Valuing defense,
reproduced

A from-scratch rebuild of the research paper “Better Prevent than Tackle” (Kim et al.). It uses AI to measure something normal soccer stats miss: how much a defender is worth for the danger they quietly shut down, not just the tackles they make.

The scoreboard

Seven small AI models, each scored on match data it never trained on, and ranked by how close I got to the paper (my score ÷ the paper's score). Most learn from one match and get tested on another. A few handle rare moments like blocks and interceptions; those I retrained on 7 games from the 2022 World Cup (marked +PFF below). The paper had 564 matches to learn from and I have a handful, so this is really a test of whether my gaps came down to having less data rather than a worse method. (The “How close” column is my score divided by the paper’s: 1.00 means I matched it exactly, above 1.00 means I beat it, and below means I came up short. AUC and MRR are just accuracy-style scores: higher is better, and 1.0 is perfect.)

#	Model	Score type	Mine	Paper	How close	Status
1	c1 · Scoring soon chance the attacking team scores · GAT	AUC	0.749	~0.68	1.10	matched the paper
2	c3 · Shot quality chance an unblocked shot scores · logistic	AUC	0.785	0.775	1.01	matched the paper
3	b2 · Shot blocked chance a defender blocks the shot · GAT · +PFF	AUC	0.672	0.672	1.00	matched the paper
4	b1 · Pass completed chance the pass reaches a teammate · GAT	AUC	0.823	0.917	0.90	matched the paper
5	a1 · Next action pass, dribble, or shoot · GAT	MRR	0.670	0.800	0.84	beat the simple version
6	d1 · Ball won back which defender makes the stop · GAT · +PFF	MRR	0.436	0.694	0.63	getting closer
7	c2 · Conceding soon chance the defending team concedes · GAT · +PFF	AUC	0.48	~0.62	n/a	needs more data

How close I got

Every model's score plotted against the paper's. Land on the dashed line and I matched it exactly; above the line I did better, below it I fell short. Where I fell short, the problem is usually too little data, not a broken method, and I checked that directly: retraining the three hardest models on 7 World Cup games put b2 right on the line and nearly doubled d1 (0.272 to 0.436). Only c2 (predicting goals conceded, which happens in barely 1% of moments) still needs far more data to work.

My score vs. the paper's score

further up and to the right is better · the dashed line means a perfect match

matched the paper, or beat a simple baseline held back by too little data (improving as I add games)

Which method wins

The paper tried several AI methods and said its fanciest one (called GAT, an “attention” network) came out on top. On the two public matches I could test, it didn't. Simpler methods matched or beat it. The likely reason: the fancy method only pays off with far more data than I have. Best score in each row is in blue.

Task	Score type	GAT	GCN	GIN	XGBoost	CatBoost
b1 · Pass completed	AUC ↑	0.823	0.880	0.701	0.869	0.862
a1 · Next action	MRR ↑	0.670	0.694	0.278	n/a	n/a

A World Cup test of the paper’s main claim: defenders are paid for positioning, not for tackling

I took 7 knockout games from the 2022 World Cup, looked at 34 defenders, and paired each one with their market value (roughly what a club would pay for them, from Transfermarkt in November 2022). The question: what kind of defending does the transfer market actually pay for? I compared two different skills against each player’s value.

Winning the ball · tackles, interceptions

r = +0.02 almost no link to value

Winning the ball back more often does not make a defender worth more.

Positioning · cutting off danger early

r = +0.51 a clear positive link

Defenders who shut down danger before it develops are valued more highly.

In short, the market pays defenders for where they stand, not for how often they tackle. That is exactly the paper’s main claim. (r just measures how strongly two things move together: 0 means no link, and +1 means they rise in perfect step.)

Treat this as a hint, not proof. The models were trained on a different data source, so the exact numbers are rough, and 34 players over 7 games is tiny next to the paper’s 564 matches. The flip side of the claim, that letting danger through should count against a defender, did not show up cleanly with this little data.

Pictures from the code

Every image here is made by the actual code, running on real match data.

World Cup Final defender responsibility with real player names — WORLD CUP · **Real, named players.** Who is responsible for guarding whom, on a real frame from the **2022 World Cup Final**. Argentina moves the ball through De Paul with Álvarez as the forward threat, and France’s cover lands mostly on Varane, who is marking Álvarez, with Tchouáméni behind him. This uses a simple geometry-based method on the real tracking data, not the trained model, which would not be fair to apply to a different data source.

Prevention credit vs market value scatter — RESULT · **Stopping danger vs. market value.** How much danger each player prevents (per 90 minutes) against their market value, for 34 players across 7 World Cup games, colored by position. Defenders trend upward. Forwards like Mbappé sit low because prevention is not their job, and goalkeepers sit at zero. Overall link: r = +0.43.

EPV heatmap — FIG 1 · **Where the ball is dangerous.** Move the ball to every spot on the pitch and the value is highest right in front of goal. It tracks distance to goal almost perfectly (**+0.91**).

Defensive credit stacked bar — FIG 2 · **Defensive credit per player.** Good defending (intercepting, disrupting, deterring) goes up; danger allowed goes down; the net is the total. This recreates Figure 6 from the paper.

Tracking frame — FIG 3 · **One frame, read from the data.** The raw tracking file turned into 11-a-side positions, with the ball at a player’s feet and the keepers at each end.

Responsibility heatmap — FIG 4 · **Who gets the credit.** On a pass that gets broken up, most of the credit lands on the nearby defenders, and the player who won the ball takes 0.59 of it.

Pairwise defender-attacker credit matrix — FIG 5 · **Defender vs. attacker.** How much credit each defender earned against each attacker (blue = came out ahead). Basically, who shut down whom.

Spatial credit heatmaps by category — FIG 6 · **Where it happens on the pitch.** Danger allowed clusters near the attacking goal; interceptions and deterrence happen more through midfield.