Developing a Data Driven Player Ranking in Hockey Using Predictive Model Weights

​

(Ranking Players By Passing Ability)

​

​

RITSAC

University of Toronto Hockey Analytics Club

• Student club at the University of Toronto
• OTTHAC, Ivey Sports Management Conference
• Developed a draft Model for the Mississauga Steelheads and a NHL Salary Cap Projection model

​

Credit to Original Study

• Original work completed by Joel Brooks, Matthew Kerr and John Guttag at the Massachusetts Institute of Technology.
• Developing a Data Driven Player Ranking in Soccer Using Predictive Model Weights [1]

​

Introduction

Current Playmaking Evaluations Are Unreliable

• What can player tracking data tell us about player passing ability in hockey?
• Current counting stats are flawed, not repeatable and lack context
• “Using assists to evaluate a player’s playmaking ability is ridiculous.” [2]

​

Not all assists are the same

[3],[4]

Great passes are not always counted as assists

[5]

A New Method of Evaluating Passes With Player Tracking Data

​

1. Model that relates pass origin and destination locations with their probability of resulting in a shot attempt
2. Use the resultant model weights to evaluate players passing ability

​

​

​

Methodology

Defining a Possession and a Pass

• Possessions are comprised of any sequence of events where a team retains the puck.
• Successful passes included puck movements followed by a reception and back to back receptions.
• Model was trained on possessions with a minimum of two successful passes.
• The result of a possession is categorically binary; either a shot attempt or not

Discretizing the Ice Rink

• Discretize the rink into 20 zones that we felt captured difficult and common passing origin and destination locations.

​

Feature Vectors For Origin and Destination Locations

• A vector of length 20 is calculated for the origin and destination locations of a pass.

​

• Ci = 1 if zone i is one of the two closest zones
• 0 otherwise
• max(d(l,zi),1) is the euclidean distance from the puck to the centroid of zone i

Feature Vectors For Origin-Destination Pairs

• Calculate outer product of the origin and destination vectors
• Flatten 20x20 matrix into vector of length 400
• Concatenate origin, destination and origin-destination pair vectors
• Average feature vectors of all passes in a possession

​

​

Example of a Feature Vector for A Single Pass

Outer Product

​

Concatenation

​

Origin:

Destination:

​

​

Model Building and Results

Training the Model and Results

• LinearSVC kernel classifier from python Sklearn library.
• 440 inputs, output is binary
• 80% training and 20% testing
• 10-fold cross validation on training data to find optimal hyperparameters

Model Coefficients - Top 10 Origin Zones

Model Coefficients - Top 10 Destination Zones

Model Coefficients - Top 10 Origin Destination Pairs

Model Coefficients - Bottom 10 Origin Destination Pairs

Pass Shot Value (PSV)

&

Average Pass Shot Value (APSV)

Pass Shot Value (PSV)

• Using model coefficients to evaluate a players passing ability
• Passes are assigned a PSV (pass shot value) by summing the model coefficients from the origin and destination locations along with the origin-destination pair

​

​

Where i represents the nearest zone to the pass origin

j represents the nearest zone of the destination of the pass.

PSV Example 1

[6]

PSV Example 2

Evaluating Players From the Dataset

• Top 80 players by event involvement; 52 Forwards and 28 Defensemen
• APSV calculated by averaging the PSV of all the passes completed throughout the season.

APSV Versus Traditional Assists

Utility of Model with Player Tracking Data in the NHL

• Identify playmakers that are underrated by current assist-driven evaluations.
• Finding players that fit a system or a specific passing role.
• Could be used as a proxy metric to evaluate vision and/or player IQ.

​

Improving the Model

• More data to train and test
• APSV repeatability, APSV usage in different game states
• Optimal zone discretion
• Include speed of a pass, time between passes, distance between players in the passing sequences as inputs into the model

​

Thank you!

• RIT for hosting
• ICEBERG for providing data
• Ryan and Matt for organizing

​

Yousuf LaHaye

University of Toronto,

Toronto, ON

yousuf.lahaye@mail.utoronto.ca

​

Alexander Profiti

University of Toronto

Toronto, ON

alexander.profiti@mail.utoronto.ca

References

[1] J. Brooks, M. Kerr, and J. Guttag, “Developing a Data-Driven Player Ranking in Soccer Using Predictive Model Weights,” Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD 16, 2016.

[2] R. Stimson, “Expected Primary Points are a better predictor of future scoring than Shots, Points,” Hockey Graphs, 19-Jan-2017. [Online]. Available: https://hockey-graphs.com/2017/01/19/expected-primary-points-are-a-better-predictor-of-future-scoring-than-shots-points/. [Accessed: 07-Sep-2019].

[3] N. H. L. Highlights, “Dallas Stars vs Minnesota Wild - Full Game Highlights | April 18, 2016 | NHL 2015-16 Playoffs,” YouTube, 19-Apr-2016. [Online]. Available: https://www.youtube.com/watch?v=NiC4iMx9skY. [Accessed: 07-Sep-2019].

[4] D. Highlights, “NHL Best Passes Of All Time,” YouTube, 03-Jul-2017. [Online]. Available: https://www.youtube.com/watch?v=elqcgbbxQJk. [Accessed: 07-Sep-2019].

[5] Jens95, “NHL: Open Net Fails,” YouTube, 18-Jan-2019. [Online]. Available: https://www.youtube.com/watch?v=fmGCi7Pl2iw&t=113s. [Accessed: 07-Sep-2019].

[6] N. H. L. Finland, “Auston Matthews spins, sets up Komarov for PPG,” YouTube, 04-Jan-2017. [Online]. Available: https://www.youtube.com/watch?v=oPoNw4bn6ho. [Accessed: 07-Sep-2019].