1 of 20

RIPP: Holistic Player Evaluation with Region-Based Isolated Player Performance

Ethan Baron, Daniel Hocevar, Kabir Malik, Aaron White��University of Toronto Sports Analytics Student Group (UTSPAN)

2 of 20

Motivation

3 of 20

Use play-by-play data to evaluate individual player actions

4 of 20

Develop a metric to isolate individual player performance

5 of 20

Framework

Assign a value to every location on the ice: Value(x,y) = P_Goal(x,y)

Calculate the value of a player’s action: RIPP = value(state B) - value(state A)

A

B

6 of 20

7 of 20

Proof of Concept

8 of 20

Set of states and their values

9 of 20

Start and end state for each action type

10 of 20

Introduction to xG

Expected goals: probability of scoring

Use contextual data like shot distance, shot angle, etc.

Commonly implemented with logistic regression

Example: Craig Smith’s expected goal map

11 of 20

Note the relative

Importance of shot

angle and distance

12 of 20

Methodology

Visible Angle

The portion of the net the shooter can “see” from their position

Distance to first post

Width of net

= 6 feet

Distance to second post

Shooter

13 of 20

Expected Goals Model

This area is likely overrated

This area is likely underrated

14 of 20

State Transitions

15 of 20

Results

16 of 20

Results - Players

17 of 20

Results - Top 10 Players

18 of 20

Results - Marie-Philip Poulin

19 of 20

Limitations & Next Steps

Off-puck actions, defense

Tracking data

Expected goals as state value

Time interval (V-ICE)

Drive analytics in women’s hockey

Release code open-source as R package

Lastly, we’d like to discuss a few limitations of our current implementation and some promising areas for next steps.

One flaw of our current proof-of-concept is that only on-puck actions are valued, meaning we don’t reward any off-puck movement and only a limited number of defensive plays. To build a more robust player valuation model, we’d ideally like to account for all their actions on the ice, both on and off the puck. One promising idea on how to use this is by using tracking data, which is now becoming more readily available to analysts thanks to providers like Stathletes.

Secondly, we use an expected goals model for the values of the states in our framework. However, most states of a hockey game aren’t really well-measured by the immediate probability of scoring, since shots are only taken from a small part of the ice. A better state value model might instead consider the probability of scoring within in a certain time interval, as was proposed by a previous Big Data Cup submission called V-ICE.

We hope that our work can help drive forward analytics in women’s hockey. For that reason, we’re planning to release our code as an open-source R package so that other analysts can use RIPP and expand upon our research.

20 of 20

Thanks!

Ethan Baron, Daniel Hocevar, Kabir Malik, Aaron White��University of Toronto Sports Analytics Student Group (UTSPAN)