Pause the match. Ask the coach.
Fortuna Düsseldorf vs SSV Jahn Regensburg, 2. Bundesliga 2022/23, replayed from optical tracking of every player and the ball. Pause when a player has the ball: the coach ranks his passes, carries and shot in expected goals. Follow it or refuse it, then watch what the professional did. Switch to 3D at any moment: stadium camera, tactical view or a player's eyes.
The match loads when you hover over, touch or focus this panel (about 3.5 MB for the first half).
The coach's options
Faint teammates: a run that would raise a pass's value (+xG). Faint defenders: a move that would lower the coach's best option (−xG).
Model suggestion, not ground truth: in real matches only 3 % of the predicted gain was still there at the pass, because defenders react.
| Tally | Choices | Expected value |
|---|---|---|
| You | 0 | +0.000 xG |
| Coach | 0 | +0.000 xG |
| Pros | 0 | +0.000 xG |
You followed the coach 0 of 0 times · Keys: Space play / pause · ← → 5 s · F follow · 1–9 pick · V 3D view · C camera · in the 3D view: Q/E R/F look, 0 re-centre
Match report: what the players chose, what the coach preferred
For every pass of this match that could be placed on the tracking, the coach (trained on the other matches, never on this one) looked at the options the passer had and valued each in expected goals. Here is how the choices compare with its opinion. It is an opinion of a weak model, not a verdict on anyone.
The report loads when you scroll here.
Before you read the numbers
The match in numbers
| Team | Decisions | Coach's first option | Coach's top 3 | Mean gap |
|---|
Players
Only the passer is counted, one decision per pass. Press a column title to sort. The gap is how much more the coach valued its best option than the pass chosen, in expected goals (xG).
Moments to look at
Press a moment: the replay above jumps there, paused, and draws two arrows.
green: what the player chose orange: what the coach valued most
10 biggest missed opportunities
The largest gaps where the coach's option would likely have worked (chance of at least 50 %). "Missed" means the model valued another option more, not that the player made a mistake.
10 best decisions
The player chose the coach's first option, and it scored well above the typical option.
How to read it
- Arrows: the coach's top three options, the best in orange; once you choose, your option turns green. The % is the chance that the pass arrives (pass model), that the team still has the ball after a 6 m carry (pitch control), or that a shot scores.
- Value: the expected change in net expected goals (xG) over the next 15 s if the option is chosen, counting the cost of losing the ball where it would be lost.
- Tally: the sum of the coach's value for the options you, the coach and the pros chose. It is the coach's opinion of every choice, not a result.
- Player cards: under the options, a small card for the player with the ball, and for the target of a pass while you point at or focus that option. It says how the player's passes compared with what the pass model expected in the other matches of the dataset, and how many passes that is. "Slightly more than expected" is a small difference; "too few to say" means less than 20 passes or shots. If the player has no history in the other matches (the opponent in this replay plays in none), the card says so. The cards are shown for information anddo not change the advice: see the limits below.
- 3D view (button or V): the same moment in 3D; C or the camera buttons switch between the stadium camera, a tactical view from above and a player's eyes. Tap a player, or pick him in "Whose eyes", to stand at his head height (1.7 m) looking where he is running, or at the ball when he stands still. The arrows lie on the grass; the full list of options stays below the pitch.
Honest limits
- The models were trained on the 6 other matches of the dataset; this match (J03WOH) was never seen in training.
- The pass model is solid (held-out AUC 0.77–0.84). The value model is weak: it explains about 2 % of what happens in the next 15 s, barely more than "closer to goal is better", so differences of a few thousandths of a goal mean little.
- The coach plays bolder than the pros: passes it ranks first are completed 66 % of the time, against 84 % for its 7th choice and below, and it has the pro's receiver in its top 3 about as often as chance (33 % vs 30 %). It recommends a pass as its best option 92 % of the time (84 % before the cut-back fix). It gives a value opinion, not a prediction of what a player will do.
- Cut-backs (a pass from the byline back to the centre) used to be ranked in the bottom few percent, although they are followed by a goal within 15 s about one time in four. Passes now get their own "arrival" value, and on held-out matches completed cut-backs are ranked near the top (median 99th percentile in the A-League, 98th in the Bundesliga, from the 4th and 6th). The coach score now agrees with what happens next (Spearman +0.09, it was slightly negative), although the value model behind it is still weak. With only 7 Bundesliga cut-backs, treat that league's figure as indicative.
- The shot option uses our own small shot model, now trained on 663 shots from two leagues: the other Bundesliga matches and 20 A-League matches (SkillCorner open data). On held-out Bundesliga matches this lowered its log-loss from 0.325 to 0.306 and raised its AUC from 0.63 to 0.68, but that rests on 16 goals; the data provider's xG reaches 0.82.
- The player cards are for information only. In a held-out test (profiles and models built without the match being scored) they did not improve the pass advice on this league, where only 7 matches are available and few players appear in several of them; they gave a small gain on the A-League, with 20 matches; and finishing showed no signal in 160 shots. So the coach ignores them (their weight is 0), and with so little history the estimates are pulled strongly towards "as expected". They describe past matches, not this one.
- The ball carrier is the nearest player of the team in possession within 2 m of the ball (ball height is not used); speeds come from positions 0.16 s apart, as in training. "What the pro did" comes from the official events when they could be synchronised with the tracking (about two thirds of the passes), otherwise from the tracking.
- The 3D figures come from the tracking positions only: there is no body pose, so the player's-eyes camera follows his movement, not his head. The ball's height comes from the same optical tracking.
- Everything runs in your browser; your choices stay in memory and are gone when you leave.
How the coach was built and evaluated →
Data credit
Bundesliga tracking and event data: Bassek et al. (2025), DFL / Sportec Solutions, CC BY 4.0. Bassek et al. (2025), Scientific Data: paper, data; licence CC BY 4.0. Changes: positions subsampled from 25 to 12.5 frames per second and rounded to the centimetre; events synchronised to the tracking. The shot model was also fitted on shots from SkillCorner open data(MIT licence), 20 A-League 2024/25 matches; only the fitted weights are used here, none of that data is shown.