Controlling for Context S. - PowerPoint Presentation

Slide1

Controlling for ContextS. Burtch © 2014

delta

CorsiSlide2

How have we traditionally assessed players?

Traditional Points

Plus/Minus

Faceoffs

Real Time Stats (hits, blocked shots, takeaways/giveaways)

Ice TimeSlide3

How has that changed in the Behind the Net Era? (2007 – Present)

Rate Statistics

Possession Metrics (On-Ice)

Shots For / Against

Shot Attempts For / Against(aka

Corsi

)

Unblocked Shot Attempts For / Against (aka Fenwick)

Shot Attempt Differentials

Shot Attempt PercentagesSlide4

Realizations over Time

Not all minutes/situations are created equal – we are aware of Contextual impacts resulting from usage.

We’re working within a dynamic system – teasing out impacts of individuals will never be a simplistic process.Slide5

Contextual Factors We Can Account For and Examine

Zone Starts

Time on Ice

Quality of Teammates

Quality of Competition (more on this later)

Face Off Wins/Losses

Aging

Score Effects

Time Effects (in game)

Shot Type / LocationSlide6

How Have We Examined These Things?

Eyeballing all the various components

REL Measures (On Ice - Off Ice differential) - Desjardins

WOWY Charts – Tango (via Johnson)

U

sage Charts –

Vollman

Aging Trajectories / DELTA – Tango

Heavy Lifter Index -

Poplichak

Score Adjusted Metrics –

Tulsky

Zone Start Adjusted Metrics –

Tulsky

Hextally

– Thomas / VenturaSlide7

Efforts to Reify Metrics Into Single Values For Comparison Between Players

THoR

–

Schuckers

Expected Goals –

Parkatti

/

Pfeffer

Delta SOT –

Awad

Goals Above Baseline – Thomas / Ventura

delta

Corsi

–

BurtchSlide8

SDI to delta Corsi – organic development

Need to assess defenders better

Defenders don’t seem to have a massive amount of impact on SV% or SH%

Defenders DO seem to

repeatably

impact on Shot Attempt Differentials

Initially patterned after HLI (

Poplichak

) – but specific to D men

Unhappy with index values / basis for comparison / weighting

Shifted to a regression model to predict outcomes

Multivariate Linear Regression Model to predict Expected

CorsiSlide9

Results of Regression

Corsi

For and

Corsi

Against are very weakly correlated – Offense does NOT equate to Defense at the team or individual skater level.

This means we should regress CF and CA separately

Position is Relevant (F and D impact CF/CA differently)

QoT

effects are very significant

QoC

effects are marginal to nonexistent

This aligns with the work done by

Tulsky

and Johnson

TOI is a significant correlate to results (better players tend to play more)

Zone Starts are significant

Faceoffs

matter – but less than people thinkSlide10

Results of Regression

Yearly Team Effects are very significant

This is likely an indication of team quality / systems / coaching

Age effects are hard to detect but appear to be present

CONTEXT MATTERS VERY MUCH – It Explains Over 64% of observed results for

Corsi

.Slide11

Expected vs. Observed Corsi Results from 2007 – 2014 (original Regression)Slide12

dCorsi – What Is It?

dCorsi

represents the residual (differential) between a player’s Observed

Corsi

and their Expected

Corsi

resulting from the discussed regression

.

This is an improvement on

Corsi

REL because it is determined directly from contextual factors while the player is ON the ice (the OFF Ice results aren’t weighted equivalently to the other factors).Slide13

dCorsi – Is It Meaningful or Reliable?

1 SD of

dCorsi

for NHL skaters is ± 2.0404, population mean µ = 0.019

dCorsi

is repeatable at a higher level than individual skater SH% or goaltender SV%.

Year to Year 5v5 SV% for goalies with 500+

mins

has an autocorrelation coefficient of r = 0.0646, which translates to an r

2

of 0.004171 – the prior year explains 0.4% of the following year’s result.

The r

2

year over year for Expected

Corsi

exceeds 43%, and for

dCorsi

is approximately 15%.

dCorsi

accounts for yearly team effects so players are not impacted negatively/positively by transitioning from one team to another year to year.

Impacts of Coaching Effects Can Plausibly Be Observed Year to Year.Slide14

dCorsi – Where Can We Get It?

Tableau Visualizations have been created for tracking individual skaters and to make team comparisons.

http://public.tableausoftware.com/shared/5MKNXWTX4?:display_count=yes

http://public.tableausoftware.com/shared/G6HYCG29P?:display_count=yes

THANKS!