Fig 1.
Feature importance of all features in xgboost value model.
Distance of the ball to goal is the highest importance feature followed by out-of-possession players then in-possession players closer to the defending team’s goal than the ball.
Fig 2.
Start of run player positions, 1.32% goal probability.
The still image of the start of a run, the blue dots represent the locations of the in-possession players, the red dots the out-of-possession players and the black dot the ball. The value model predicts the blue team has a 1.32% chance of scoring at this moment the blue player is about to begin his run along the highlighted trajectory.
Fig 3.
Highest value moment in run player positions, 8.63% goal probability.
The still image of near the end of a run, the blue dots represent the locations of the in-possession players, the red dots the out-of-possession players and the black dot the ball. The value model predicts the blue team has an 8.63% chance of scoring at this moment the blue player is nearing the end of his run along the highlighted trajectory. This is the highest probability of scoring the blue team has over the course of this run.
Fig 4.
End of run player positions, 4.92% goal probability.
The still image of the end of a run, the blue dots represent the locations of the in-possession players, the red dots the out-of-possession players and the black dot the ball. The value model predicts the blue team has a 4.92% chance of scoring at this moment the blue player is at the end of his run along the highlighted trajectory.
Fig 5.
Tortuosity example, player moving along path C.
In this example the player moves along the path C from point A to point B over three seconds. Tortuosity is a measure of the ratio of the length of the actual path travelled C and the length of hypothetical straight line distance L.
Table 1.
Summary of all runs in sample split by in-possession and out-of-possession, note that some runs include both in and out of possession segments and some include neither in nor out of possession segments.
Table 2.
Percent of all runs in the sample that fit into each time duration bucket.
Fig 6.
Distribution of value accrued across in-possession runs.
Distribution centred at 0 with long tails, cut off at 5th percentile (-0.05 goal probability) and 95th percentile (+0.14 goal probability).
Fig 7.
Distribution of value accrued across in-possession runs by bucket.
Value buckets are divided into negative value (V<0), positive negligible value (0 ≤ V < 0.014) and high value (V ≥ 0.014).
Fig 8.
Distribution of opponent value accrued across out-of-possession runs.
Distribution centred at 0 with long tails, cut off at 5th percentile (-0.05 opponent goal probability) and 95th percentile (+0.14 opponent goal probability).
Table 3.
Correlation table between player movement metrics and value accrued for in-possession runs and opponent value accrued for out-of-possession runs.
Speed is the most correlated with value for both in and out-of-possession runs, then tortuosity (negatively) and then acceleration. The direction of correlation for all metrics for in and out-of-possession runs is the same but the magnitude is higher for in-possession speed and acceleration, but higher for out-of-possession tortuosity.
Fig 9.
Box and whisker plot for speeds by value bucket.
Box and whisker plot (horizontal line showing median, box showing 25th and 75th percentiles, vertical line showing highest and lowest points within 1.5 of inter-quartile range) showing the difference in speed across the three bands of run value. For both in and out-of-possession runs the distribution of max speeds is skewed higher for high value runs.
Fig 10.
Box and whisker plot for accelerations by value bucket.
Box and whisker plot (horizontal line showing median, box showing 25th and 75th percentiles, vertical line showing highest and lowest points within 1.5 of inter-quartile range) showing the difference in acceleration across the three bands of run value. For both in and out-of-possession runs the distribution of max absolute accelerations is skewed slightly higher, but the difference is negligible due to the acceleration outliers (acceleration in this graph is cut off lower than -5 m·s-2 and higher than 20 m·s-2).
Fig 11.
Box and whisker plot for tortuosity by value bucket.
Box and whisker plot (horizontal line showing median, box showing 25th and 75th percentiles, vertical line showing highest and lowest points within 1.5 of inter-quartile range) showing the difference in tortuosity across the three bands of run value. For both in and out-of-possession runs the distribution of tortuosity is skewed lower for high value runs.
Fig 12.
Box and whisker plot for number of concurrent runs by value bucket.
Box and whisker plot (horizontal line showing median, box showing 25th and 75th percentiles, vertical line showing highest and lowest points within 1.5 of inter-quartile range) showing the difference in concurrent runs across the three bands of run value. Higher value runs tend to coincide with more players on the same team making high intensity runs, suggesting potential coordination.
Fig 13.
Box and whisker plot for number of high speed runs by position.
Box and whisker plot (horizontal line showing median, box showing 25th and 75th percentiles, vertical line showing highest and lowest points within 1.5 of inter-quartile range) showing the difference in number of high speed runs by position. Fullbacks, wingers and forwards make more high intensity runs in-possession, while the distributions out-of-possession are much more evenly distributed across positions.
Fig 14.
Box and whisker plot for in-possession run value buckets by position.
Box and whisker plot (horizontal line showing median, box showing 25th and 75th percentiles, vertical line showing highest and lowest points within 1.5 of inter-quartile range) showing the difference in run value by positions in-possession. Wingers and forwards have the most high value in-possession runs, but also the most negative value runs.
Fig 15.
Box and whisker plot for out-of-possession run value buckets by position.
Box and whisker plot (horizontal line showing median, box showing 25th and 75th percentiles, vertical line showing highest and lowest points within 1.5 of inter-quartile range) showing the difference in run value by positions out-of-possession. Centre backs and fullbacks make the most runs while their opponents are accruing high value.
Fig 16.
Sample player’s high intensity runs by match minute, value and speed.
Timeline of all a player’s in-possession runs over the course of a match. The colour scale shows the speed of the run and the y-axis position shows the value of the run. This shows the majority of the player’s high value runs came later in the game and their negative value runs all came earlier in the match.
Table 4.
The sample player’s run summary.
This table demonstrates the additional level of detail in a player’s run profile by using value accrued numbers to analyse the players runs rather than just reporting the total of 80 high-intensity efforts.