[Paper Review] Development of an expected possession value model to analyse team attacking performances in rugby league

Thomas Sawczuk, Anna Palczewska+1 more|arXiv (Cornell University)|May 5, 2021

Sports Analytics and PerformanceEconomics, Econometrics and Finance17 references10 citations

TL;DR

This study develops an expected possession value (EPV) model for rugby league using 59,233 plays from 180 Super League matches, evaluating three zone-based models: EPV-308 (5m×5m zones), EPV-77 (10m×10m zones), and EPV-19 (aggregated variable-sized zones based on match return). The EPV-19 model demonstrated the highest reproducibility (95% similarity in attacking performance distributions over six prior matches), making it optimal for tactical analysis and team performance evaluation.

ABSTRACT

This study aimed to provide a framework to evaluate team attacking performances in rugby league using 59,233 plays from 180 Super League matches via expected possession value (EPV) models. The EPV-308 split the pitch into 308 5m x 5m zones, the EPV-77 split the pitch into 77 10m x 10m zones and the EPV-19 split the pitch in 19 zones of variable size dependent on the total zone value generated during a match. Attacking possessions were considered as Markov Chains, allowing the value of each zone visited to be estimated based on the outcome of the possession. The Kullback-Leibler Divergence was used to evaluate the reproducibility of the value generated from each zone (the reward distribution) by teams between matches. The EPV-308 had the greatest variability and lowest reproducibility, compared to EPV-77 and EPV-19. When six previous matches were considered, the team's subsequent match attacking performances had a similar reward distribution for EPV-19, EPV-77 and EPV-308 on 95 +/- 4%, 51 +/- 12% and 0 +/- 0% of occasions. This study supports the use of EPV-19 to evaluate team attacking performance in rugby league and provides a simple framework through which attacking performances can be compared between teams.

Motivation & Objective

• To develop a framework for evaluating team attacking performances in rugby league using expected possession value (EPV) models.
• To compare three EPV models with different pitch zone systems: fixed 5m×5m (EPV-308), fixed 10m×10m (EPV-77), and aggregated variable-sized zones (EPV-19) based on match return.
• To assess the reproducibility of attacking performance distributions between matches using Kullback-Leibler (KL) divergence.
• To propose a practical framework for identifying key attacking zones per team using z-score analysis relative to league averages.

Proposed method

• The pitch was divided into 308 ~5m×5m zones (EPV-308), 77 ~10m×10m zones (EPV-77), and 19 variable-sized zones (EPV-19) based on total value generated during matches.
• Attacking possessions were modeled as Markov Chains, with zone values estimated from try, conversion, penalty, and drop goal outcomes.
• Kullback-Leibler (KL) divergence was used to compare the reward distribution (match return) of each zone across matches, assessing reproducibility.
• Z-score analysis was applied to compare each team’s zone-specific match return to the league average, identifying zones of tactical dependence.
• Data from 180 Super League matches (59,233 plays) were used, with only the starting location of each play included due to data limitations.
• The model evaluated performance reproducibility over six prior matches, using the subsequent match as the true distribution for KL divergence.

Experimental results

Research questions

• RQ1 Which EPV model (EPV-308, EPV-77, or EPV-19) provides the highest reproducibility of attacking performance distributions between matches?
• RQ2 How does the number of prior matches used in performance evaluation affect the reproducibility of attacking performance patterns?
• RQ3 Can a zone aggregation method based on match return generate a more sport-specific and reproducible EPV model than fixed-size grids in rugby league?
• RQ4 To what extent can z-score analysis of zone-specific match returns identify key tactical zones for individual teams?
• RQ5 How do different zone systems affect the generalizability and interpretability of EPV models in rugby league performance analysis?

Key findings

• The EPV-19 model demonstrated the highest reproducibility, with 95 ± 4% similarity in attacking performance distributions between a team’s subsequent match and the average of its six prior matches.
• The EPV-77 model showed moderate reproducibility, with 51 ± 12% similarity in reward distributions across subsequent matches.
• The EPV-308 model had the highest variability but the lowest reproducibility, showing 0 ± 0% similarity in reward distributions across subsequent matches.
• Z-score analysis revealed that Team 4 relied heavily on wide zones (zones 3 and 4), while Team 6 focused centrally (zones 5–7), highlighting distinct tactical profiles.
• Team 9 generated value across a broad range of zones, particularly from 80–100m from its own try line, suggesting a balanced attacking approach that may contribute to high scoring.
• The study confirms that EPV-19 provides a practical, interpretable, and reproducible framework for evaluating team attacking performance in rugby league.