It kind of got buried, what with the “Zack Kassian should be in jail” post, but there were some interesting comments on the post below that about score effects and how things change when one team has the lead. It’s pretty well known at this point that teams with a lead tend to do worse in terms of shots than they do when the game is tied. This is kind of difficult to wrap your head around because coaches will frequently insist that they don’t tell their teams to sit back or anything when they get a lead but the evidence that something happens is overwhelming.
Jesse Dahl expressed one theory pretty well:
Assuming there is some data out there that the number of total shots goes down for both sides (i.e. low event hockey), and I’m pretty sure there is, do you think the following could be true?
At some point in the game it would be more optimal to just sit back even though in the long run it is a losing strategy. In other words, a goal for is less valuable than a goal against is damaging when you’re near the end of the game. If you can lower the variance over that small period of time you will get 2 points more often (and hopefully?) more points on average.
To use an extreme example to illustrate the point, with only 30 seconds left (or less) in the game, up by one, prevent defense is obviously optimal over a more aggressive style even though it is more likely you will score a goal with a more aggressive style. A two goal lead is not as valuable here as preventing a goal.
The reasoning is correct, I think, so let’s look at the data. I grabbed the data for while game play is tied and while a team has a one goal lead for the past six years. Does the data support the theory that the total number of shots for both sides is reduced when one team has a one goal lead?
So, nope, that doesn’t really work. It’s a reasonable hypothesis because it would be justify getting outshot if you could drive down the volume of shot attempts, but the data says it doesn’t really happen. We know that shooting percentage increases for teams that are leading:
The bump in shooting percentage doesn’t make up for getting outshot by as much as teams leading by one goal do though:
It’s not a huge change but, if anything, it favours the teams that are trailing, not the teams that are leading.
I wondered if maybe this was a bad team thing. Maybe the bad teams screw it up and can’t hold leads. Maybe teams that are good when things are tied are strong when they’re leading too? I grabbed the data for the top five teams based on goal share in each of the past six seasons to check. Here are the same graphs:
So a huge difference in goal difference between when the game is tied and when they’re one goal up.
And again, we see that there’s a big decline in terms of the share of the shots that they get from when they’re tied to when they go one up.
This is a bit different from the norm. The teams that did best in terms of goal difference when things were tied were really good (or lucky) at keeping down their opponents’ shooting percentage and that disappeared when they were one up. I suspect that that’s just some regression and not a coaching thing.
Basically though, it shows the same thing: the best teams in the NHL over the past six years in terms of outscoring when the score is tied get fewer than 50% of the shots when they go a goal up despite crushing their opposition in shots when the score is tied. Their 5v5 goal difference in tie games was 0.77; when they went one up, it fell to 0.13. Note that, as with the league-wide data, there’s no reduction in the total number of goals scored at 5v5; it actually increases slightly.
In the absence of some confounding factor, there’s something not right here. We expect people to make decisions that maximize their outcomes. If the best 5v5 teams in the NHL when the score is tied become much less dominant when they go up one goal, either they’re making improper tactical decisions when they go up a goal or their opposition are making improper tactical decisions when the score is tied or some combination of the two.
Some of this may relate to the data not being sufficiently granular. We know, for example, that teams will play to the score late in a tie game – if a game is tied with, say, five minutes to go in the third, both sides are generally happy to take their point and head to the lottery. This probably starts earlier than that.
It’s worth digging into this a little more deeply, I think. The first step would probably be doing it on a time stamped basis. That would allow you to cut out the score effect of teams playing for OT/SO. Even if you cut that out, I’m having a hard time imagining that it would explain the results that we see. Odd.Email Tyler Dellow at email@example.com