Scoring Goals: Shot Generation or Shooting Percentage?

There are two things that must occur to score a goal.  The first way is to get an opportunity to score and the second is to capitalize on that opportunity to score.  There are a number of statistics that we can use as a proxy for opportunity to score but one of the most common is Fenwick numbers which are shots + missed shots (some call this Corsi but I define Corsi as shots + missed shots + blocked shots).  We can then define the ability to cash in on opportunities as shooting percentage, or in this case fenwick shooting percentage.  So let me define the following:

Opportunity Generation = Fenwick shots per 20 minutes of ice time.

Capitalization Ability = Fenwick Shooting Percentage = Goals Scored / Fenwick shots

So the question I pose today is this:  What is more important in scoring goals, generating opportunities or the ability to capitalize on those opportunities.  To answer this I calculated each teams Fenwick per 20 minutes (opportunity generation) and each teams Fenwick Shooting Percentage (capitalization ability) and compared them to the number of goals they generated per 20 minutes of ice time and I did this for each of the past three seasons (I only considered even strength five on five data).  I also did this for both the offensive and defensive ends of the ice for a total of 90 data points offensively and defensively.

First for the offensive end of the game:

As you can see, shooting percentage (opportunity capitalization) has a much stronger relationship with scoring goals than getting shots (opportunity generation).  What about the defensive end of the game?

Again, opposition capitalization rates are much more correlated with scoring goals than opportunity generation.  In fact opportunity generation appears to have no correlation with giving up goals at.

The conclusion we can draw from these four charts is when it comes to scoring goals, having the ability to capitalize on opportunities (shots) is far more important than having the ability to generate opportunities (getting shots).  Controlling the play and generating shots does not mean you’ll score goals (just ask any Maple Leaf fan), having the talent to capitalize on those opportunities is what matters most.  From my perspective, this means the usefulness of ‘Corsi Analysis’ to be minimal, at least for the purpose of evaluating players and teams.  For evaluating goaltender workload, as it was initially intended by its originator former NHL goalie and Buffalo goalie coach Jim Corsi, it still has merit.

This article has 24 Comments

  1. Very interesting. Am I right to assume that this would go a long way in explaining why the Leafs’ Corsi rating of the past 2 seasons seems to be such an outlier?

    1. I thought the Toronto Maple Leafs have a higher Corsi because they were often trailing, and teams that are trailing take more shots (especially in the 3rd). No?

  2. Agreed, shooting percentage drives outcomes. Nobody questions this.

    But what you should look at next is 1) the persistence of shooting percentage and 2) the persistence of Corsi. You’ll find that team shooting percentage is mostly transient, while Corsi is much more persistent.

    That’s the key attribute of Corsi and why people use it.

      1. No, you have to actually run the numbers to figure that out, which I encourage you to do.

        Corsi EV score tied predicts ~40% of scoring and is something like 90% persistent (depends on the data set you use.)

        Shooting percentage is only 35% persistent, so it would have to account for 100% of scoring in order to have as much predictive value as Corsi.

        Also, you should be using z-scores so that you can fit a line with no constant.

        1. Ok, I’ll accept your numbers and your theory. Corsi is 40% predictive and 90% persistent which equates to Corsi explaining about 36% of a teams talent. Am I going to base any conclusions about how good or bad a team is knowing only 36% of the story? Nope. Corsi still doesn’t cut it for me.

          1. That’s just silly.

            38% of winning percentage is luck.
            36% is due to the persistent portion of Corsi.

            And then we have all the small stuff:

            – Goaltender talent
            – Talent related to taking/drawing penalties and PP%/PK%
            – Finishing talent
            – Whatever else you might like to add

            Goaltending, special teams and shooting talent *together* account for 26% of winning.

            It makes zero sense to reject something worth 36% in favor of something that – at maximum – is worth 26%. But that’s what you want to do.

  3. The conclusion we can draw from these four charts is when it comes to scoring goals, having the ability to capitalize on opportunities (shots) is far more important than having the ability to generate opportunities (getting shots). Controlling the play and generating shots does not mean you’ll score goals (just ask any Maple Leaf fan), having the talent to capitalize on those opportunities is what matters most. From my perspective, this means the usefulness of ‘Corsi Analysis’ to be minimal, at least for the purpose of evaluating players and teams.

    This is silly and, to be honest, poor analysis. It’s not particularly a secret that shooting percentage correlates more tightly with goal scoring and goals allowed than does Corsi or SF and SA.

    The problem that you run into is that, over the course of an NHL season, randomness swamps talent in terms of “finishing ability” or “puck stopping ability.” Shots for/against are, in the context of an NHL season, a skill, in the sense that you can observe teams that are better at it doing better at it. Percentages aren’t really, in the sense that a single season doesn’t really tell us that much about a team or individual’s ability in that regard.

    In order for your statement to make sense, “finishing ability” would have to have some predictive value. Let me know when that post goes up because that’s the post that will stick a knife in Corsi.

    1. Tyler:

      Does measuring “finishing ability” as a corsi shooting percentage (Goals / corsi shots) rather than traditional shooting percentage (Goals / shots) smooth out any of the randomness? Intuitively to me it seems like it would somewhat. If I miss the net 9 times in a game and then pot a goal, I get 100% shooting for that game by the traditional measure. Next game, I bury 9 shots into the goalie’s midsection, and then pot a goal I am 10% for that game. But really, my “finishing ability” should be considered the same for both games.

      I’ve seen yours and Vic’s posts on the randomness of shooting pct, but those used traditional shooting pct, no? (Correct me if I’m wrong) I would be curious to see if the randomness is reduced at all by using a Corsi shooting pct. Even just having the larger sample size that comes with corsi shots should smooth some of the randomness. But I also wonder how much of the randomness is directly attributable to not counting the missed shots in traditional shooting pct. I would fully expect there to still be significant randomness, but I would also expect there to be less randomness than the traditional shooting pct. The question would be would the randomness be reduced to the point where it has any predictive value. I suspect not, but it would be interesting to see the comparison.

      1. Alan,

        I looked at finishing talent including missed shots (look up shooting talent on There’s not much talent there at the team level. Tom Awad has some pieces on this too.

  4. Tyler, I don’t disagree that getting/giving up shots is a skill, it is just that that skill doesn’t correlate well with the scoring goal, which is what really matters. So, while it may be a skill to get shots, it really doesn’t matter much.

    I also don’t necessarily disagree with your second paragraph with about randomness trumping ability over short periods of time (three years seems to be a good starting point for reliably evaluating talent), but the point is with Corsi, uselessness trumps ability.

  5. Two questions, David.

    First, how big is your sample size? The ability to capitalize on a shot involves getting yourself in position, taking a quality shot, and “other.” “Other” can be quite significant and luck-driven – so the larger the sample size the better.

    Second, could you elaborate on how your findings adversely affects Corsi analysis?

    The primary attraction of Corsi analysis in the first place include:
    – There is far less luck (and greater consistency) in attempted shooting than there is on the ability to capitalize on an attempted shot.
    – There are ten attempted shots for every capitalized attempted shot (aka goals), diluting the effect of factors outside the player’s control even further.
    – These attractions are even greater when you account for zone starts, and game situation (ie. teams take more shots when trailing).

    I don’t see how either of these attractions has been undermined by your findings.

    1. The average is about 141 goals for on 2464 Fenwick shots for per team season of which there are 90 (30 teams, 3 seasons).

      The importance is, if getting shots does not correlate well with scoring goals (which is what matters) then the value of shot statistics might not be that much greater than know how much gatorade a team goes through in a season even if drinking gatorade is a repeatable and persistent skill. Of course I am being a little facetious here in comparing gatorade drinking to shots taking as shot taking does have some correlation to scoring and drinking gatorade probably doesn’t, but if the corellation isn’t that great its usefulness is limited.

      1. If X Then Y.

        X = “getting shots doesn’t correlate well with scoring goals”
        Y = Only interesting if X is true.

        Is X true? Is that what you’ve proven? If so, I’m totally missing it.

        Now let Z = “Ability to capitalize”, but regardless of Z …

        You say Z is better, but:
        1. Even if you found something better, it doesn’t mean attempted shots don’t correlate “well”
        2. Even if Z is better, is it consistent enough to be preferable to Corsi even over small sample sizes?
        3. Is Z even a thing?

        1. 1. True, there is a correlation, just not much of one as shown above.

          2. True, Z does poorly over small sample sizes. That doesn’t mean we should all jump full force onto the Corsi bandwagon (especially without understanding its limitations). The better goal, IMO, should be to attempt to study larger sample sizes, especially at the player level. Besides, you would never actually use Z because you’d be better off using goal data.

          3. I believe it is, and when I present player data (probably tomorrow) I think you’ll find some interesting results.

          1. We already know that there’s identifiable shooting talent at the player level. Just not at the team level.

          2. Lots of people have analyzed this. Likens had a piece over a year ago on the role of luck in team-level shooting percentage (quite high). Tom Awad looked at “shot quality” and found that there was minimal talent at the team level. Good teams create more opportunities – they don’t create better opportunities on average.

            Like I said above, you’ve got 26% to allocate to goaltending, special teams, team shooting talent and whatever else. That’s not a lot of real estate for finishing ability.

          3. Thanks for taking the time to respond, and to present your work in the first place.

            In my opinion, you haven’t discredited Corsi. It’s still a lot easier for a player/team to get a single lucky goal than ten lucky attempted shots. The argument that there are superior options in large sample sizes is correct, but already well-established.

            I agree that we conduct too much of our analysis on small sample sizes. I like to use 3-4 seasons, but having that much data is a luxury, and more accurately reflects what a player was, not what he currently is.

            Still, there are a lot of misperceptions about Corsi. If I can say anything about it:
            1. Most useful in small-medium sample sizes, due to the larger volume of data and smaller influence of luck.
            2. Fundamentally its a measure of puck possession.
            3. Therefore is affected by zone starts.
            4. And the score: teams in the lead generally surrender more shots (and puck possession) in the third period.
            5. Is of course still subject to the influences of QualTeam and QualComp.

  6. So again, we’re on the same page:

    1) Corsi is a skill, but it doesn’t correlate perfectly with scoring.
    2) Finishing is random (ie – not a skill) but it drives observed performance.

    Next step, David: put some numbers on this. How persistent is Corsi? How much predictive value does it then have? How persistent is finishing? How much predictive value does it have?

    What you’ll find is that Corsi has more predictive value than Shooting percentage. If you find the opposite, then like Tyler said, it will stick a knife in Corsi. I look forward to seeing your results.

    1. Actually, I am not sure I would agree with either of those statements.

      1. Corsi is a skill, but doesn’t correlate well with scoring.
      2. Not sure it’s random, its just that you can’t identify the skill very well with one year of data.

      More to come in the upcoming days/weeks.

      1. Again, you should look at regulation winning percentage vs shot differential and use Z-scores. That’s the right analysis to start with.

        If finishing is a team-level talent, then you’ll be able to identify it with a single year of data. If things aren’t identifiable, then they’re random.

Comments are closed.