Machine learning

Machine learning in MD1Q

Model, measure, limit.

MD1Q works with three trained models on coordinate event data. For each we state what it measures, what it learned from, how accurate it is on tournaments the training never saw, and where its limits are.

Request a demo Pricing

Independent test, AUC

Chance of scoring0.83

Chance of conceding0.85

Pass success0.95

0.5 is chance, 1.0 is a perfect ranking
Test: 2022 World Cup, Euro 2024

Three models

Action value

What it measures: How much an on-ball action (pass, carry, dribble, shot, recovery, duel) changes the chance that the team scores or concedes within the next ten actions.
Learned from: Public coordinate event data, 1,073,172 actions from three competitions.
Independent test: AUC 0.83 for scoring and 0.85 for conceding, on 263,061 actions at the 2022 World Cup and Euro 2024.
Limit: Runs only where the match has coordinate event data.

Pass success

What it measures: How likely a pass is to reach its target in the situation it was played from.
Learned from: Public 360 freeze-frame data, 98,190 passes from two tournaments.
Independent test: AUC 0.95 on 43,820 passes at Euro 2024.
Limit: The 360 frame shows only the camera's field of view: the model cannot see players outside it.

Decision quality

What it measures: The value of the chosen pass compared with the other options visible at that moment.
Learned from: Built on the pass success model and a threat-value grid computed on the same data.
Independent test: Relies on the pass success model's test (AUC 0.95).
Limit: It rates passes; carries, dribbles and shots are not options.

AI in the interface

MD1Q Proof (AI)

Ask the system and get a chart or a table. It handles 14 question types, answers only from calculated numbers, and lists the facts behind every answer.

AI text

14 text types across the interface. It may only write measured numbers, is always labelled, and can be switched off per organisation.

How to read the number

AUC measures how well the model ranks cases: 0.5 is chance, 1.0 is a perfect ranking. The test used matches from tournaments the training never saw, so the number shows how the model performs on new data. The model gives a number, not a verdict: the decision stays with the staff.