cs.AIApr 13, 2026

GenTac: Generative Modeling and Forecasting of Soccer Tactics

Authors: Jiayuan Rao, Tianlin Gui, Haoning Wu, Yanfeng Wang, Weidi Xie

Organizations: Shanghai Jiao Tong University, Shanghai, China · Nanjing University, Nanjing, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Modeling open-play soccer tactics is a formidable challenge due to the stochastic, multi-agent nature of the game. Existing computational approaches typically produce single, deterministic trajectory forecasts or focus on highly structured set-pieces, fundamentally failing to capture the inherent variance and branching possibilities of real-world match evolution. Here, we introduce GenTac, a diffusion-based generative framework that conceptualizes soccer tactics as a stochastic process over continuous multi-player trajectories and discrete semantic events. By learning the underlying distribution of player movements from historical tracking data, GenTac samples diverse, plausible, long-horizon future trajectories. The framework supports rich contextual conditioning, including opponent behavior, specific team or league playing styles, and strategic objectives, while grounding continuous spatial dynamics into a 15-class tactical event space. Extensive evaluations on our proposed benchmark, TacBench, demonstrate four key capabilities: (1) GenTac achieves high geometric accuracy while strictly preserving the collective structural consistency of the team; (2) it accurately simulates stylistic nuances, distinguishing between specific teams (e.g., Auckland FC) and leagues (e.g., A-League versus German leagues); (3) it enables controllable counterfactual simulations, demonstrably altering spatial control and expected threat metrics based on offensive or defensive guidance; and (4) it reliably anticipates future tactical outcomes directly from generated rollouts. Finally, we demonstrate that GenTac can be successfully trained to generalize to other dynamic team sports, including basketball, American football, and ice hockey.

Figures & tables

Figure 1: (a) Overview of GenTac. We illustrate the two primary capabilities of the framework: multi-player trajectory forecasting and tactical event recognition. Trajectory forecasting encompasses five distinct conditioning configurations based on the number and identity of the teams being predicted. Meanwhile, tactical event recognition is evaluated under two settings: event grounding (inferred from observed data) and event forecasting (anticipated from generated futures). (b) Geometric accuracy of trajectory prediction. The x-axis represents the future time horizon (up to 5 seconds), and the y-axis reports the prediction error measured by Average Displacement Error (ADE) and Final Displacement Error (FDE) in meters. The top and bottom rows display results for unconditioned and opponent-conditioned forecasting, respectively. Lower values indicate superior geometric accuracy. (c) Collective structural consistency of forecasted trajectories. We report the absolute differences between structural metrics computed from the predicted trajectories and the ground truth over a 5-second future horizon. Circles and squares denote the mean and minimum errors, respectively. Red markers correspond to unconditioned forecasting, while blue markers indicate opponent-conditioned forecasting. The evaluated structural metrics include surface area ( Δ∣Surface Area∣\Delta|\mathrm{Surface\ Area}| ), stretch index ( Δ∣Stretch Idx.∣\Delta|\mathrm{Stretch\ Idx.}| ), team length ( Δ∣Team L.∣\Delta|\mathrm{Team\ L.}| ), team width ( Δ∣Team W.∣\Delta|\mathrm{Team\ W.}| ), Frobenius norm of pairwise distances ( Δ∣Frob. Norm∣\Delta|\mathrm{Frob.\ Norm}| ), synchronization order parameter ( Δ∣Sync. Ord.∣\Delta|\mathrm{Sync.\ Ord.}| ), and team centroid displacement ( Δ∣Cent. Disp.∣\Delta|\mathrm{Cent.\ Disp.}| ). Detailed definitions of these metrics are provided in Supplementary 12.2 .
Figure 2: (a) Performance of team- and league-conditioned trajectory forecasting. We evaluate the geometric accuracy and structural consistency of GenTac when generating collective trajectories conditioned on specific team or league identities. The results demonstrate the model’s ability to simulate the distinct playing style of a specific team (Auckland FC) over a 5-second horizon, as well as the broader stylistic differences between competitions (the Australian A-League versus German leagues) over a 1-second horizon. (b) Performance of objective-conditioned trajectory forecasting. We illustrate the deviation in tactical metrics when guiding trajectory generation toward specific strategic objectives ( offense versus defense ). The evaluated metrics quantify both attacking threat and defensive solidity: off-ball expected threat , width threat , and depth threat measure attacking intent (where higher values indicate greater offensive threat), while defensive shape disruption and defensive dominant region characterize the structural integrity and spatial control of the defending team. Detailed definitions of these metrics are provided in Supplementary 12.2 .
Figure 3: Geometric accuracy of multi-player trajectory forecasting in other team sports. Demonstrating GenTac’s ability to generalize across diverse multi-agent dynamics, we train and evaluate the framework on tracking data from basketball, American football, and ice hockey. We report the geometric accuracy of the forecasted trajectories under both unconditioned and opponent-conditioned settings, evaluated across varying historical observation windows (causal window sizes).
Figure 4: (a) Examples of tactical events with diverse semantics. The TacBench-Event dataset encompasses 15 distinct event subtypes, each capturing unique tactical behaviors on the pitch. (b) Distribution of event types and subtypes. Numerical values indicate the total count of each event subtype, while colors correspond to the broader event categories (types). (c) Overall accuracy. We report the top-1, top-3, and top-5 accuracy metrics for both event type and subtype classification. (d)-(e) Overall recall. We report the top-1, top-3, and top-5 recall metrics for both event type and subtype classification.
Figure 5: (a) Pipeline for tactical event forecasting. Given a historical multi-player trajectory, GenTac encodes the sequence and samples its future evolution KK (=20) times. Each generated future trajectory is then classified to predict a distribution over tactical events. The box plot illustrates the variance of these predictions across the KK independent rollouts, summarizing the top-5 predicted tactical events via their median probabilities, 10–90% percentile intervals (shaded boxes), and min/max ranges (whiskers). (b) Qualitative examples of trajectory and event forecasting. Given historical observations, GenTac accurately forecasts both the future spatial movements of the players and the corresponding tactical events across a 15-class action space.
MethodGS-HOTAGS-DetAGS-AssA
Playbox & MIXI58.141.381.6
Metrica-Sports58.244.476.2
KIST-GSR61.548.578.0
SoccerFactory [ 12 ]64.151.579.9
Ours (SoccerFactory + post-refinement)64.251.380.5
Supplementary Table 1 : Comparison on GSR Task according to SoccerNet 2025 Challenges Results [ 35 ] . Compared to the previous sota model SoccerFactory [ 12 ] , our refinement keeps promoting the performance of GSR task on SoccerNet-GSR [ 9 ] test set. The best and second-best results are bolded and underlined .
Supplementary Figure 1 : Manual data annotation interface. The original Chinese layout is overlaid with red English labels to highlight core components. Volunteers execute the annotation process through the following steps: Step 1: Watch tactical plays in the video playback view “1"; Step 2: Identify target sequences based on the specific task and description “3" Step 3: Utilize detailed metadata “2" to pinpoint exact start and end frames within the segment editing panel “4". To ensure dataset quality, a consistency check feature “5" allows annotators to filter out anomalous or noisy data, such as abnormal positional jumps.
Sport CategoryDatasetSourceTrainValidTestTotalUnit
SoccerMetrica Sports [ 22 ]Public Dataset10.5 *0.5 *2matches
 SkillCorner Open Data [ 20 ]Public Dataset81110matches
 DFL Bundesliga [ 21 ]Public Dataset5117matches
 SoccerNet-GSR [ 9 ]Broadcast-derived6162–123clips
 SoccerFactory+refinement [ 12 ] †Broadcast-derived–––353clips
BasketballNBA SportVU [ 23 ]Public Dataset42,05416,0468,39866,498clips
Supplementary Table 2 : Quantitative distribution of trajectory datasets. The table summarizes the allocation of data across training, validation, and test splits for all team sports. As for the "source" column, "Public Dataset" represents the data from website that has already collected in the form of pitch coordinates, while "Broadcast-derived" represents the trajectory data generated from soccer broadcast video. The temporal scope of each dataset is defined in the "Unit" column. "Matches" indicates complete games, "clips" denotes shorter tracking sequences such as specific events or fixed-length segments, and "periods" refers to the standard 20-minute intervals specific to ice hockey.
Macro-categorySubtypeTotalTrainValidTestAvg. Duration (s)
Build-upBuild47233194473.04
TransitionBall Win1,5661,0973131562.09
Progression898630179893.20
ThreatGoal3324636.32
Shot Off Target1178323115.94
Shot Saved70491472.86
Supplementary Table 3 : Quantitative distribution of event subtypes. The table summarizes the allocation of the 4,284 extracted events across training, validation, and test splits (approximating a 7:2:1 ratio). To prevent severe class imbalance at the individual match level, partitioning is executed globally across the aggregated event pool.
Supplementary Figure 2 : Visualizations of Collective Structure Consistency. Example from a progression phase (Red: attacking team, Blue: defending team). (a) Stretch index maps player distances to the team centroid (yellow cross), contrasting the defensive block’s compactness with the attacking team’s expanded shape. (b) Surface area uses convex hulls to illustrate spatial occupation, showing the attacking team utilizing greater pitch width and depth. (c) Length & width quantifies these formations: the attacking team measures 49.5349.53 m in width and 60.4460.44 m in length, compared to the defending team’s 49.3649.36 m by 40.6940.69 m. (d) Kuramoto order uses velocity arrows to visualize the Kuramoto order parameter, indicating the directional cohesion of both teams’ movements.
Supplementary Figure 3 : Visualizations of Offense and Defense Performance. Example from a deep attacking phase (Red: attacking team, Blue: defending team). (a) Off-ball expected threat maps dynamic spatial control, showing the attacking team dominating part of the penalty arc while the defending team is highly compressed. (b) Depth threat evaluates vertical penetration by segmenting the pitch into vertical strips; the intense red bands near the goal line indicate successful offensive progression into the final third, highlighting where possession yields the highest threat coming. (c) Width threat assesses lateral distribution via horizontal strips; the prominent red band along the central axis demonstrates that despite the presence of wide players, the core offensive threat is concentrated in the middle, forcing the defense to prioritize central passing lanes. (d) Defensive dominant region isolates the defending team’s spatial dominance, demonstrating their constrained control restricted primarily to the penalty box.
More in this topic

Soccer Analytics

See all 26 papers

Explore similar work

CardsList
Apr 20, 2026

TacticGen: Grounding Adaptable and Scalable Generation of Football Tactics

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/.

Sheng Xu, Guiliang Liu, Tarak Kharrat +12

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 ModelingSoccer Analytics
AskPDF
Apr 2, 2026

PlayGen-MoG: Framework for Diverse Multi-Agent Play Generation via Mixture-of-Gaussians Trajectory Prediction

Multi-agent trajectory generation in team sports requires models that capture both the diversity of possible plays and realistic spatial coordination between players on plays. Standard generative approaches such as Conditional Variational Autoencoders (CVAE) and diffusion models struggle with this task, exhibiting posterior collapse or convergence to the dataset mean. Moreover, most trajectory prediction methods operate in a forecasting regime that requires multiple frames of observed history, limiting their use for play design where only the initial formation is available. We present PlayGen-MoG, an extensible framework for formation-conditioned play generation that addresses these challenges through three design choices: 1/ a Mixture-of-Gaussians (MoG) output head with shared mixture weights across all agents, where a single set of weights selects a play scenario that couples all players' trajectories, 2/ relative spatial attention that encodes pairwise player positions and distances as learned attention biases, and 3/ non-autoregressive prediction of absolute displacements from the initial formation, eliminating cumulative error drift and removing the dependence on observed trajectory history, enabling realistic play generation from a single static formation alone. On American football tracking data, PlayGen-MoG achieves 1.68 yard ADE and 3.98 yard FDE while maintaining full utilization of all 8 mixture components with entropy of 2.06 out of 2.08, and qualitatively confirming diverse generation without mode collapse.

Kevin Song

Amazon Web Services, Seattle, Washington

Trajectory GenerationMulti-Agent Trajectory Prediction
AskPDF
Mar 16, 2026

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

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.

Miru Hong, Minho Lee, Geonhee Jo +4

University of Seoul, Seoul, Republic of Korea · Saarland University, Saarbrücken, Germany · KAIST, Daejeon, Republic of Korea +2

Autoregressive TransformerCounterfactual PredictionSports Outcome PredictionCausal Counterfactual Generation
AskPDF