Corsi/Fenwick/Shooting Efficiency as future predictors

sanitysrequiem

Registered User
Nov 14, 2009
1,426
115
Hello all

Please bear with me, it's my first post in this sub-forum and I don't really have a strong background in statistics. But I was just thinking about something and was wondering if an analysis such as the one I will describe below has been performed. And if not, whether it would be useful to do it.


Background

So from my understanding, we have the following:

Corsi (for/against) - includes shots, missed shots, blocked shots
Fenwick (for/against) - includes shots, missed shots
Shots (for against) - self-explanatory, only shots that hit the net
Goals (for/against)
Wins (based on differential between goals for and against)

Analysis


With the above noted, we can see that as we decrease the granularity, we are removing one variable at each step (i.e. Fenwick removes blocked shots from Corsi, shots removes missed shots from Fenwick, etc).

So what I am wondering is, has there been an analysis of the efficiency percentages of these stats? See below for an example - let's say only for events FOR.

Fenwick efficiency % = Fenwick FOR / Corsi FOR
Shot efficiency % = Shots FOR / Fenwick FOR
Shooting % = Goals FOR / Shots for (yes I know this stat is tracked already)

Potential Use

Once we have the efficiency percentages above, it would allow us to disaggregate between the different levels of puck possession and shooting which lead to a goal. If you look at one team, you could see how much posession they generate, what percentage of puck possession results in shots directed to the net, what percentage of those events actually hit the net, and what percentage of those go in for goals.

As it relates to predicting future performance, this is where the statistics bit would come in, that I am not qualified for. I'm assuming you would have to run a regression/analysis on each of the metrics and see their correlation and consistency over time?

I think the toughest part would be to determine how much the metrics correlate closest to scoring goals or keeping them out? And how much of it is pure randomness? Also I guess how they correlate with each other. But if we can figure out what "stabilised" values are for each metric (is that possible?) then I think it could be a great indicator of future performance.


I'm not sure about this part. But what spurred my thought was, if you just use Corsi close 5v5 as a predictor for future results, I think it is flawed, as it doesn't give credit for quality and efficiency of scoring, just possession. If you look at Pittsburgh last year they were 18th in Corsi For %, but because they have such talented and creative players, they were more effective in making the most of the possession they have. And I think it's clear that they are better than 18th in the league.
 

Caesium

Registered User
Apr 13, 2006
7,525
184
The statistics required to build a good model for predicting success in hockey do not exist yet. Corsi and fenwick are the closest models that exist today and it's clear that they aren't good enough. The updated model that some people are working on to gauge shot quality is interesting but even that isn't good enough. Simply charting where shots are taken from does not tell the whole story. There is a huge difference between someone taking a shot from 5 feet out and someone taking a shot from 5 feet out after receiving a cross crease pass. One scenario leaves an open net, the other doesn't. Unfortunately the stats that are recorded today are not high enough resolution to model the plays that happen on the ice. Until it's possible to accurately break down why a goal is scored, it won't be possible to create a mathematical model to build a team around, which is the end goal of statistical analysis. That said, just because today's stats aren't good enough, it doesn't mean that stats should be ignored. There will be many incorrect statistical models created before a good one is discovered, but making the incorrect models does hold some value.
 

sanitysrequiem

Registered User
Nov 14, 2009
1,426
115
The statistics required to build a good model for predicting success in hockey do not exist yet. Corsi and fenwick are the closest models that exist today and it's clear that they aren't good enough. The updated model that some people are working on to gauge shot quality is interesting but even that isn't good enough. Simply charting where shots are taken from does not tell the whole story. There is a huge difference between someone taking a shot from 5 feet out and someone taking a shot from 5 feet out after receiving a cross crease pass. One scenario leaves an open net, the other doesn't. Unfortunately the stats that are recorded today are not high enough resolution to model the plays that happen on the ice. Until it's possible to accurately break down why a goal is scored, it won't be possible to create a mathematical model to build a team around, which is the end goal of statistical analysis. That said, just because today's stats aren't good enough, it doesn't mean that stats should be ignored. There will be many incorrect statistical models created before a good one is discovered, but making the incorrect models does hold some value.


Thank you for the reply. I guess what I'm struggling with is, intuitively I feel like Corsi numbers stabilise themselves quite well over a period of time. But the true wildcard here is the efficiency of converting on those Corsi events. If you break it down, I believe it comes down to a few things to score a goal:

a) Events which aren't blocked
b) Events which aren't missed shots
c) The inherent quality (independent of player skill) of the shot which was directed to the net (both in terms of position and player openness/time, as well as difficulty of shot to handle - e.g. a tipped shot)
d) The ability of the shooter to score given those opportunities
e) That transient, ineffable, mysterious "luck/randomness"


Does it make sense that we should be isolating these variables? If we can get a good stabilised number for the percentage of shots which are blocked or missed (I think this is the easy part), and then isolate the luck/randomness bit, we are left with the true talent of converting Corsi events.

I was reading somewhere that to isolate luck/randomness, someone ran a regression on shooting percentage at odd/even seconds during the season. This was supposed to remove all randomness, and the difference between a perfect correlation and what the result was, indicated the level of luck/randomness. Does this make sense?

Logically this all makes sense to me but I just don't have the expertise and know-how to follow through on doing the analysis.
 
Last edited:

MVP of West Hollywd

Registered User
Oct 28, 2008
3,491
951
The NHL should track time of possession just to get rid of these Corsi fights.

Optimal information:

Time of possession
Where a team's shots come from and how many from where

Get that widely available and I think we could get a great grasp. We could look at the Leafs and see eg bottom 5 in possessions and they create shots in cold zones, but give up shots from hot ones... well I guess they're screwed
 

Ad

Upcoming events

Ad

Ad

-->