Interpretable Low-Dimensional Modeling of Spatiotemporal Agent States for Decision Making in Football Tactics

Keisuke Fujii; Kenjiro Ide; Kohei Kawaguchi; Taiga Someya

REVIEW 5 major objections 6 minor 15 references

Interpretable Low-Dimensional Modeling of Spatiotemporal Agent States for Decision Making in Football Tactics

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Interpretable low-dimensional models of player states can capture football tactics, with ball-to-receiver distance and a speed-aware space score predicting pass success.

desk verdict The paper has a nice coach-driven feature idea, but its evaluation is broken: the model doesn't beat the trivial all-success baseline, so the SHAP-based claims about distance and space score are unsupported. read the letter →

arxiv 2506.16696 v1 pith:UYNOLR4M submitted 2025-06-20 cs.AI

Kenjiro Ide , Taiga Someya , Kohei Kawaguchi , Keisuke Fujii This is my paper · ORCID

classification cs.AI

keywords footballtacticspasssuccesspredictionspatiotemporaldatainterpretablemachinelearningXGBoostSHAPspacescorelow-dimensionalmodeling

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether a small set of interpretable, rule-based state variables can capture football tactics well enough to predict whether a pass succeeds. The authors define four off-ball receiver features—space score, distance to the ball, time for the nearest defender to reach the receiver, and time for the nearest defender to reach the pass line—drawing on interviews with a football manager. Using StatsBomb event data and SkillCorner tracking data from the 2023/24 LaLiga season, they train an XGBoost model on the top three receiver candidates per event. The model reaches an F1 score of about 0.56, and SHAP analysis shows that the distance between the ball and the receiver is the most influential factor, with the receiver's space score and defensive arrival times also contributing. If correct, this supports the practical claim that coaches can understand and act on passing decisions through low-dimensional, interpretable features rather than opaque black-box models.

What carries the argument

The load-bearing object is the space score: a speed-aware Voronoi region for each player, weighted by field position so that areas closer to the opponent's goal and the center of the pitch count more. The scoring of a passing option uses four off-ball receiver variables: fast space vel (the score of space to move), dist ball (distance between ball and receiver), time to player (nearest defender's arrival at the receiver), and time to passline (nearest defender's arrival at the pass line). XGBoost, a scalable tree-boosting classifier, trains on the top n receivers by distance to the ball (n=3 chosen on the coach's view that three players suffice), and SHAP (Shapley additive explanations) attributes feature importance. The mechanism works by converting raw tracking coordinates into these few human-readable quantities, so the model's predictions and importance rankings can be inspected and discussed with coaches.

What would settle it

A concrete test would be to train the same XGBoost pipeline with match-context features added (e.g., score difference, match period) or with the on-ball state variables from Section 2.3; if the feature-importance ranking of ball-to-receiver distance shifts substantially or the F1 score rises markedly, the claim that the four off-ball variables capture the decisive tactical state is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that football tactics can be captured by a low-dimensional, rule-based state representation that keeps decision-relevant information interpretable. On the paper's terms, the discovery is that a receiver's distance from the ball, together with a speed-aware Voronoi 'space score' and the arrival times of the nearest defender to either the receiver or the passing lane, carry enough signal to predict pass success in professional league data. The authors report that distance to the ball is the dominant feature in SHAP values for both n=1 and n=3 receiver-selection settings, and that space score and defensive arrival times also rank among the influential variables. This is offered as evidence that interpretable low-dimensional modeling can support tactical analysis and decision-making without the computational cost of full spatial-kinematic models or the opacity of reinforcement-learning agents.

Load-bearing premise

The whole analysis rests on the assumption that four off-ball receiver features—including a speed-aware Voronoi space score whose field weighting is not explicitly specified and a top-3 receiver selection made on the coach's opinion—are enough to represent the tactical state that determines pass success.

Editorial extensions

If this is right

If the four-feature model holds, analysts can rank passing options by distance, space, and defensive arrival time without heavy spatial-kinematic computation.
The dominance of ball-to-receiver distance in SHAP suggests that short, safe passes are the model's primary successful pattern, which coaches can use to design training around creating closer receiving options.
The relatively small influence of the space-score change on movement ('fast space vel') implies that the immediate space a receiver currently holds matters more than the space they could create by moving one meter in eight directions.
Because the variables are defined for both attacking and defending teams, the same representation can be extended to defensive risk analysis and to situations where possession is lost.
Adding the on-ball and absolute state variables described in Section 2.3 is the paper's stated next step for covering shooting phases, off-ball actions, and defensive analysis.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the field-position weighting of the space score is stated verbally and visually but never given an explicit formula; a natural next step is to estimate this weighting from data, which would test whether the manager-informed prior is actually the one that maximizes predictive power.
Editorial inference: the top-3 receiver selection is fixed by the coach's opinion; a data-driven alternative would be to let the model choose the number of plausible receivers per event, and to check whether feature-importance rankings survive that change.
Editorial inference: the same four-variable scheme could transfer to other invasion sports (basketball, handball, hockey) where time and space competition define the pass decision, provided the space weight is adapted to the scoring geometry of that sport.
Editorial inference: the paper's claim that low-dimensional rule-based models are a practical decision-support tool could be tested prospectively by asking whether coaches who see the space-score visualization make different passing decisions than those who do not.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this paper is a classic case of a decent idea undermined by a missing baseline. The authors build an interpretable XGBoost model to predict pass success from four off-ball features (distance to ball, a Voronoi-based 'space score', time to opponent, time to pass line), using LaLiga event and tracking data and features informed by a real head coach. That part is genuinely nice: the feature set is intuitive, the data are substantial, and the paper is honest about its limitations (XGBoost misses interactions, no game context).

The soft spot is load-bearing. The data have 24,416 successful and 6,325 unsuccessful passes, so the majority-class baseline is 0.794. The paper reports XGBoost accuracy of 0.685 (n=1) and 0.752 (n=3) — both below that baseline. The n=3 model is 4.2 points worse than always predicting 'success'. The paper never reports ROC-AUC, PR-AUC, calibration, or any comparison to a constant predictor. The authors do include precision/recall/F1 (F1 = 0.56 for n=3), but those numbers don't redeem a model that can't beat a trivial classifier. This isn't a minor reproducibility issue; it means the model has not demonstrated predictive skill, and the SHAP rankings in Section 3.2.1 are then not evidence that distance and space score are key factors. The central claim of the paper is unsupported as written.

I also agree with the reader that the space score formula is underspecified — we're told it's a speed-aware Voronoi area multiplied by a field-importance weight, but no explicit equation is given. The choice of n=3 is justified by 'the coach's opinion', which is fine for interpretability but not for evaluating model performance. And the key empirical finding that pass distance predicts pass success is already in Power et al. (2017), so the novelty is mainly in the feature-engineering wrapper, not the result.

One caveat: the stress-test's arithmetic assumes one instance per pass event. If the authors built a multi-instance protocol (e.g., one candidate receiver per pass), the class distribution and baseline would be different. But the paper never specifies the instance construction, so the burden is on them to clarify. As it stands, the reported numbers are consistent with a model that hasn't learned anything useful.

Who is this for? Sports analysts interested in interpretable models might find the feature idea worth exploring, but the paper is not yet a reliable source. I'd still send it to peer review — the flaw is potentially fixable with better evaluation, and the coach-driven feature approach deserves a serious look. But if the authors can't show the model beating a trivial baseline, the paper should be rejected.

Recommendation: engage with it, but push hard for baseline comparisons, clear instance definitions, and the full space-score formula. If those come through, the paper could be a modest but useful contribution.

Referee Report

5 major / 6 minor

Summary. The manuscript proposes an interpretable low-dimensional model of football tactical states. The authors define state variables for the ball-holder and potential pass receivers—including a speed-aware Voronoi 'space score,' distance to the ball, and defensive arrival times—based on interviews with a football manager. Using StatsBomb event data and SkillCorner tracking data from 200 La Liga matches, they train an XGBoost model to predict pass success and use SHAP values to identify distance and space score as key factors. The central claim is that rule-based, low-dimensional state variables can effectively capture pass-related tactics while remaining interpretable and useful for decision-making.

Significance. If the central claim were fully supported, the paper would offer a practical, interpretable alternative to computationally expensive spatial models and black-box reinforcement learning approaches, with direct value for coaches and analysts. The paper has notable strengths: it uses a large real-world dataset (200 La Liga matches), derives features from domain-expert interviews, and attempts a quantitative validation with cross-validation. However, the current evidence is undermined by an unclear evaluation protocol, the absence of a trivial baseline, and missing formal definitions for the central 'space score,' so the significance can only be assessed after these issues are addressed.

major comments (5)

[§2.1 and §3.1, Table 1] The evaluation protocol and baseline are the most serious issue. The dataset contains 24,416 successful and 6,325 unsuccessful passes, so an always-predict-success classifier achieves 79.4% accuracy. If each sample is one pass event, the reported accuracies of 68.5% (n=1) and 75.2% (n=3) are below this trivial baseline, and the model has not demonstrated predictive skill. Moreover, the reported precision, recall, and accuracy in Table 1 are internally inconsistent with those class counts under standard definitions: using the n=1 row, precision 0.54 and recall 0.55 imply about 11,400 false positives, which exceeds the total number of unsuccessful passes (6,325), so the reported metrics cannot all be correct. If instead the evaluation uses all n candidate receivers per pass event, the class counts and the labeling rule must be specified, because the majority baseline changes. The paper needs to state the exact sample construction, report the majority baseline, and add ROC-AUC, PR-AUC, and calibration metrics.
[§2.3] The space score is a central input to the model but is never defined formally. The text says only that each player's space is a Voronoi region that takes speed into account, and that the region's area is multiplied by a field-weight function that increases closer to the opponent's goal and the center. There is no equation for the speed-aware Voronoi partition, no specification of the weight function, and no statement of how offside players and goalkeepers are handled in the computation. Without these definitions, the space score cannot be reproduced, and the SHAP-based conclusion that space score is a key factor cannot be audited. Please provide the exact formulas and parameter values.
[§3.2.1] The SHAP analysis is used to support the abstract's claim that distance and space score are 'key factors in determining successful passes.' SHAP values only explain the fitted XGBoost model; they do not establish a causal or independent role of these variables in real pass outcomes. Given that the model's accuracy is at or below a trivial baseline, the feature-importance ranking is not trustworthy evidence. The authors should temper the causal language and validate the rankings with an ablation study or a simple interpretable baseline (for example, a logistic regression on distance alone), and ideally test on held-out matches.
[§2.2 and §2.5] Two methodological choices weaken the reliability of the evaluation. First, the synchronization of event and tracking data is based on a heuristic: within 50 frames around the kickoff, the kickoff is set four frames before the frame of maximum ball acceleration, with visual confirmation. This alignment step is not quantified and could propagate errors through every feature. Second, the k-fold cross-validation does not appear to group by match or team, so samples from the same game may appear in both training and test folds, inflating generalization estimates. Please report a match-grouped or temporally split evaluation.
[§2.5] The number of candidate receivers is fixed to n=1 or n=3 based on the coach's opinion, and the choice is not validated with a sensitivity analysis. Because the model input changes with n and the reported accuracy changes substantially (0.685 for n=1 vs. 0.752 for n=3), the paper should justify this hyperparameter empirically, for example by showing accuracy as a function of n and verifying that the qualitative conclusions are stable across reasonable choices.

minor comments (6)

[§2.5] The term 'repeatability' is not a standard classification metric; the table later uses 'Recall.' Please correct the terminology for consistency.
[§2.3 and §3.1] The four off-ball variables (fast space vel, dist ball, time to player, time to passline) are introduced informally; please define each with a symbol and units, and state the exact formula for 'fast space vel.'
[§2.3] The phrase 'the direction with the largest change in space score when moving 1 m in 8 directions' is ambiguous: does it mean the maximum absolute change or the maximum positive change? Please clarify.
[§3.2.2] The attached video is referenced but is not accessible in the manuscript; please include a URL or supplementary material link so that reviewers and readers can view it.
[References] The SHAP method is cited only as an arXiv preprint; please cite the peer-reviewed version (Lundberg and Lee, NeurIPS 2017) and provide complete bibliographic details.
[General] The paper does not state whether code and derived features will be released. A reproducibility statement would be helpful, especially since the raw data are licensed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the interpretable-model claims rest on an independent empirical fit, and the SHAP-based feature analysis is a standard post-hoc description of the fitted model rather than a construction-level equivalence.

full rationale

I walked the paper's derivation chain looking for places where a claimed prediction or discovery reduces by construction to its own inputs. The central pipeline is: (1) define rule-based state variables from manager interviews (space score, distance to ball, time-to-player, time-to-passline); (2) train an XGBoost classifier to predict pass success from those variables using StatsBomb and SkillCorner data; (3) report classification metrics; (4) use SHAP to identify which features the fitted model relies on. None of these steps is circular. The features (distance, space score, defensive arrival times) are measured from tracking data, and the target (pass success/failure) is an independent event label. The model is fitted to the data, and the SHAP values describe that fitted model; this is not a case of defining a variable in terms of the outcome, nor of fitting a parameter to a subset and then 'predicting' the same subset by construction. The top-n receiver selection is based on distance to the ball, which restricts the feature distribution, but the model still has to learn the mapping from all four features to the outcome; the SHAP ranking is not preordained by the selection rule. There is no load-bearing self-citation: the cited prior work from the same group (Nakahara et al., 2023) is only mentioned as future direction and is not used to justify any assumption or uniqueness theorem. The space score, while verbally defined, is not defined by an equation involving pass success, so it does not smuggle in the target. The paper's serious empirical weakness—reported accuracies (0.685 and 0.752) falling below the trivial all-success baseline (0.794)—is a validity and correctness concern, not a circularity concern; it does not indicate that the derivation is equivalent to its inputs. Therefore, no circular step can be exhibited with the required specificity, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The model relies on several hand-set parameters (n, space-weighting, Voronoi speed parameters, XGBoost hyperparameters) and domain assumptions. The space score is an invented metric without independent evidence, and the data alignment heuristic is unvalidated.

free parameters (4)

n (number of top receivers) = 3
Set to 3 based on coach's opinion and accuracy comparison; Section 2.5.
Space weight function = not fully specified
Field importance weighting is defined verbally in Section 2.3; no explicit formula given.
Speed-aware Voronoi parameters = not specified
Speed weighting in the Voronoi diagram is mentioned but not quantified.
XGBoost hyperparameters = grid-searched values not listed
Optimal hyperparameters determined via grid search, but values are not reported.

assumptions (3)

domain assumption Soccer is a sport where teams compete for time and space; this framing justifies the space score.
Section 2.3 introduces the principle without external validation.
domain assumption The event and tracking data can be aligned via kickoff detection heuristics.
Section 2.2 describes visual confirmation but no error analysis.
standard math XGBoost and SHAP are appropriate for modeling and interpreting pass success.
Used without derivation; standard tools.

invented entities (1)

Space score
purpose: Quantify the open space available to each player, used as a feature for pass success prediction.
Defined in Section 2.3; no independent validation other than the model's performance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Interpretable Low-Dimensional Modeling of Spatiotemporal Agent States for Decision Making in Football Tactics." pith.science (2026). https://pith.science/paper/UYNOLR4M

@misc{pith2026250616696,
  author       = {Pith},
  title        = {Pith review of: Interpretable Low-Dimensional Modeling of Spatiotemporal Agent States for Decision Making in Football Tactics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UYNOLR4M}},
  note         = {Machine review of arXiv:2506.16696}
}

abstract

Understanding football tactics is crucial for managers and analysts. Previous research has proposed models based on spatial and kinematic equations, but these are computationally expensive. Also, Reinforcement learning approaches use player positions and velocities but lack interpretability and require large datasets. Rule-based models align with expert knowledge but have not fully considered all players' states. This study explores whether low-dimensional, rule-based models using spatiotemporal data can effectively capture football tactics. Our approach defines interpretable state variables for both the ball-holder and potential pass receivers, based on criteria that explore options like passing. Through discussions with a manager, we identified key variables representing the game state. We then used StatsBomb event data and SkillCorner tracking data from the 2023$/$24 LaLiga season to train an XGBoost model to predict pass success. The analysis revealed that the distance between the player and the ball, as well as the player's space score, were key factors in determining successful passes. Our interpretable low-dimensional modeling facilitates tactical analysis through the use of intuitive variables and provides practical value as a tool to support decision-making in football.

Figures

Figures reproduced from arXiv: 2506.16696 by the authors.

**Figure 3.** SHAP for n = 3. The types of variables are the same as in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 2.** SHAP for n = 1. fast space vel: the score of space to move. dist ball: the distance between the ball and the offball player. time to player: the time it takes for the nearest opponent to reach the off-ball player. time to passline: the time it takes for the nearest opponent to reach the pass line. Variables of the off-ball player with the closest distance to the ball are shown [PITH_FULL_IMAGE:figures/full_fig_p00… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...
[2]

Bransen, L., Van Haaren, J., and van de Velden, M. (2019). Measuring soccer players’ contributions to chance creation by valuing their passes. Journal of Quantitative Analysis in Sports , 15(2):97--116

work page 2019
[3]

and Guestrin, C

Chen, T. and Guestrin, C. (2016). Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages 785--794

2016
[4]

and Sugihara, K

Fujimura, A. and Sugihara, K. (2005). Geometric analysis and quantitative evaluation of sport teamwork. Systems and Computers in Japan , 36(6):49--58

work page 2005
[5]

R., Kempe, M., Meerhoff, L

Goes, F. R., Kempe, M., Meerhoff, L. A., and Lemmink, K. A. (2019). Not every pass can be an assist: a data-driven model to measure pass effectiveness in professional soccer matches. Big Data , 7(1):57--70

work page 2019
[6]

and Liashchynskyi, P

Liashchynskyi, P. and Liashchynskyi, P. (2019). Grid search, random search, genetic algorithm: a big comparison for nas. arXiv preprint arXiv:1912.06059

arXiv 2019
[7]

Lucas, B. (2020). All bets are off: Preempting major league baseball’s monopoly on sports betting data. Washington Law Review , 95:1537

work page 2020
[8]

Lundberg, S. (2017). A unified approach to interpreting model predictions. arXiv preprint arXiv:1705.07874

arXiv 2017

Show all 15 references

[9]

Nakahara, H., Tsutsui, K., Takeda, K., and Fujii, K. (2023). Action valuation of on-and off-ball soccer players based on multi-agent deep reinforcement learning. IEEE Access , 11:131237--131244

2023
[10]

Power, P., Ruiz, H., Wei, X., and Lucey, P. (2017). Not all passes are created equal: Objectively measuring the risk and reward of passes in soccer from tracking data. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages...

2017
[11]

G., Berkovics, F., and Toka, L

Rahimian, P., da Silva Guerra Gomes, D. G., Berkovics, F., and Toka, L. (2022a). Let’s penetrate the defense: a machine learning model for prediction and valuation of penetrative passes. In International Workshop on Machine Learning and Data Mining for Sports Analytics , pages...

2022
[12]

Rahimian, P., Kim, H., Schmid, M., and Toka, L. (2023). Pass receiver and outcome prediction in soccer using temporal graph networks. In International Workshop on Machine Learning and Data Mining for Sports Analytics , pages 52--63. Springer

2023
[13]

Rahimian, P., Van Haaren, J., and Toka, T. A.-L. (2022b). Beyond action valuation: A deep reinforcement learning framework for optimizing player decisions in soccer. In 16th Annual MIT Sloan Sports Analytics Conference

2022
[14]

Robberechts, P., Van Roy, M., and Davis, J. (2023). un-xpass: Measuring soccer player's creativity. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 4768--4777

2023
[15]

Taki, T., Hasegawa, J.-i., and Fukumura, T. (1996). Development of motion analysis system for quantitative evaluation of teamwork in soccer games. In Proceedings of 3rd IEEE International Conference on Image Processing , volume 3, pages 815--818. IEEE

1996

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

[2] [2]

Bransen, L., Van Haaren, J., and van de Velden, M. (2019). Measuring soccer players’ contributions to chance creation by valuing their passes. Journal of Quantitative Analysis in Sports , 15(2):97--116

work page 2019

[3] [3]

and Guestrin, C

Chen, T. and Guestrin, C. (2016). Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages 785--794

2016

[4] [4]

and Sugihara, K

Fujimura, A. and Sugihara, K. (2005). Geometric analysis and quantitative evaluation of sport teamwork. Systems and Computers in Japan , 36(6):49--58

work page 2005

[5] [5]

R., Kempe, M., Meerhoff, L

Goes, F. R., Kempe, M., Meerhoff, L. A., and Lemmink, K. A. (2019). Not every pass can be an assist: a data-driven model to measure pass effectiveness in professional soccer matches. Big Data , 7(1):57--70

work page 2019

[6] [6]

and Liashchynskyi, P

Liashchynskyi, P. and Liashchynskyi, P. (2019). Grid search, random search, genetic algorithm: a big comparison for nas. arXiv preprint arXiv:1912.06059

arXiv 2019

[7] [7]

Lucas, B. (2020). All bets are off: Preempting major league baseball’s monopoly on sports betting data. Washington Law Review , 95:1537

work page 2020

[8] [8]

Lundberg, S. (2017). A unified approach to interpreting model predictions. arXiv preprint arXiv:1705.07874

arXiv 2017