Model-based scouting: two thousand data points per match and the comparability problem

Data availability is the starting point. A Chinese review of big data in football reports that a single professional match generates over two thousand data points covering running, passing, and physical contact. Machine learning models combine multiple sources to estimate a player's future form. This is no longer a record of a single event; it is a stream of measurable quantities.

The second pillar is expected metrics. An expected goals (xG) model assigns a probability of scoring to every shot. Summing these values shows how many goals a team should have scored from the chances it created, regardless of the actual tally.

The same shot, different numbers

Here lies the problem analysts know well. Modern xG models differ in their feature selection: the body part used for the shot, the type of preceding action, defender pressure, and position. These choices lead to systematically different estimates for the same shots. Consequently, xG values from different providers are not directly comparable, especially when one model carries contextual information the other lacks.

The practical conclusion is a warning for any club comparing players across two leagues: the same footballer's goals minus xG can shift when the data provider changes. With short samples, the gap is often noise. It reflects a model's choices and biases, not a player's ability.

A concentrated market of providers

The market for data providers is concentrated and worth billions. Opta, founded in 1996, supplies data for more than 30 sports in more than 70 countries and has belonged to Stats Perform since 2019. Genius Sports, listed on the New York Stock Exchange, employs around 1 500 people and serves more than 300 betting brands, delivering data and broadcasts from more than 240 000 sporting events a year.

Consolidation continues. In February 2026 Genius Sports announced it would buy Legend, the owner of the Casino.org and Covers.com brands, for 1.2 billion dollars. The company's shares fell 23 percent in response, which shows how nervously the market now prices purchases in this segment.

A club no longer buys a single model. It buys a provider, its event definitions and its way of counting, and with all of that, its errors.

For model-based scouting, then, the question is not whether to use data but whose data, and whether we are comparing the same thing. The answer is no, not until the provider and the definition of the metric are settled before two names are set side by side.