating every shot as equivalent, an expected goals model estimates the probability of a shot becoming a goal.

This provides a different perspective on performance.

A 2-0 win with 0.6 expected goals may tell a different analytical story from a 2-0 win with 2.8 expected goals. The scoreline is identical. The underlying chance creation is not.

Turning performance into prediction

Describing what happened is only one part of football analytics. The more difficult problem is estimating what could happen next.

A prediction model needs some representation of team strength and attacking and defensive performance. It then needs to translate those estimates into probabilities.

One established approach is to model goals. A Poisson-based model can estimate the probability of different numbers of goals being scored by each team. Those scoreline probabilities can then be combined to calculate probabilities for home win, draw and away win.

Football Data Lab uses a Dixon-Coles approach as part of its modelling. Dixon-Coles builds on Poisson modelling while accounting for characteristics of football scores that a basic independent Poisson model does not fully capture.

The important point for a broader understanding of football analytics is not the mathematics itself. It is the transition from performance data, to estimated scoring rates, to possible scorelines, to outcome probabilities.

That is where descriptive football data becomes predictive analytics.

Team strength is another part of the picture

Goals and expected goals are not the only information that can be used to estimate future performance. A football team also has an underlying level of strength that changes over time.

Elo ratings provide one way of representing that strength. A team's rating changes according to its results relative to expectations, allowing the model to maintain an evolving estimate of how strong the team is.

This provides a different signal from simply looking at recent goals.

For example, a team that has won several matches against strong opponents may receive a different assessment from a team with the same number of wins against weaker opposition.

Football Data Lab combines different modelling signals rather than treating a single statistic as a complete description of team strength.

Why football predictions need probabilities

A common mistake in football prediction is treating a prediction as a binary statement.

If a model says that the home team has a 60% probability of winning, that does not mean the home team is expected to win every time. It means that the model considers the home win the most likely of the possible outcomes while still assigning substantial probability to the alternatives.

For example:

The home team can lose. In fact, according to the model's own probabilities, it should lose or draw 40% of the time in matches with exactly those probabilities.

This is why individual prediction results can be misleading.

A model can make a perfectly reasonable 60% prediction and still be wrong. Conversely, a model can make an extremely confident prediction that happens to be correct without that proving the underlying probability was well estimated.

The quality of a probabilistic model has to be judged over a sufficiently large collection of predictions.

Accuracy is useful, but it is not enough

The simplest way to evaluate a football prediction model is accuracy. If the model's highest-probability outcome occurs, the prediction is counted as correct.

Suppose a model makes 100 three-way match predictions and the highest-probability outcome occurs in 55 of them. Its classification accuracy is 55%.

That number is easy to understand, but it throws away much of the information contained in the probabilities.

A prediction of 51% home, 27% draw and 22% away, and one of 90% home, 6% draw and 4% away, can both count as simply correct if the home team wins. Yet they express completely different levels of confidence.

This is why serious evaluation needs proper probabilistic scoring rules as well as accuracy. Football Data Lab uses measures including the Brier score and log loss to evaluate the quality of its probabilities.

What the Brier score tells us

The Brier score measures the difference between predicted probabilities and the actual outcome.

For a three-outcome football prediction, the model assigns a probability to home win, draw and away win. Once the match is complete, those probabilities can be compared with what actually happened.

Lower scores are better.

The advantage is that the Brier score evaluates the probability distribution rather than simply asking whether the top prediction was correct.

A model that repeatedly assigns sensible probabilities can therefore be distinguished from one that makes the same number of correct classifications while being poorly calibrated.

What log loss tells us

Log loss places particular emphasis on the probability assigned to the outcome that actually occurred. This means that confident mistakes are punished heavily.

If a model gives a team a 95% probability of winning and that team loses, the error is considerably more serious than if the model had assigned the team a 45% probability.

That is useful because a prediction model should not receive the same evaluation for "I think this is slightly more likely" as for "I am extremely confident this will happen".

Log loss makes that distinction explicit.

Calibration: do the probabilities mean what they say?

Another important concept is calibration.

Suppose a model produces 100 predictions where one outcome is assigned a probability of approximately 70%. If the model is well calibrated, that outcome should occur roughly 70 times across a sufficiently large sample. Likewise, outcomes assigned approximately 20% probability should occur around 20% of the time.

Calibration therefore asks whether a model's stated confidence corresponds to reality.

It is different from accuracy. A model can correctly predict the most likely outcome reasonably often while still systematically overestimating or underestimating its probabilities.

That distinction matters particularly when predictions are being used as quantitative forecasts rather than simple winner selections.

A real example: 36 Scottish Premiership matches

The principles become much easier to understand when applied to real predictions.

Football Data Lab has published pre-match probabilities for Scottish Premiership fixtures during the 2026/27 season. The first 36 completed matches provide an early sample for evaluating how those probabilities performed.

The evaluation uses the predictions that were stored before each match kicked off.

That distinction matters. A prediction generated after a result cannot be used to honestly evaluate a pre-match forecasting system. It would potentially contain information that was unavailable before the match.

For this analysis, only prediction snapshots dated before kick-off were used. All 36 matches had a qualifying prediction snapshot.

The headline results were:

The model's Scottish Premiership predictions therefore got the most probable outcome right in 20 of the first 36 matches.

But the more informative figures are the Brier score and log loss, because they evaluate the probabilities themselves.

What the Scottish Premiership data shows

Across the 36 matches, the model made the home side favourite 22 times and the away side favourite 14 times. It did not make the draw the most likely individual outcome in any of the 36 matches.

The actual results were 15 home wins, 6 draws and 15 away wins.

The average probabilities assigned to each outcome, against the frequency with which each outcome actually occurred:

The most obvious difference is the away result. Away teams won 41.7% of the matches, compared with an average predicted probability of 33.0%.

That is an observation, not a conclusion about model quality.

Thirty-six matches is too small a sample to establish that the model systematically underestimates away wins. A gap of this size can occur through ordinary variation.

This is one of the central principles of football analytics: an interesting pattern is not automatically a statistically meaningful finding.

How model performance changed during the sample

The results were not evenly distributed throughout the opening period. The cumulative figures developed as follows:

The model began poorly, improved through August and reached 60.6% accuracy after the matches of 9 September.

Three matches on 15 September were then all incorrect. That single matchday reduced the cumulative accuracy from 60.6% to 55.6%, while the Brier score moved from 0.526 to 0.569.

This illustrates why short-term football prediction statistics should be treated cautiously. A model can move substantially in either direction after only a handful of matches.

Why confidence matters

The individual probabilities also reveal why accuracy alone is insufficient.

The strongest example in the Scottish sample came from Celtic's home matches. Football Data Lab assigned Celtic probabilities of 78.2% against Dundee, 77.4% against Falkirk and 74.1% against Aberdeen. Celtic won all three.

Those matches produced very low Brier errors because the model assigned high probability to the outcome that actually occurred.

The opposite happened with Hibernian v Kilmarnock on 15 September. Hibernian were assigned a 63.8% probability of winning. Kilmarnock won 1-0, and the away win had a probability of only 16.3%.

That result was therefore a significant miss according to the model's own probability distribution.

This is the difference between saying that a prediction was wrong and analysing how wrong the probability was.

What was the biggest surprise?

The biggest probability surprise in the first 36 Scottish Premiership matches was Hibernian 0-1 Kilmarnock. The model assigned Kilmarnock a 16.3% probability of winning.

Rangers 1-2 Hibernian was another substantial surprise, with the away win assigned a probability of 17.7%.

These are useful observations precisely because they come directly from the model's probabilities.

They do not tell us why those results happened.

An analyst should resist the temptation to invent an explanation after the fact. The model can tell us that an outcome was unlikely according to its information and assumptions. It cannot, from the probability alone, establish which specific factor caused the unexpected result.

That distinction between evidence and explanation is important in football analytics.

The draw problem

The Scottish Premiership sample also illustrates a particularly interesting modelling problem. The model never made the draw the most likely outcome in any of the first 36 matches.

That does not mean the model ignored draws. The average draw probability was 23.2%. Six of the 36 matches ended in draws, meaning the actual draw rate was 16.7%.

The issue was therefore not simply the amount of probability assigned to draws. It was the model's ability to make the draw the highest-probability outcome in a particular match.

This is a useful distinction. A model can assign meaningful probability to an outcome without ever making that outcome its top prediction.

In this sample, the draw probability never exceeded both the home and away probabilities. That creates a structural limitation for simple classification accuracy: every draw is automatically counted as an incorrect top-outcome prediction.

The Scottish data therefore provides a practical example of why analysts should look beyond a single headline accuracy figure. The wider difficulty of predicting draws is covered separately.

What this example does not prove

The Scottish Premiership sample is deliberately small.

• It does not establish the long-term quality of the model.
• It does not establish that the model is fully calibrated.
• It does not establish profitability.
• It does not establish that away teams are systematically underestimated.
• It does not establish that the model will maintain a 55.6% accuracy rate over the remainder of the season.

Those conclusions would require substantially more data.

The useful conclusion is narrower: the model can be evaluated objectively because its pre-match probabilities are stored and can be compared with subsequent results.

That is an important property of any forecasting system. Without an archived prediction made before the event, retrospective evaluation becomes much harder to trust.

Why sample size matters in football analytics

Football contains considerable variance. A single match can be determined by a small number of events, and even a good model will produce incorrect predictions.

That means performance statistics become more informative as the sample grows.

Thirty-six matches can reveal interesting patterns. Several hundred matches can provide substantially more evidence. Multiple seasons can provide more again, particularly if the model is being evaluated across different teams, competitions and periods.

This is why a serious football analytics process should not celebrate every short winning run or panic after every short losing run.

The objective is to identify persistent patterns rather than react to noise.

Football analytics is not the same as betting analysis

Prediction and betting are related but distinct subjects.

A model can be evaluated on whether its probabilities are accurate without asking whether those probabilities produce a financial return.

Brier score and log loss measure probability quality. They do not measure betting profitability.

Profitability introduces additional questions, including the available market price, bookmaker margin, staking method, liquidity, timing and whether the model's probability estimate differs sufficiently from the market.

Football Data Lab's prediction model is therefore better understood first as a probabilistic forecasting system. Its probabilities can be tested against actual outcomes independently of whether someone chooses to bet on them.

Where Monte Carlo simulation fits

Once a model has produced probabilities, those probabilities can also be used to simulate possible futures.

Monte Carlo simulation repeatedly generates outcomes according to the estimated probabilities.

Instead of asking who will win, the analysis can ask how often each outcome occurs across thousands of possible seasons.

That makes simulation useful for questions involving league positions, tournament progression and other multi-match scenarios.

The important principle remains the same: simulation does not eliminate uncertainty. It represents uncertainty repeatedly and measures the resulting distribution of outcomes.

What good football analytics should do

A useful football analytics process should have several characteristics.

It should start with a clear question

Data is only useful in relation to a question.

"How many shots did the team have?" is descriptive. "Did the team create better chances than its opponents?" is analytical. "Does that underlying performance improve its expected probability of winning the next match?" moves into prediction.

It should use appropriate data

Different questions require different measurements.

Goals may be appropriate for some questions. Expected goals may be more informative for others. Team ratings, historical results or other variables may provide additional information when estimating future outcomes.

There is no single statistic that completely describes football.

It should preserve uncertainty

A model should distinguish between a 70% probability and a certainty. They are not equivalent.

Good probabilistic analysis makes uncertainty visible rather than hiding it behind a single predicted outcome.

It should be tested before conclusions are drawn

A model should be evaluated against data that was not available when the prediction was made.

This is why the timing of the Scottish Premiership prediction snapshots matters. The prediction existed before the match. The result happened afterwards.

The two can therefore be compared without using the result to construct the prediction retrospectively.

It should distinguish observation from explanation

If a team wins five matches despite receiving lower predicted probabilities, that is an observation. Saying exactly why it happened requires additional evidence.

Analytics becomes weaker when an observed pattern is immediately turned into an unsupported explanation.

How Football Data Lab approaches football analytics

Football Data Lab's approach is built around the same distinction between prediction and evaluation.

The prediction system combines modelling approaches to estimate the probabilities of home win, draw and away win.

For the Scottish Premiership analysis above, the model used a hybrid ensemble in which a Dixon-Coles Poisson model carried 65% of the weight and an Elo system carried the remaining 35%.

The resulting scoreline distribution was converted into three-way outcome probabilities, with calibration and market-prior blending applied where appropriate.

Most importantly, the predictions are stored before matches take place. That creates an auditable historical record.

The predictions can then be compared with actual results using accuracy, Brier score, log loss and other measures. This turns the prediction system into something that can be tested rather than simply asserted to work.

You can see the current Football Data Lab predictions and follow how the model performs as more matches are completed.

The wider football analytics process

The methods discussed throughout football analytics are not isolated tricks. They form parts of a larger process.

A simplified version runs through the following stages:

• Football data
• Performance measures
• Team strength and attacking or defensive estimates
• Statistical modelling
• Scoreline probabilities
• Home, draw and away probabilities
• Predictions
• Actual results
• Brier score, log loss, accuracy and calibration
• Model evaluation and improvement

Each stage answers a different question.

The data describes what happened. The performance measures attempt to quantify it. The models turn those measurements into estimates. The probabilities express uncertainty. The evaluation compares those estimates with reality.

And the results of that evaluation provide evidence about where the model works and where it does not.

That feedback loop is the core of quantitative football analysis.

What football analytics can tell us

When enough data is available, football analytics can investigate questions such as:

• Which teams consistently create more scoring opportunities than their results suggest?
• Which teams have the strongest underlying attacking numbers?
• How does home advantage affect expected outcomes?
• How accurate are football prediction probabilities?
• Are high-confidence predictions actually more reliable?
• Are certain outcomes systematically over- or under-estimated?
• How difficult is one competition to predict compared with another?
• Does model performance remain stable over time?
• How much does a small number of unexpected results affect cumulative model metrics?

These are questions that can be tested with data. They are more useful than simply asking whether a prediction came in.

Explore the individual methods

Football analytics covers a wide range of techniques. The specialist articles below explain individual components in more detail:

• Expected Goals Explained
• Poisson Distribution Explained
• Dixon-Coles vs Poisson
• Elo Ratings Explained
• How Monte Carlo Simulations Work
• Brier Score Explained
• Log Loss Explained
• How Accurate Are Football Prediction Models?
• Football Model Calibration
• Why Draws Are Hard to Predict

For a real-world example of the process, see the Scottish Premiership prediction model performance analysis.

Frequently Asked Questions

What is football analytics?

Football analytics is the use of football data, statistics and mathematical models to analyse performance and estimate future outcomes.

What is the difference between football statistics and football analytics?

Statistics are measurements such as goals, results, shots or expected goals. Analytics uses those measurements to answer specific questions about performance, team strength, prediction and uncertainty.

What data is used in football analytics?

Depending on the question, analysts can use results, goals, shots, expected goals, team ratings, home advantage and other match information. Different models use different combinations of data.

How are football matches predicted?

Statistical models can estimate team strength and expected scoring rates before converting those estimates into probabilities for different scorelines and match outcomes.

Are football predictions certain?

No. A probability is an estimate of likelihood, not a guarantee. A team with a 70% predicted probability of winning can still lose.

How do you measure whether a football prediction model works?

Accuracy measures how often the highest-probability outcome occurs. Brier score and log loss evaluate the quality of the probabilities themselves. Calibration measures whether predicted probabilities correspond to observed frequencies.

Why is accuracy not enough?

Accuracy ignores the difference between a low-confidence prediction and a highly confident prediction. Two models can have identical accuracy while producing very different probability estimates.

Why do football models get predictions wrong?

Football contains substantial uncertainty. Individual matches can be affected by finishing variance, penalties, red cards, defensive mistakes, goalkeeper performance and many other events. A good probabilistic model is expected to make incorrect predictions.

How many matches are needed to evaluate a football prediction model?

There is no single universal threshold. More matches generally provide more evidence, while small samples can be heavily influenced by random variation. Thirty-six matches can reveal useful patterns, but they are not enough to establish long-term model performance.

What is the difference between prediction and betting?

A prediction model estimates probabilities. Betting analysis additionally requires market prices, bookmaker margins, staking assumptions and other financial considerations. A model can have useful probability forecasts without making any claim about betting profitability.

Conclusion

Football analytics is not about finding a formula that removes uncertainty from football.

It is about measuring what can be measured, modelling what can be estimated and being honest about what remains uncertain.

Expected goals can provide a better measure of chance quality than raw goals alone. Team-strength models can provide a changing estimate of relative ability. Statistical models can convert those estimates into outcome probabilities. Brier score, log loss and calibration can then test whether those probabilities correspond with reality.

The Scottish Premiership example demonstrates why this process matters.

After 36 matches, Football Data Lab's model had correctly identified the most likely result in 55.6% of matches, with a Brier score of 0.569 and log loss of 0.962. Those figures provide evidence that can be evaluated, but they do not justify a claim about long-term performance. The sample is too small, and a handful of unexpected results materially changed the numbers.

That is precisely what football analytics should make visible.

A prediction is not a promise. A result is not necessarily a complete measure of performance. And a short-term pattern is not automatically a durable finding.

The purpose of football analytics is to turn those uncertainties into quantities that can be measured, tested and improved over time.