The problem with Corsi
It's no secret. There's a problem with Corsi. It's not news either. The analysts have been aware of it's limitations since the very beginning: Corsi doesn't account for context. Perhaps this problem was well, although coarsely, described by former Edmonton Oilers' head coach Dallas Eakins when he somewhat infamously stated:
An individual does not have a Corsi. His Corsi is a whole bunch of players that are with him, right.
Make no mistake about it, Eakins was right, but his statement was also ignorant. Analysts have struggled with this problem in many different ways, and have proposed several tools for dealing with it. Individual players do not have a Corsi but they do have a dCorsi.
A few weeks ago I reintroduced Copper & Blue readers to Stephen Burtch's dCorsi metric and used it to analyze the Oilers' defence. Subsequent to that analysis I took a hard look at the Taylor vs Tyler debate using dCorsi. During the few days that have trickled along since McDavid day our defence has once again become a popular topic of conversation. The purpose of this post is to bring dCorsi into that conversation, and to explore the dCorsi Impact metrics.
The "d" in dCorsi stands for "delta". Although the measure itself is complicated and involves intense calculation, the concept is simple. dCorsi accounts for the contexts that influence Corsi: it provides a measure of independent performance relative to other players in the league.
The dCorsi process first involves the calculation of Expected Corsi scores against which Observed scores are then compared. Expected scores are, for the sake of simplification, average references. Observed scores are regular Corsis, the ones Eakins says individuals don't have, attempted shots that happen on the ice. The delta measure is subsequently calculated as the difference between the Observed and Expected measures .
Although it has yet to achieve universal acclaim within the analytics community dCorsi appears to be a more than adequate tool for comparing individual contributions to possession within teams, across teams, across seasons, and eventually perhaps across leagues and eras. While I'm certain it's not perfect, and I would like too see an extended peer-review of its regression parameters, I believe that Burtch has done something very special and important with this measure. If dCorsi does what it claims to do, and at this point I believe it does, it's useful for grading individual performances regardless of context: it puts the numbers on a level playing field and properly assigns a Corsi number to individuals.
A dCorsi tells us very little about how a player achieved good or bad possession play and nothing about the context that led to it. Context is best investigated through other means such as usage charts an line-mate analysis, the tools that many analysts have developed to account for the known and established issues of Corsi measures. If you want to know about context, don't look to dCorsi. All of that context information is under the hood hidden in Burtch's magnificent engine but the numbers themselves tell us little about context. By comparing dCorsi scores across time and by comparing Expected measures we might be able to glean some of this contextual information but generally speaking what dCorsi tells us about is the player and how the player compares to his colleagues.
Although it doesn't present us with context, dCorsi does separate For and Against measures and such separation is critical. It's various measures tell us that there are several ways of being a possession player. Some players achieve positive scores by moving pucks while others are better at blocking them. The "Impact" measures combine time-on-ice with delta Corsi in order to give us an idea of the influence a player has on the ice.
dCorsi Against Impact (dCAI) is the difference between a player's Observed and Expected Against scores, multiplied by average five-on-five time-on-ice per 60 minutes of play. Negative dCAI is good, and means that a player suppressed attempted Corsis relative to his colleagues. This metric is shown on the horizontal axis for the visualizations in the following section.
dCorsi For Impact (dCFI) is the difference between a player's Observed and Expected Corsi For scores, multiplied by average five-on-five time-on-ice per 60 minutes of play. Positive dCFI is good. dCFI is a relative measure of offence (not to be confused with scoring). This metric is shown on the vertical axis for the visualizations in the following section.
dCorsi Impact (dCI) is the difference between a player's dCAI and dCFI. dCI is an overall relative measure of a players effect of on-ice possession. This metric corresponds to bubble size for the visualizations in the following section.