Evaluating Movement Initiation Timing in Ultimate Frisbee via Temporal Counterfactuals

Shunsuke Iwashita, Ning Ding, Keisuke FujiiarXiv 2025 · DOI · arXiv

Introduction

In invasion sports like soccer, basketball, and Ultimate Frisbee, "off-ball" movement is the engine of offensive success. In Ultimate specifically, where the player with the disc is stationary, the timing of a teammate's "cut"—a sudden sprint into open space—is the primary factor in advancing the disc. Despite its importance, evaluating the timing of these movements has historically been a subjective exercise for coaches. This research introduces the VTCS (Valuing Timing by Counterfactual Scenarios) framework to quantify the value of movement initiation through data-driven spatial analysis.

The VTCS Framework

The VTCS framework operates by generating "what-if" scenarios to evaluate whether a player's movement was optimal. The process involves three distinct stages:

  1. Movement Initiation Detection: The system identifies the start of a "cut" using kinematic rules, such as a sudden burst of acceleration exceeding $4 m/s^2$ while the player is not in possession of the disc.
  2. Temporal Counterfactual Generation: For every detected cut, the system generates 30 alternative scenarios. These are created by shifting the initiation time forward or backward by up to one second ($\pm 15$ frames).
  3. Spatial Evaluation (wUPPCF): Scenarios are scored using a modified "Pitch Control" model called wUPPCF (Weighted Ultimate Potential Pitch Control Field). This model accounts for Ultimate-specific factors like distance weighting (penalizing long, difficult passes), marker obstruction (defensive blocking lanes), and defender reaction times.
Temporal counterfactual generation: shifting initiation earlier (a) or later (b) to evaluate alternative movement timings.

The final Timing Score ($V_{timing}$) represents the difference in spatial value between the actual play and the most advantageous counterfactual scenario.

Data and Methodology

The study utilizes UltimateTrack, the first publicly available dataset for Ultimate Frisbee providing continuous positional data. Captured via drone footage of 7v7 matches, the dataset includes over 18,000 frames of manually tracked player and disc coordinates. This high-fidelity data allows the VTCS model to simulate defensive responses and passing lanes with precision.

Key Results and Insights

The researchers validated the model by comparing spatial evaluation scores of actual target receivers against other players. A Kolmogorov–Smirnov test confirmed a significant separation ($D > 0.31, p=0$), indicating that the metric accurately reflects real-game outcomes.

The spatial value ($V_{frame}$) plotted over time for a specific cut, highlighting the peak potential advantage ($V_{scenario}$) identified by the model.

A notable case study demonstrated the model's practical utility: a receiver who initiated a cut into a congested lane (Value: 0.407) could have nearly doubled their spatial advantage (Value: 0.751) by moving just one second earlier. By shifting the timing, the player could have reached the open spot before the defender obstructed the passing lane.

Interestingly, the study revealed an "Elite Player Paradox." While novices tended to follow the model's spatial optima more closely, elite players showed a wider distribution of timing offsets. This suggests that high-level players may prioritize complex tactical goals—such as baiting defenders or timing rotations—over simple spatial maximization.

Limitations and Future Directions

While VTCS provides a robust foundation for timing analysis, it currently assumes defenders follow their original paths in counterfactual scenarios. Future iterations aim to incorporate reactive defensive models where defenders adjust to the attacker's shifted movement. Additionally, expanding the dataset to professional leagues and automating player tracking via computer vision will be essential for scaling the framework to broader applications in sports analytics.

Summary updated

BibTeX
@misc{iwashita2025evaluating,
  author = {Shunsuke Iwashita and Ning Ding and Keisuke Fujii},
  title = {{Evaluating Movement Initiation Timing in Ultimate Frisbee via Temporal Counterfactuals}},
  year = {2025},
  eprint = {2508.17611},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.17611},
}
Figure