Last season, as hockey analysts struggled to explain how a possession-dominant Kings team failed to make the playoffs while Anaheim and Vancouver topped 100 points, there was a lot of discussion surrounding the role of one-goal games in the standings. LA’s disappointing season was largely dismissed as bad luck, with an argument that went something like this: the outcome of a one-goal game is effectively random, and the Kings’ 13-9-15 record in these games (against the 33-1-7 and 22-4-5 performances of the Ducks and Canucks, respectively) was the difference in keeping them out of the postseason. I wasn’t entirely convinced by this, but it got me thinking about the randomness of close contests. How random are one-goal games, and how significant a problem is this for people trying to use numbers to understand why some teams win and others don’t?
Primarily using WAR on Ice, I gathered game-level data from the nine 82-game NHL seasons from 2005-06 to 2014-15 (because I wanted to look at single-season counts of one-goal games, I excluded the short 2012-13 season), and characterized games as non-1GGs, regulation one-goal games, and OT/SO games. Except in unusual cases where they turned out to be game-winners, I dropped empty-net goals from final scores and characterized the game accordingly. So, for example, a 4-2 game where Team A’s fourth goal was an ENG would still be considered a regulation 1GG. In no particular order, here are five things I found out about one-goal games.
1. One-goal games are increasingly common.
There are a lot of reasons for it (the Bettman Point and decreased scoring league-wide are two big ones), but the NHL of today has more close games than ever. Back in 2006-07 and 2007-08, NHL fans were treated to 667 and 668 one-goal games, respectively; in 2013-14, we were up to 709 such games, and last season, there were 730. If you’re not a fan of increasingly random factors determining who gets into the NHL playoffs, this is something that should bother you.
2. Stronger teams win non-one-goal games pretty consistently, but regulation one-goal games aren’t entirely random.
If you’re the sort of person that enjoys using numbers to understand the results you see in the NHL (or, you know, you just gamble a lot), you should love non-one-goal games. For one thing, the year-on-year correlation in an NHL team’s non-1GG winning percentage is 0.5; put another way, about 25% of the variation in teams’ win % in non-close games can be explained by last year’s win %. And, as the following scatterplots show, that win % correlates closely to most measures of “this team gets good results”. (Note: These are full-season measures of GF%, CF% and PDO.) The GF% result isn’t surprising (the best way to end up with a big positive goal differential is winning a lot of blowout games, and vice versa), but possession and PDO each have a strong role in driving these outcomes.
Now, if one-goal games are random, you’d expect zero autocorrelation from season to season, and you’d expect the graphs below to look like clouds with a flat trend line. And when you look at overtime and shootout games, that’s exactly what you see. But one-goal games ending in regulation fall somewhere between random and predictable.
The autocorrelation in teams’ win percentages in these games is about 0.22; not terribly strong, obviously, but also pretty far from zero. What this all suggests is that 5v5-based hockey analytics are awesome in games that aren’t close, useless in OT/SO situations, and marginally useful in regulation one-goal games. It also suggests an interesting endogeneity question (how much of observed shot differentials is attributable to inferior teams playing conservatively to force overtime?) that I’ll let go for now. Given that non-close games are becoming less common, of course, none of this is great news for analytic game prediction.
3. In a lopsided match-up, the stronger team still has an excellent chance of winning a close game.
While the above graphs imply that the outcome of a regulation one-goal game is fairly unpredictable, there are situations where analytics still have considerable explanatory power. When one team’s score-adjusted Corsi % is 5 percentage points or more higher than their opponent’s, they win 68% of non-one-goal games, but they also win 60% of regulation 1GGs. When one team’s large-sample even-strength GF% is 5 percentage points or more better than their opponent’s, they win 71% of non-close games and 60% of regulation one-goal contests. In other words, so long as you have a strong sense of the relative underlying strengths of each team, you can be fairly certain that a much stronger team will win any game that doesn’t go to overtime.
4. Home ice advantage and back-to-back games don’t really matter in close games.
For someone who’s spent a ton of time writing about home-ice advantage and back-to-back effects, I was a little surprised to see that these effects are almost entirely confined to non-close games. That is, home teams win 59% of all non-1GGs, and teams on a back-to-back win just 40% of the time. In regulation one-goal games, however, the home-ice advantage is just 53% (close to coin-flip territory), and teams on back-to-backs win 48% of the time (both factors are irrelevant in OT and shootout games).
5. It’s largely impossible to guess how many one-goal games a team will play each season.
Of course, all of this would be super-useful information if there was a way to predict which games were likely to be one-goal and non-one-goal games. Unfortunately, this isn’t the case. On average, about 42.5% of games are non-1GGs; the only scenario in which the probability of a non-1GG differs much from this is a back-to-back situation, when about 45% of games are non-1GGs. More generally, as shown in the graphs below, teams that drive either a very good or a very bad goal differential will tend to play more non-one-goal games, but the effect is not as strong as you might think.
So, what are we left with? On one hand, hockey games are a lot less predictable when the score is tight, and close games are becoming more and more common, but it’s largely impossible to anticipate which games will be close and which teams might play a lot of them. On the other hand, it looks like regulation one-goal contests are less random than some people think. (Shootouts are either a crapshoot, or they’re not.) These results can help us to make educated guesses about how often specific teams might face each type of game scenario, and how they might fare. For example, non-one-goal games might be more common in the Eastern Conference, where back-to-backs are more frequent. And, to the extent that forecasts of teams’ possession differential and 5-on-5 goal differential are becoming increasingly refined, this information may help in guessing how specific match-ups might play out. And, on the assumption that PDO will tend to regress strongly to average for most teams, it stands to reason that teams getting crushed in possession might find themselves on the wrong end of a lot of non-1GGs. Still, more work needs to be done before these can be much more than guesses.
One thought on “Some Things to Know about One-Goal Games”
Have you thought about trying to account for the path of how the game comes to a one goal result? For example a game that is tied until decided by a late goal vs a game where one team goes ahead early then the other team scores one or more late goals to make the game close. I expect that when the winning team scores the last goal the results will be nearly random since the probability of scoring is nearly the same for each team. If the losing team scores the last goal(s) then the results will be more nearly like non-1GG results.
The degenerate case would be games that end 1-0 (or 2-0, with an ENG). I’m not sure there are enough of those to create a valid data set.