Sports Analytics

Insights from watching a lot of sports and digging through data.

Empty Nets

I don’t claim to know everything about Ice Hockey tactics, but one thing I do know is that the last couple of minutes of a game very often make or break your bets. The reason is that when a team is losing, they will try to substitute the goalkeeper for an extra attacker, going all-in with an empty net to make a comebak. This is typically only done very late in the game and the increased volatility in goals has the potential to turn losers into winners, but frustratingly, also winners into losers - often in the last moments of the game.

The decision to pull the goalie is made by the manager or coach, and it seems to me this decision is kind of like 4th down decisions in American Football - managers and coachers see things one way, while analysts and gamblers see them another way. Usually, the guys in charge are too risk-averse, anxious not to make an active decision that could lead to a sure loss. If they pull the goalie and concede an open net goal, they might look silly, but if they just play on as usual, it will look like a regular loss, and they can’t be blamed for messing up. So they wait and wait, most often much too long to optimally increase the chances of success. What they really should be doing, though, is going for it. Generally, goal difference isn’t a factor, so they have nothing to lose when pulling the goalie early.

Not only do they have nothing to lose, but risk-averse coaches are actually leaving value on the ice with their conservative approach. There have been many studies on the optimal time to pull the goalie, and the results have been conclusive: if coaches want to maximize their chances of success, they should go for it much earlier than they currently do. But the question is, when is the actual optimal time to pull the goalie?

Using my Ice Hockey Simulation Model, which does a good job of quantifying the effects of empty nets, we can force the losing team to pull the goalie at any given time. Above we see an example where the two teams are evenly matched in terms of skill, but we are one goal down with 10:00 left on the clock in the third period. Let’s define success as coming back to at least drawing the game. The way to read the above plot is this: when at 10:00, what would our chances of success be if we waited to pull the goalie until x left on the clock? In this specific scenario it turns out it’s optimal to do so with around 4:00 left, much earlier than what most coaches tend to do, and much better than not pulling the goalie at all. But note that this assumes we actually can pull the goalie when we want to do so, which isn’t always the case. We want to be in control of the puck, not under significant pressure, not have any penalties interfering, have the right players on the ice etc - this isn’t always easy, thus we need to introduce different p[empty net]’s. Note that these values also apply if we go two or three goals down sometime after 10:00.

Comparing different probabilites of actually being able to pull the goalie we see some interesting things. First, we see that a certain empty net at 10:00 is actually slightly better than not pulling at all. Second, pulling too early is almost as bad as not doing it at all. There is a sweet spot around 07:30 - 04:00 when we should try to pull the goalie, exactly when depending on how good chances we think we have of actually getting that extra attacker on the ice.

This analysis is of course only a rough one, to be able to predict games for profitable betting we need to dig further. After all, the decision on when to pull the goalie is on the coach, so data on their individual tendencies are of course valuable. Most are still fairly conservative, but others are starting to catch on…