A software program designed by two computing science researchers has proven successful in predicting player performance.

When sport scouts look for young talent, they have to observe players across a number of games in order to assess their skills and techniques. However, it is not possible for the scouts to keep track of all games in the National Hockey League (NHL).

SFU professor Oliver Schulte and Zeyu Zhao, an undergraduate alumnus, developed statistical machine-learning models to combat these challenges. The assessment system is able to capture the similarity in roles and styles of hockey players.

“The overall objective of this assessment system is to be able to foretell how successful a NHL player is in securing a win for their team,” said Schulte.

The model did just that when it predicted Ottawa Senators defenseman Erik Karlsson’s key role, according to analysts, in advancing the team to the playoffs. Before the start of the 2016–17 hockey season, Schulte and Zhao released a paper in which Karlsson scored the highest of all the players.

“Karlsson’s score was 6.09 and the next best player, according to the model, [was] Sidney Crosby, [who] scored 4.475,” said Schulte. Karlsson was the only NHL player in the model to achieve a score above six.

Nearly a year later, and Karlsson is being credited with catapulting the Ottawa Senators back into the Stanley Cup playoffs.

By utilizing new game events and location data, Schulte and Zhao introduced a player performance assessment system that supports coaching, drafting, and trading decisions in the NHL. The assessment system clusters players who tend to play in similar locations into one cohort.

“Clustering players avoids apples-to-oranges comparisons, like comparing forward to defensive players,” said Schulte. Within each cluster, players are ranked according to how much their actions impact their team’s chance of scoring the next goal.  

“Basically, our model can predict how a player’s actions affect their scoring. We look at all the actions the player makes — block, hit, and pass — and assign a value to every action they make. At the end, we total up the value,” explained Schulte.

The system builds on the work of Schulte’s former master’s student Kurt Routley who devised the model for valuing player actions.

In addition, players are clustered according to the heat map concept — the ice rink is divided into 12 regions and each region is colour-coded. Players who have similar playing styles end up sharing the same heat map.

Afterwards, the model compiles players with similar heat maps into one category, which then allows for fair comparison of the players within each respective group, said Schulte.

Although the assessment system is not 100% accurate in predicting a player’s scoring ability, when Schulte looks at the top 20 players and his model’s prediction, there is usually an agreement between the media’s ranking and his. 

Currently, the assessment system does not take into account the game time and arena space. Schulte expressed that deep learning and match prediction will be the two components he will be working on for the next phase of the project.