Real-time expected possession value estimation in ice hockey
Murto, Frans (2024)
2024
Tietojenkäsittelyopin maisteriohjelma - Master's Programme in Computer Science
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-12-02
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2024112810595
https://urn.fi/URN:NBN:fi:tuni-2024112810595
Tiivistelmä
The purpose of this thesis was to design and implement an ice hockey-specific methodology for the estimation of the expected possession value framework of any given moment of puck possession in real time. Existing metrics used to evaluate the performance of players in ice hockey are dependent on shots being taken or goals being scored, and are unable to take the context of the match properly into account for each action a player takes. Being able to quantify the value of a possession would enable creating novel metrics for player evaluation and facilitate analysis at a much higher level of granularity than was possible before.
The data used to implement the framework was provided by the Wisehockey system and consisted of the synchronized event and tracking data from 872 regular season matches of the Liiga seasons 2021--22 and 2022--23, the top-level men's ice hockey league in Finland. We framed the problem of estimating the value of a possession as a Markov decision process, and defined the action space to consist of shots, carries, passes, dump-ins, and dump-outs. We analyzed the possession sequences in our dataset and determined that a goal being scored or conceded in the next 10 seconds following an action was the optimal length for determining the future reward and risk of a possession.
We implemented the models of the expected possession value framework using XGBoost with hand-crafted features based on domain knowledge, and a graph neural network architecture with simple spatial features. Both classes of models were shown to be viable for estimating the expected possession value and produced well-calibrated probabilities. The graph neural network approach either outperformed or achieved comparable performance with the XGBoost models across all tasks except for the shot reward. The completed framework was integrated into the Wisehockey system and several new practical applications based on the expected possession value were developed for match analysis and player evaluation.
The data used to implement the framework was provided by the Wisehockey system and consisted of the synchronized event and tracking data from 872 regular season matches of the Liiga seasons 2021--22 and 2022--23, the top-level men's ice hockey league in Finland. We framed the problem of estimating the value of a possession as a Markov decision process, and defined the action space to consist of shots, carries, passes, dump-ins, and dump-outs. We analyzed the possession sequences in our dataset and determined that a goal being scored or conceded in the next 10 seconds following an action was the optimal length for determining the future reward and risk of a possession.
We implemented the models of the expected possession value framework using XGBoost with hand-crafted features based on domain knowledge, and a graph neural network architecture with simple spatial features. Both classes of models were shown to be viable for estimating the expected possession value and produced well-calibrated probabilities. The graph neural network approach either outperformed or achieved comparable performance with the XGBoost models across all tasks except for the shot reward. The completed framework was integrated into the Wisehockey system and several new practical applications based on the expected possession value were developed for match analysis and player evaluation.