Advanced Hockey Stats 101: Fenwick (Part 2 of 4)
Here we take a closer look at Fenwick and compare it to Corsi while making some visualizations along the way. This also serves as a primer for future articles delving into the validity and usefulness of hockey stats to make machine learning-based predictions.
Agenda
- Fenwick definition and theory
- Correlation between Fenwick and Corsi
- Which is better for predicting points over the course of a season?
Data
2019–2020 advanced stat data was downloaded from hockey-reference.com. 2018–2019 team data was downloaded from IcyData.
Fenwick definition
Fenwick is another measure of shot attempts, a proxy of offensive pressure. More specifically, it counts any shot that is on goal, hits the post or misses the net during 5-on-5 play. That should sound awfully similar to Corsi, discussed here, as the only difference in the calculation is that Fenwick does not include blocked shots.
Team Fenwick
Like Corsi, a team’s Fenwick is often measured as a percentage of the game total Fenwick. Therefore, a Fenwick For % (FF%) above 50% indicates that team had more unblocked shot attempts than their opponent.
Player Fenwick
Fenwick, and most other advanced stats, also apply to individual players. It simply compares FF% while a player is on the ice to FF% while the player is off the ice. For example, if Nathan MacKinnon has a good game and generates a lot of offense and is responsible in his own end, then his FF% on would be high, say 65%. If his team was dominated when he was off the ice, his FF% off would be low, say 40%. This would leave Mackinnon with a large FF% rel of 25%. These numbers would indicate he was a difference maker on the ice.
Correlation between Fenwick and Corsi
As expected, FF% and CF% are strongly and positively correlated. In fact, the median values from the 2019–2020 shortened season were 49.8 and 49.9 for FF% and CF%, respectively. Interestingly, Nicolas Roy of the Golden Knights had the highest value of 61.4% in both categories. The biggest difference between metrics was 4.5%, observed in Nathan Gerbe of the Blue Jackets, who had a FF% of 50.2 and a CF% of 45.7%. Ultimately, these results suggest that Corsi and Fenwick are capturing the same information at the individual player level.
Are blocked shots really that common/important?
A quick look at the 2018–2019 season shows there was an average of ~100 shot attempts and ~15 shot blocks a game during the regular season. If shot blocks are evenly distributed between opposing teams, then there would not be a major difference between Corsi For % (CF%) and FF%. However, some players block far more than others, for example, defensemen like Mark Giordano and Chris Tanev. As a result, there are some instances where CF% and FF% tell two different stories.
Why bother differentiating blocked shots?
For some players, blocking shots is part of their defensive game, therefore it can be argued that blocked shots shouldn’t be rewarded as shot attempts, as they are in Corsi, since they are (sometimes) deliberately thwarted. A major criticism of Corsi is that it doesn’t discriminate between high and low quality shots and that the volume of shots doesn’t actually reflect offense and possession.
Since Fenwick considers shots that actually make it through the defense, it is sometimes thought as more stringent than Corsi but more lenient than shots on goal (SOG), in which hitting posts don’t count. However, a blocked shot doesn’t necessarily mean it was a poor scoring opportunity to begin with, so like Corsi, Fenwick also has its limitations.
Does Fenwick or Corsi better predict team wins?
We have discussed the theory behind Fenwick and Corsi, but the real test is which one provides more predictive value.
A simple linear regression of each variable reveals that both metrics are statistically significant, however, FF% is slightly more informative. This is also captured by the correlations in the scatter plot above that show FF% is more strongly correlated with points than CF%. It is also worth noting that the range for CF% is much narrower than FF% over the course of a full season. Over time, blocked shots begin to add up. Teams like the Senators, Kings and Devils likely had a higher fraction of their shots blocked while teams like the Lightning, Bruins and Flames blocked more shots.
Conclusion
Fenwick is measure of unblocked shot attempts. As expected, it is highly correlated with Corsi at the player level.
We’ve previously seen that on a game-by-game basis, Corsi is not particularly useful for predicting wins. In fact, a higher CF% for the home team actually predicts a loss. Here, we’ve shown that a higher Corsi over a full season does in fact correlate with more points, and that this relationship is actually stronger for Fenwick.
These results may sound contradictory, however, in this analysis, we considered all 82 games for each individual team. Previously, the game-by-game analysis was performed across all 31 teams together. As a result, we could make general statements about the average team. However, averages can be easily skewed by a subset of teams; recall, the medians were still close to 50% and there were still hundreds of games in which the team with CF% < 50 lost and the team with CF% > 50 won.
With that aside, if I had to pick one between Corsi and Fenwick, I would go with Fenwick. Let me know if you agree!
