Quintessa

Quintessa presents new sports rating algorithm at OR Society conference

Simon Rookyard presented Quintessa’s new algorithm for rating sports teams and individuals at the Operational Research Society’s annual conference, held 3rd-5th September 2019 at the University of Kent.

The Operational Research Society’s 61st annual conference was held at the University of Kent in September 2019. The conference was attended by a wide range of delegates from across industry and academia.

Simon Rookyard presented a new algorithm that Quintessa has developed to rate sports teams or individual players by ability. Rating sports teams/individuals is a complex and dynamic problem with a wide range of applications, from competition organisers and governing bodies seeding teams to ensure competitive matches (or to separate the better teams/individuals in knock-out tournaments), to sports fans and pundits predicting results for up-coming matches.

Quintessa's new “N-estimates” algorithm draws on our mathematical modelling expertise. It differs from most existing rating algorithms by acknowledging that every team/individual has good and bad days. This variability in performance is modelled specifically for each team/player, meaning the calculated ratings are both data-driven and appropriately account for how well (or otherwise) a team’s or individual’s performance can be predicted. See Figure 1. The N-estimates algorithm is versatile: With only limited changes it could be adapted to any of a wide range of team and individual sports.

Simon also presented a test case comparing the algorithm’s predictions of Premier League football results with those of a leading football-specific algorithm; the pi-rating algorithm. The N-estimates predictions are shown to be more accurate on average. See Figure 2. The difference is especially pronounced at earlier times, suggesting N-estimates is also quicker to react to new data.

Graph showing overlapping probability distributions for the ratings of two teams, illustrating how the mean and variability impact impact on the estimated difference in performance.
Figure 1: The N-estimates algorithm starts by using each match result to calculate the difference in performance between the two teams during the match. It then uses mathematical modelling and probability theory to estimate the absolute performance level of both teams. The rating of each team is updated by combining this performance estimate with other estimates from that team’s previous matches.
Line graph plotting improvement of the new algorithm over the existing algorithm for 12 seasons of Premier League football data, measured in goals. The improvement starts from zero, increasing rapidly to a peak of approx. 60 goals by 4 seasons, and then reducing slowly to approx. 10 goals by 12 seasons.
Figure 2: Comparison of N-estimates and the pre-existing “pi-rating” algorithm for Premier League football data. The cumulative difference (number of goals) between the predictions and actual match results are calculated for each algorithm; the vertical axis shows the difference between the algorithms, with positive values denoting higher accuracy for N-estimates. The new algorithm is more accurate over the 12-year dataset.