TL;DR
Monte Carlo Pass Search (MCPS) moves set-piece analysis from "what happened" to "what could have happened" by simulating thousands of counterfactual futures using a generative world model. By optimizing ball delivery, player positioning, and run timing, the researchers found that teams could increase their set-piece expected goals (xG) by approximately 28%.
What Problem Does This Paper Address?
In football (soccer) analytics, set pieces like corners and free kicks are usually evaluated based on their actual outcome—a "point estimate" of reality. This approach suffers from three major flaws:
- Outcome Bias: A brilliant delivery might be cleared by a lucky defender and scored as a "bad" play, while a poor kick might result in a goal due to a keeper's mistake.
- Lack of Counterfactuals: Traditional data only tells you about the zones players aimed at, not the empty zones that might have been more effective.
- Interaction Complexity: The value of a pass depends on how all 22 players react. You can't simply "offset" a past play by 3 meters and assume everyone would have stood still.
The authors propose a simulation-based framework to judge decisions against the full distribution of available alternatives at the moment of the kick.
Methods
The authors adapted Monte Carlo Tree Search (MCTS) into a three-part pipeline:
- The Policy: Proposes counterfactuals. It samples ~180 feasible target points for the ball and alternative "set-ups" (where attackers stand before the kick).
- The World Model: A multi-component simulator that plays the scenario forward to the first touch. It includes 3D ball physics (reconstructed from 2D tracking data), a "come-or-stay" model for the goalkeeper, and a motion-token transformer that generates realistic player trajectories based on their "intent" (destination).
- The Value Model: Scores the resulting state based on the probability of a shot occurring within 3 seconds and the xG of that shot.
A standout technical contribution is the "Causal Calibration." To prevent the simulator from over-reacting to counterfactuals (e.g., simulated players chasing a "what-if" ball more perfectly than real humans would), the authors used double machine learning to match simulated responses to real-world causal estimates of player movement.
Datasets
The study utilized a dataset of 30 Serie A 2025/26 matches involving six clubs. This included:
- StatsBomb event data and 25 Hz tracking data.
- 249 lofted deliveries (170 corners, 79 free kicks).
- Pre-training on 26,000 open-play situations to help the motion models learn realistic player movement.
Key Insights and Ideas
- The "Keeper Race": The most valuable deliveries aren't necessarily the most accurate ones; they are the ones that reach an attacker just before the goalkeeper can intervene. The "best" balls were found to be flatter and faster than observed ones.
- The Optimizer’s Curse: If you search for the "best" option using a single model, you will inevitably pick an option that the model happens to overestimate. The authors solve this by using an independent judge—an ensemble of models that didn't participate in the search—to provide an honest evaluation of the gains.
- Setup vs. Delivery: While designing clever runs adds some value (+0.0007 xG), the biggest levers are the delivery itself (+0.0085 xG) and the initial set-up of where players stand (+0.0027 xG).
- 3D from 2D: The authors successfully reconstructed 3D ball flights from tracking data that never actually measured the ball's height, by fitting physics models to the kick and the eventual contact point.
Key Results
- Global Search Gains: The "best available" delivery is worth 0.0085 xG more than the one actually kicked, a 28% improvement. Over a full season, this equates to roughly 1.4 extra xG per team.
- Combined Optimization: Searching for the best ball and the best attacker set-up together yields a total gain of +0.0108 xG per set piece.
- Reliability: In 82% of the studied set pieces, the search found a delivery significantly better than the one observed.
Limitations and Future Work
- Missing Physics: The simulation currently lacks body contact, aerial duels, and header physics. It treats the "first touch" as a probability rather than a physical collision.
- No Fouls: The model does not account for holding, blocking, or penalties, which are frequent in crowded set-piece situations.
- Counter-attack Risk: The value model only looks at the attacking team's prospects. It doesn't "price in" the risk of being caught on the break if you commit too many players forward.
- Data Needs: The authors estimate that while some components (like player-to-touch) saturate quickly, the goalkeeper and value models would benefit from several more seasons of data to reach professional-grade reliability.
Summary updated