I was born into a family of Islander fans, so I never had a chance to avoid the sadness that comes with that fandom. While Islander fans are sad for a lot of reasons, one constant complaint over the past several years has been their inability to protect a lead.
However, this is not a unique complaint of Islander fans alone. Fans of other teams have similar gripes. For example, the Leafs have been criticized this season on the same grounds. And here’s fellow Hockey Graphs write Asmae when I suggested doing some research on blown leads:
So, are some teams particularly bad at holding leads? Asked another way, is keeping a lead a skill distinct from the rest of the team’s performance, or is it just a function of the team’s overall skill and luck?
In theory, I could see cases either way.
On one hand, people believe something is more common if they can remember multiple instances of it. I watch a lot of Islander games, so of course I remember more of their blown leads than any other team. That doesn’t necessarily mean that they have a unique problem, I’m just a biased analyst. On the other hand, maybe there is some defensive system or player trait that is particularly effective at holding onto a lead.
In this post, I introduce some measurements of blown leads and demonstrate the following:
- There are ways to descriptively measure how well teams protect leads, and they are small improvements on typically used points like “record when leading after 2 periods”
- Holding leads is a repeatable skill that will stay consistent for a team, BUT
- Protecting leads is almost entirely a function of a team’s Goals For % and does not add any additional information to whether a team will win or lose
My key point is that we can measure how well teams have protected leads in the past, but there is no “protecting a lead” skill or system for a team that is separate from their general ability to outscore their opponent in any other situation. Good teams protect leads because they tend to score more than their opponents all the time. Bad teams tend to blow leads because they always tend to get scored on.
Descriptive Data, or, 43% of the Time It Works Every Time
To calculate how well teams protect leads, I looked at every regular season regulation lead obtained in from 2007-2008 to 2015-2016 (to my knowledge, every lead except one has been a regulation lead). This gave me 270 team-seasons of data. On average, a team will have 72 leads in a full season, and will successfully protect 41 of them until the end of the game, so our sample sizes are useable but not enormous.
I calculated 3 different metrics to measure lead protection. The primary one is Lead Protection Rate (LPR), which is simply the percent of all obtained leads that the team maintains until the end of the game. Note that this treats all leads equally, regardless of how many goals they lead by or when in the game it was obtained.
Second, I made 3rd Period Lead Protection Rate (LPR-3), which only looks at 3rd period leads. These includes both leads from earlier that carry over into the start of the 3rd plus new leads that come from scoring in a tied game during the 3rd. Finally, we have Late Game Lead Protection Rate (LPR-Late), which is analogous to LPR-3 but looks at only the final 5 minutes of a game.
The distributions of these stats are shown below. The average team will have an LPR of 43%, meaning that 43% of the time that they get a lead, they will hold on to that lead and win the game. Unsurprisingly, when you narrow the view to leads later in the game, the protection success rate goes up; the average LPR-3 is 59% while the average LPR-Late is 89%.
And just for fun, here are the teams that have been best and worst at protecting leads
Relationship with Winning, or, Don’t Always Use Protection
Now we know lead protection can be measured, but should it be? There are some things we need to check. First, we need to know if these stats are repeatable. We can measure lead protection in the past, but it is not very valuable if it doesn’t tell us anything about the future. Second, we want to see if it provides additional information about winning beyond what we already have available. I’ll show that LPR-Late fails the repeatability test, while LPR and LPR-3 do not expand win predictivity beyond what we know from GF%.
To test repeatability, I split each team’s games each season into even and odd buckets and checked the correlation:
I also tested repeatability with year-to-year tests and by grouping seasons into buckets of 3, and these had similar results. LPR is moderately repeatable, as is LPR-3. But LPR-Late is not. Goals are fluky, and protecting leads in a timeframe as short as five minutes is basically luck; knowing who was good at it last season won’t help you predict who can do it again this season.
This is particularly important because many teams have specific strategies or role players in mind for the final high-leverage minutes of a game. These findings suggest that looking exclusively at performance in a time sample that small is inadequate (or at the very least, we don’t have the data to tell which teams are reliably good at it).
LPR and LPR-3 remain, but how useful are they? To answer this, we should take a step back and see how related they are to metrics we already have:
Unsurprisingly, the three lead protection metrics are correlated with one another and have an inverse relationship with the raw number of blown leads each team gives up. More interestingly is LPR’s close relationship with more general performance measures: Corsi For % and Goals For %. In fact, LPR’s in-sample correlation with GF% (.71) is even stronger than the one between CF% and GF% (.66).
This is important and is the first hint of the flaw that will be LPR’s undoing. LPR is only valuable as a predictive tool if it adds more information that what we already have. Given its close relationship with GF%, we have to study whether we can better predict winning by including LPR or ignoring it.
To test that, I ran two regressions: LM0 assessed total standings points for each team based only on GF%. LM1 was similar but also added in a new coefficient for LPR. Here are the results:
|Model||GF% p-value||LPR p-value||Adjusted R2|
GF% is very significant if you want to predict how many points a team earned that season. Once you include that information, LPR does not offer any additional value. Comparing these two models in an ANOVA test gives a p-value of 0.53, again showing that lead protection does not provide additional value.
Put another way: if we were in a contest to guess the standings, and you only knew each team’s GF% while I knew that and their LPR, I wouldn’t be expected to do any better than you.
Conclusion and Next Steps
These negative results may sound disappointing, but they’re not. Rather, they teach us something about hockey: protecting a lead is not different than trying to outscore your opponent any other way. A team that’s bad probably got lucky to have the lead, so they are more likely to lose it.
Imagine my beer league team played the San Jose Sharks and automatically started with a 1-0 lead. My team would blow that lead almost immediately, but it wouldn’t be because we’re particularly bad at protecting leads compared to other skills. It’s just because the Sharks are way better than us at hockey.
Whether or not the team protects leads should not be used as a diagnostic tool. If a team is blowing a lot of their leads, the solution isn’t necessarily to get more shutdown players who can prevent the next goal; rather it might be to get better offensive players so that the lead becomes larger. The route you take doesn’t matter as long as you build a team that consistently outscores their opponents.
There are plenty of ways to both improve on this work and to expand it. This initial look could be improved with additional predictivity tests and refinements around the size of the lead. Even better would be to add in a measure of expected points to get the unique strength of any particular lead.
In addition, new ways of looking at this data could include studying leads at the player level or assessing the impacts of penalties on protecting leads (thank you to Asmae for this suggestion)
All of the code used to produce this work is available on DropBox here.