Read on Twitter
I like thinking of Expected Goals in this way: Watch a play and then freeze frame at the exact moment that a puck is released. An expected goal value should closely reflect an experienced hockey mind's intuitive sense of shot danger in that instance.
The objective isn't to predict the chance of a goal being scored on a shot as well as we possibly can. If that were the case, we could theoretically add in post-release data like shot placement, goalie reflexes and perfectly predict shot outcomes and end up with... well... goals
Why do we exclude post-release data? Why this arbitrary line in the sand? Because the idea is that things like variations in shot placement and goalie save efforts are not sustainable paths to scoring goals. The belief is that if you depend on them, your well is going to run dry
This theory is supported, frankly. In general,
Goals vs. xGoal differences are not repeatable at the skater, team or goalie level. So why is Corsi more predictive than Expected Goals in some contexts?
Goals vs. xGoal differences are not repeatable at the skater, team or goalie level. So why is Corsi more predictive than Expected Goals in some contexts?
If you start to investigate the high xG-value opportunities in hockey, you start to notice something. Most of them were facilitated by a 50/50 play breaking in the right way, at the right time to an attacking team who happened to be perfectly situated to take full advantage.
This is perhaps different from soccer, the sport in which xG was introduced, where quality chances follow a slow-building string of skillful touches and playmaking. In fact, in public hockey xG models, the majority of high-value opportunities tend to come off of rebounds.
In fact, @MoneyPuckdotcom has shown that xG is actually more predictive when rebounds are excluded entirely. You could argue that rebound generation, not unlike the post-release factors mentioned earlier, fails to qualify as a sustainable means of scoring goals at the NHL level
Corsi out-predicting xG ~ in some cases ~ may suggest that controlling play and placing yourself in a better position to deal with hockey's inevitable dumb bounces is a better model for long-term success
That said, this is not an either/or situation. Neither Corsi or xG are going away anytime soon in hockey - their relative value is circumstantial
And FWIW, at @HockeyAnalytics, we lean more heavily on our xG model than Corsi for predictive modeling. The improved predictive value of xG is likely due to some combination of the additional variables at our disposal, better data quality and having expanded info on blocked shots
