There’s pretty broad agreement that coaches of NHL teams have their teams play more conservatively when protecting a lead. It’s reflected in the data – virtually every team in the NHL sees its share of the shot attempts at 5v5 get worse when it’s winning. The thinking goes that coaches have carefully sussed out that they can maximize their chances of victory by playing a defensive game, clogging up the neutral zone and taking advantage of any odd-man rushes that might present themselves as the other team pushes. TACTICS.
I’ve copied the table that Gabe has on Behind The Net, just to illustrate what I’m talking about, in case any of you are less familiar with score effects. 29/30 teams have a lower Fenwick share when up a goal than they do when they’re tied. The reverse is true. The same effect is also visible when up or down two goals, although it’s less consistent, likely because of sample size. Not a lot of hockey is played with a two goal deficit.
NHL teams that were up two goals entering the third last year were pretty dominant: they had a regulation time record of 257-6-27, taking a very impressive 93.3% of points available in regulation. (Aside: if you mostly watch the Oilers, this might surprise you. They took 12 points out of a possible 16 in this situation.) We have pretty solid evidence that those teams, as a collective, became less aggressive when they had the lead and, presumably, in crunch time, were less aggressive still. Get it out, get it deep, one man forecheck (if that), repeat.
We need something to compare this too. There’s a useful way of doing this by assuming that goal scoring is consistent with a Poisson function or distribution (I’ve never learned the exact way of saying this): basically, that it’s random, in accordance with underlying probabilities. Update: Commenter “E” points out that many non-nerd types may be encountering this idea for the first time. I recommend the paper linked here. There’s lots of research that supports this idea, with some caveats – if you apply it to predict a team’s record as a whole, you tend to significantly under-predict the number of games that will end up tied after regulation, because teams respond to the incentives and will play for the tie if they’re tied in the third period.
What I did was take each team’s GF/G and GA/G and calculate the probabilities of winning, losing and ending up tied in a game in which they led by two after two periods. As you can see, even terrible teams are awfully likely to win. I then multiplied those probabilities by the number of games in which they actually had a two goal lead after two. It’s the row at the bottom that’s really of interest.
On this way of calculating things, all of that Get it out, get it deep, one man forecheck, repeat resulted in teams winning virtually exactly as many games as we’d expect with a two goal lead after two, losing two and a half fewer and drawing two and a half more. In 290 games. It doesn’t seem like much of a reward for playing less fun hockey.
There are, of course, a couple of confounding factors here. It’s entirely possible that there are other factors at play. Maybe, for example, referees are way more likely to give penalties to the leading team and third periods for teams up by two have a much tougher row to hoe, in that they face more PP against. Wouldn’t shock me. It’s also possible that teams trailing by two shorten the bench significantly and that using their overall goal differential isn’t really appropriate, in that the players who are on the ice are a sort of a better team.
That runs both ways though – teams protecting a lead are known to shorten the bench as well. Moreover, it’s implicit in the math here that teams are playing against the average team that they faced; in fact, the average game where one team leads by two entering the third features opposition that’s weaker than usual.
I ran the same table for teams with one goal leads entering the third period. It’s a little more interesting.
You can see that the Poisson Excel function significantly overestimates the number of wins and underestimates the number of ties. It also overestimates the number of losses, although not by a huge amount. This is, presumably, teams responding to the incentives that are before them – if the team that’s trailing scores, everyone’s happy to go to overtime and put a point in the bank. My suspicion is that the team leading the game becoming more conservative probably changes the odds of a tying goal from what they otherwise would be, making it more likely that the trailing team will tie to score the game than they otherwise would.
The really funny thing about this is that playing conservatively seems to have little in the way of benefits for the team leading. If you calculate the points on the basis of regulation win being worth two points and a regulation tie being worth 1.5, you would expect the teams leading by one to average 1.66 points. Actual average points collected by teams leading by one heading into the third: 1.67. The real beneficiaries are the teams trailing by one heading into the third period. Their expected points per game is 0.56. They actually averaged 0.64 points per game.
As with two goal leads though, I’m not sure that this suggests that playing a more defensive game is actually an optimum strategy or that it’s worth doing, when you consider that hockey’s not just a game but also supposed to be entertainment. On last year’s data, teams pick up an extra 1 point per 100 games in which they have a one goal lead (subject to all my caveats above about how the data might be influenced by other factors.) They give up an extra point per ten games though which can’t be good – I’ve found before that it was weaker teams who benefited disproportionately from the OTL rule; this might help explain why.
In any event, the point I’m really interested in this: we’re told that defence wins games and that settling into a defensive shell is the way to go about doing things when you have the lead. If this was really true, shouldn’t it show up in the data? (Subject, again, to the caveats above ways in which the data might be misleading.)