Modeling Matches as Language: A Generative Transformer Approach for Counterfactual Player Valuation in Football

Miru Hong; Minho Lee; Geonhee Jo; Hyeokje Cho; Hyunsung Kim; Pascal Bauer; Sang-Ki Ko

cs.AIMar 16, 2026

Modeling Matches as Language: A Generative Transformer Approach for Counterfactual Player Valuation in Football

Authors: Miru Hong, Minho Lee, Geonhee Jo, Hyeokje Cho, Hyunsung Kim, Pascal Bauer, Sang-Ki Ko

Organizations: University of Seoul, Seoul, Republic of Korea · Saarland University, Saarbrücken, Germany · KAIST, Daejeon, Republic of Korea · Fitogether Inc., Seoul, Republic of Korea ·

Abstract

Evaluating football player transfers is challenging because player actions depend strongly on tactical systems, teammates, and match context. Despite this complexity, recruitment decisions often rely on static statistics and subjective expert judgment, which do not fully account for these contextual factors. This limitation stems largely from the absence of counterfactual simulation mechanisms capable of predicting outcomes in hypothetical scenarios. To address these challenges, we propose ScoutGPT, a generative model that treats football match events as sequential tokens within a language modeling framework. Utilizing a NanoGPT-based Transformer architecture trained on next-token prediction, ScoutGPT learns the dynamics of match event sequences to simulate event sequences under hypothetical lineups, demonstrating superior predictive performance compared to existing baseline models. Leveraging this capability, the model employs Monte Carlo sampling to enable counterfactual simulation, allowing for the assessment of unobserved scenarios. Experiments on K League data show that simulated player transfers lead to measurable changes in offensive progression and goal probabilities, indicating that ScoutGPT captures player-specific impact beyond traditional static metrics.

Figures & tables

Figure 1: Overview of the ScoutGPT framework. Our NanoGPT-based Transformer model autoregressively predicts event tokens, enabling counterfactual ‘what-if’ simulations. For instance, replacing Kevin De Bruyne with Scott McTominay could alter actions (e.g., pass/shot) or modify the same action with a different location, outcome, or VAEP.

Group	Symbol	Description	# Tokens
Context $\mathbf{c}$	$u_{\text{H}},u_{\text{A}}$	Team IDs of the home and away teams	$2$
	$r_{k}$	Role † of each player $k$	$22$
	$p_{k}$	Player ID of each player $k$	$22$
	$\mathbf{q}$	Match state: period, minutes, home/away goals, home/away yellow and red cards	$8$
Event $\mathbf{e}_{t}$	$u_{t}$	Team ID of the acting team	$1$
	$r_{t}$	Role † of the acting player	$1$

Table 1: Components of the context block

\mathbf{c}

and each event tuple

\mathbf{e}_{t}

. Note that continuous fields (locations and elapsed time) are quantized into discrete bins.

Figure 2: Next-token prediction over the preceding event sequence. The pitch shows the final five ball-progressing actions leading up to the target, which form the preceding event sequence that we serialize into tokens and feed to the model; the star ( $\star$ ) marks the event the model predicts. At selected tokens we overlay the model’s predicted distribution, shown as pie charts for action type and outcome and histograms for the end-location coordinates $x_{t}^{\text{end}}$ and $y_{t}^{\text{end}}$ .

Split	Seasons	Matches	Episodes	Events	Events/ep	Players
Train	2021–2023	1,320	132,315	3,528,635	26.67	1,090
Valid	2024	462	43,579	1,277,169	29.31	848
Test	2025	501	47,046	1,324,363	28.15	859

Table 2: Summary of the dataset

Group	Attr.	LSTM	CatBoost	LEM- MLP	LEM- Transformer	ScoutGPT
Cat.	Team	0.94 / 0.94	0.94 / 0.94	0.66 / 0.37	0.96 / 0.49	0.92 / 0.92
	Role	0.77 / 0.67	0.76 / 0.66	0.71 / 0.32	0.45 / 0.22	0.63 / 0.47
	Type	0.76 / 0.34	0.76 / 0.33	0.71 / 0.20	0.77 / 0.43	0.78 / 0.53
	Out.	0.89 / 0.59	0.89 / 0.59	0.92 / 0.14	0.89 / 0.61	0.95 / 0.86
Cont.	Start $x$	0.74 / 2.78	0.80 / 3.88	0.91 / 2.14	0.78 / 4.59	0.96 / 0.97
	Start $y$	0.76 / 2.34	0.80 / 3.32	0.94 / 1.63	0.78 / 4.09	0.93 / 1.00

Table 3: Performance comparison across categorical and continuous attributes. For categorical attributes, Accuracy / F1 are reported; for continuous attributes,

R^{2}

/ MAE are reported.

	Goal-scoring (GS)			Goal-conceding (GC)
Method	AUC	Brier	ECE	AUC	Brier	ECE
CatBoost [ 10 ]	0.8424	0.0075	0.0003	0.8051	0.0021	0.00082
LSTM [ 16 ]	0.8294	0.0230	0.0660	0.8080	0.0029	0.01383
GRU [ 8 ]	0.8250	0.0205	0.0576	0.8091	0.0029	0.01376
Transformer [ 33 ]	0.8277	0.0210	0.0589	0.7948	0.0031	0.01519
ScoutGPT	0.8344	0.0069	0.0024	0.8153	0.0016	0.00081

Table 4: Performance comparison on goal-scoring (GS) and goal-conceding (GC) prediction within 15 seconds.

Group	Attr.	Ours	w/o Context	w/o Lineup	w/o Both
Cat.	Team	0.921 / 0.921	$-0.003$ / $-0.002$	$-0.008$ / $-0.007$	$-0.003$ / $-0.003$
	Role	0.634 / 0.469	$+0.002$ / $+0.047$	$-0.090$ / $-0.156$	$-0.069$ / $-0.128$
	Type	0.782 / 0.529	$+0.001$ / $-0.016$	$-0.002$ / $-0.006$	$-0.004$ / $-0.033$
	Out.	0.947 / 0.862	$-0.001$ / $+0.000$	$-0.001$ / $-0.001$	$+0.000$ / $+0.004$
Cont.	Start $x$	0.961 / 0.973	$-0.005$ / $+0.048$	$+0.004$ / $-0.020$	$+0.009$ / $-0.077$
	Start $y$	0.925 / 1.001	$-0.018$ / $+0.108$	$-0.029$ / $+0.164$	$-0.003$ / $+0.011$

Table 5: Ablation on lineup information and the context block. The Ours column is the full model’s absolute performance; the other columns report the change after removing each component (Accuracy / F1 for categorical attributes,

R^{2}

/ MAE for continuous). Improvement corresponds to positive changes for Accuracy, F1, and

R^{2}

, and to negative changes for MAE; green and red mark improvement and degradation.

Minute	Score state	Original	Simulated	$\Delta$
0	Drawing	0.02678	0.023477	$-$ 0.003303
	Trailing	0.02678	0.025916	$-$ 0.000864
	Leading	0.02678	0.026392	$-$ 0.000388
40	Drawing	0.02678	0.029647	$+$ 0.002867
	Trailing	0.02678	0.029864	$+$ 0.003084
	Leading	0.02678	0.029644	$+$ 0.002864

Table 6: Average VAEP under different match minute and score-state conditions. The control value corresponds to the fully controlled reference condition and is fixed at 0.02678 for all rows.

\Delta

denotes the change relative to this reference (

\Delta=\text{Simulated}-\text{Original}

Panel of Figure 3 — Figure 3: Comparison of the mean absolute delta under different numbers of samples. The left shows the per-episode mean absolute delta, while the right shows the cumulative mean absolute delta.

Model	Top-1 (%)	Top-5 (%)	Top-10 (%)
ScoutGPT (Ours)	9.20	21.97	30.90
w/o position masking	8.48	20.02	27.05
Statistics-based embedding	8.98	21.36	30.34

Table 7: Cross-season player retrieval performance using player embeddings learned from the 2021–2023 seasons. For each query embedding of a player from one season, retrieval is performed over embeddings from other seasons only. Evaluation is restricted to players who appear in at least two seasons. A query is counted as correct if an embedding of the same player from a different season is retrieved.

Figure 4: $t$ -SNE projection of ScoutGPT player embeddings from the 2024 K League season, colored by role.

Role	Player	Episode VAEP Sum			Absolute Error w.r.t. GT
		Naive	GT	Sim	\|GT $-$ Sim\|	\|GT $-$ Naive\|
Average		4.85	4.71	4.59	1.25	1.84
Left Back	Jinsu Kim	7.00	11.07	11.63	0.56	4.07
Left Wing	Reis	16.15	12.19	12.06	0.13	3.96
Center Back	Hoik Jang	3.79	4.78	5.83	1.05	0.99
Central Midfielder	Jihoon Cho	8.83	5.09	5.31	0.22	3.74

Table 8: Comparison of post-transfer episode VAEP in 2025 using naive and simulated estimates against ground truth (GT). The top 40 transferred players are selected by post-transfer minutes and used to compute the overall average. Only five players are shown for readability. Columns “Naive”, “GT”, and “Sim” report episode VAEP sums, while |GT

-

Sim| and |GT

-

Naive| report absolute errors with respect to GT.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Action Type	Count	%
On-ball	Pass	1,901,605	31.02
	Pass Received	1,742,075	28.42
	Carry	841,879	13.73
	Recovery	263,881	4.30
	Cross	70,332	1.15
	Shot	50,165	0.82

Appendix

Table S1: Action-type taxonomy (29 types), grouped into six functional categories. Count and % report the number and share of events of each type over the full dataset (6,130,167 events). Each event carries a binary outcome

a\in\{0,1\}

a=1

when the action achieves its intended objective (e.g. a pass or cross reaching a teammate, a shot resulting in a goal, a take-on or tackle retaining or regaining possession), and

a=0

otherwise. Types whose occurrence already implies their result (e.g. Carry, Recovery, Interception) are always recorded as successful, whereas terminal negative events (e.g. Foul, Error, Own Goal) are always unsuccessful.

Abbr.	Full name	Abbr.	Full name
GK	Goalkeeper	CAM	Central Attacking Midfielder
CB	Center Back	LM	Left Midfielder
LB	Left Back	RM	Right Midfielder
RB	Right Back	LW	Left Winger
LWB	Left Wing Back	RW	Right Winger
RWB	Right Wing Back	CF	Center Forward

Appendix

Table S2: List of roles used as the role tokens

r_{k}

(context) and

r_{t}

(event).

Player	Roles Played (minutes)
Jinsub Park	CB (4,383), CDM (2,081), CM (1,849)
Masatoshi Ishida	CM (2,282), CAM (1,693), RW (287), CF (237), RF (141), LM (90), LW (77), LF (67)
Sangho Na	LW (3,670), RW (1,508), LM (1,073), RM (692), CF (451), LF (405), RF (90)
Seungwon Jeong	RWB (2,355), CM (2,021), RW (450), RM (270), RB (199), CAM (90)

Appendix

Table S3: Distribution of representative multi-role players. Minutes indicate total playing time in each role.

Autoregressive Transformer

See all 29 papers

TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics
Apr 20, 2026Sheng Xu, Guiliang Liu, Tarak Kharrat +12Conditional Generative Modeling Soccer Analytics
School of Data Science, The Chinese University of Hong Kong, Shenzhen, China · Real Analytics, London, UK · Birmingham City Football Club, Birmingham, UK +3
GenTac: Generative Modeling and Forecasting of Soccer Tactics
Apr 13, 2026Jiayuan Rao, Tianlin Gui, Haoning Wu +2Soccer Analytics Multi-Agent Trajectory Prediction
Shanghai Jiao Tong University, Shanghai, China · Nanjing University, Nanjing, China
Monte Carlo Pass Search: Using Trajectory Generation for 3D Counterfactual Pass Evaluation in Football
Jun 9, 2026Andrew Kang, Priya NarasimhanCounterfactual Evaluation Multi-Agent Trajectory Prediction
Carnegie Mellon University · Pittsburgh, USA

Apr 20, 2026J/K move · Enter open · S save

TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics

Sheng Xu, Guiliang Liu, Tarak Kharrat, Yudong Luo+11

School of Data Science, The Chinese University of Hong Kong, Shenzhen, China · Real Analytics, London, UK · Birmingham City Football Club, Birmingham, UK+3

Conditional Generative Modeling Soccer Analytics

Open paper AskPDF arXiv

Success in association football relies on both individual skill and coordinated tactics. While recent advancements in spatio-temporal data and deep learning have enabled predictive analyses like trajectory forecasting, the development of tactical design remains limited. Bridging this gap is essential, as prediction reveals what is likely to occur, whereas tactic generation determines what should occur to achieve strategic objectives. In this work, we present TacticGen, a generative model for adaptable and scalable tactic generation. TacticGen formulates tactics as sequences of multi-agent movements and interactions conditioned on the game context. It employs a multi-agent diffusion transformer with agent-wise self-attention and context-aware cross-attention to capture cooperative and competitive dynamics among players and the ball. Trained with over 3.3 million events and 100 million tracking frames from top-tier leagues, TacticGen achieves state-of-the-art precision in predicting player trajectories. Building on it, TacticGen enables adaptable tactic generation tailored to diverse inference-time objectives through classifier guidance mechanism, specified via rules, natural language, or neural models. Its modeling performance is also inherently scalable. A case study with football experts confirms that TacticGen generates realistic, strategically valuable tactics, demonstrating its practical utility for tactical planning in professional football. The project page is available at: https://shengxu.net/TacticGen/.

Abstract

Autoregressive Transformer

Explore similar work

TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics

GenTac: Generative Modeling and Forecasting of Soccer Tactics

Monte Carlo Pass Search: Using Trajectory Generation for 3D Counterfactual Pass Evaluation in Football