How Machine Learning Algorithms Identify Patterns in Football Statistics

There are numerous events in just one game, and there are millions of events in a season of a whole league. It is impossible for an analyst to go through all of these events manually, nor can he analyze them in relation to many games that have been played previously. Machine learning can do this through the analysis of large amounts of data.

The current paper also adheres to the above methodology. We start by identifying the data we can analyze. Then, we consider how the algorithms learn from this data and generate output. However, just because we find a pattern doesn’t mean it will repeat in the next game.

What Football Data Can Machine Learning Analyse?

Football data comes in layers, and each layer lets algorithms see something different. The main types are:

• Results and statistics about the team: goals scored, shots, possession, and history of results – to capture the occurrences and frequency of events.
• Event-based data: a list of on-the-ball actions such as passes, shots, and tackles, along with who performed them, outcome, timing, and location.
• Positional data: the current position of the players and the ball to show where everything happened.
• Per-player data: player actions per 90 minutes, minutes played, etc.

Event data is the most common type used in football analysis, as noted in DataBallPy documentation. However, it lacks context. For example, it doesn’t show how many defenders were nearby when a shot was taken. This is where the tracking data comes into play.

Having more sophisticated inputs into the algorithm allows the system to detect more complicated patterns. The expected goals model by Opta, for example, has over 20 different variables for each shot.

How Do Machine Learning Algorithms Find Patterns in Football?

The basic process is similar for most football projects, whatever the algorithm. It moves from raw records to a tested model in four steps:

1. Collect and cleanse the data. Events need to be matched to the right players, teams, and matches. It is necessary to get rid of duplicates and other inaccuracies – for example, wrong timestamps or double passes – because small errors add up when you analyze thousands of matches.
2. Select the features. This is what the model can measure – e.g., shot distance and angle, body part used, number of passes in one action. Proper feature selection usually has more impact on your final results than the choice of algorithm.
3. Train the model. Using past examples, the model identifies which combinations of features are associated with particular outcomes, such as goals, wins, or successful passes.
4. Test it on future matches. Evaluate its performance on data not used during training (usually, from a later season).

The training data is the set of past examples the model learns from, and the algorithm is the method of learning. Opta’s xG model, for instance, uses XGBoost, which combines many small decision rules, and it was trained on nearly one million historical shots. The practical role of each method matters more than the maths, as the next two sections show.

Finding Patterns in Team and Player Performance

The simplest models link statistics to results. An xG model studies past shots to see how often each type scores. For example, a long-range header has a low score, while a close-range one-on-one gets a high score. Penalties are typically assigned a fixed xG value within a given model, often around 0.76–0.79.

VAEP, developed by researchers at KU Leuven and SciSports, evaluates on-the-ball actions by estimating how they change a team’s chances of scoring and conceding. This means that even a backward pass can have a positive value if it improves the team’s position.

Detecting Tactical and Positional Patterns

Combining tracking and event data reveals structures that basic stats hide. A 2025 study by Andrés Chacoma built passing networks from Barcelona’s 2017–18 season and found which player roles were linked to stronger passing structures.

There is other research related to line-breaking passes. Oktay Karakuş and Hasan Arkadaş developed an algorithmic model based on clusters. It analyzes line-breaking passes using 2022 World Cup data and shows how teams and individual players perform in attack.

How Are These Patterns Used in Football?

Clubs use these patterns to prepare, scout, and review. DeepMind and Liverpool’s TacticAI recommend corner setups that experts favored 90% of the time. Meanwhile, VAEP pointed out players like Kevin De Bruyne, who aren’t top scorers. At the 2026 World Cup, FIFA’s Football AI Pro analysed more than 2,000 metrics, though these tools still support decisions rather than make them.

Can Machine Learning Accurately Predict Football Results?

Models can estimate probabilities and identify tendencies, but football contains luck and many variables that data cannot capture. It helps to separate four terms that are often mixed up. The table below shows how they differ:

Football models often produce probabilistic predictions rather than definite outcomes. A 45% chance of a home win is still a prediction, but it does not guarantee that the home team will win. One arXiv study of Premier League seasons 2021–22 and 2022–23 found accuracy from about 38% to 73%, depending on the model and data, with draws hardest to predict. Its authors concluded that models alone cannot consistently beat bookmakers.

The same limit applies to games of chance, where outcomes are random by design. Slotozilla’s guide to the 10 euro deposit casino explains low-stake options and their terms for adults, but no model or pattern can remove the house edge. In football, by contrast, data does help explain performance, as long as its limits are kept in mind.

What Are the Limits of Machine Learning in Football?

Every model inherits the weaknesses of its data and of the people who use it. The main limits are:

• Inconsistent data. Data is recorded by humans, and the better leagues are recorded more extensively than the lower leagues and female soccer leagues, so something learned for one league does not necessarily apply in another.
• Changing tactics. Teams and coaches regularly adjust their approaches, so patterns learned from historical data may become less reliable over time.
• Expensive to track positions. Position tracking is done through expensive technology provided by specialized companies.
• Correlation, not causation. Possession, for example, could be an indicator of success, but only because of the team’s superior skill.

Even VAEP’s authors note that ratings are hard to compare across leagues. Machine learning works best as decision support for coaches and analysts, not as a replacement for their judgment.