Finding Expected + Unexpected Goals With R

Namita Nandakumar

@nnstats

What is hockey?

• Violent soccer on ice.�
• 31 (soon to be 32!) teams in the�NHL spend all year trying to be�good enough at this that they�win the Stanley Cup.

Why does hockey matter?

What is hockey?

• Violent soccer on ice.�
• 31 (soon to be 32!) teams in the�NHL spend all year trying to be�good enough at this that they�win the Stanley Cup.������
• Some people think it’s fun?

​

Some ways to evaluate hockey teams + players::

• Win %

• Goals For
• Goals Against

• Shots For
• Shots Against

• Feelings

Some ways to evaluate hockey teams + players::

• Win %

• Goals For
• Goals Against

• Shots For
• Shots Against

• Feelings

These have the nice statistical property of being one of the most granular things we can observe,��but people complain that not all shots are equally dangerous...

And they’re right!��So, what other features can we account for?

Let’s take a look.

A closer look...

• I randomly chose this shot from the 2019-20 NHL season.�
• gif via @NHLFlyers

What can the play-by-play data tell us?

• It was a goal.�
• It was scored by Flyers forward Travis Konecny on Kings goalie Jack Campbell.�
• It was at even strength, 5v5.�
• It was right in front of the net, ~18 feet away.�
• It was a wrist shot.�
• It caused the Flyers to go up 2-0.

What is the play-by-play data hiding?

• The shot angle looks off, which is always plausible since it’s manually tracked (for now).�
• James van Riemsdyk set up Travis Konecny with a great, quick pass.�
• Konecny was wide open.�
• There were two players screening the net.

This is where I might’ve labeled the shot.

5v5 unblocked shots have scored ~6% of the time this season…��but I think TK’s shot was better than that.

Couch to 5K xG

FACTS.

​

• It’s worth contextualizing shots with all the extra info that we have by estimating the probability of scoring.�
• Everyone with a passable understanding of hockey and R can build an expected goals model right now.�

OPINION.

​

• This is a fun way to spend a weekend afternoon.

xG Researchers, Past + Present

• Brian Macdonald (2012)�
• Dawson Sprigings,�Asmae Toumi (2015)�
• Manny Perry (2016)�
• Peter Tanner (2016)�
• Matt Barlowe (2017)�
• Micah Blake McCurdy (2018)�
• Luke and Josh Younggren (2018)�
• and many more!

• None of the ideas that follow are brand new or solely mine.�
• Except for the parts where I do nice ggplots and talk about my favorite players.

Data Source: NHL via @MoneyPuckdotcom

• Peter Tanner scrapes NHL data every day and shares season-level .csvs of every unblocked shot.�
• We’re going to use 2017-present.

Some stuff he did for us that I did not want to do::

• Label rush shots, rebounds, and more.

• Adjust x, y coordinates for recording bias.

• Clean it up so it’s ready to be #tidily analyzed.

This is all the code you need to get started!

Fun w/�Heat Maps

Unblocked Shots

Goals

ggplot(aes(arena_adjusted_x_cord_abs, arena_adjusted_y_cord)) + .

gg_rink() + .

stat_density_2d(aes(fill = ..level..), geom = 'polygon', alpha = 0.8) + .

scale_fill_viridis(option = 'D') ...

Lazy Logistic Regression xG

• Estimate P(goal) based solely on the shot’s absolute distance to the net.�
• This will be our baseline model.

glm(goal ~ poly(arena_adjusted_shot_distance, 2),� data = shots, family = 'binomial')

Slightly Less Lazy xgboost xG

• In addition to shot distance, we can incorporate shot angle, shot type, and time/distance from last event.�
• Use xgboost so as to not assume linearity.�
• Do some 5-fold CV to choose the # of boosting rounds.�
• ???�
• Profit.

Google image results for “machine learning”

Which variables am I not including?

Score State.

​

• I can’t think of a reason why scoreline would be important after controlling for everything else we already know.�
• @EvolvingWild and others have confirmed this.�

Shooters + Goalies.

​

• For simplicity, I’ll evaluate shooters and goalies by their residuals, but one could incorporate them into the model.

Some unsurprising xgboost variable importance #s.

I love using geom_smooth() to glance at calibration.

More Model Comparisons

Unexpected Goals: Shooters

shots %>%

group_by(shooter_id, shooter_name) %>%

summarize(exp_shooting_pct = mean(x_goal_xgboost),

actual_shooting_pct = mean(goal),

diff = actual_shooting_pct - exp_shooting_pct,

shot_count = n()) %>%

filter(shot_count > 300) %>%

arrange(-diff) %>%

View()

​

Thank you!!

+ thanks to @MoneyPuckdotcom, @mannyelk, @IneffectiveMath, @EvolvingWild, and all the other�researchers I’ve read + cited.��You can find me @nnstats.

Appendix: Goalies

• There’s much more spread in actual save % than expected save % once you face 500+ unblocked shots.�
• Labeled goalies deviate�> 1.5% from expectation.�
• All goalies are the same, except the bad ones?