My s**t doesn’t work in the playoffs. My job is to get us to the playoffs. What happens after that is f***ing luck.
The long 2013-14 regular season is steadily winding down, and increasingly the attention of NHL fans is going to playoff spots and potential matchups. In this spirit, I’ve been asked a few times whether I plan to continue Puck Prediction into the postseason. The answer, happily, is yes.
Of course, once I started thinking about a model for predicting playoff outcomes, it occurred to me that what I’ve been doing for regular season games and playoff forecasts probably isn’t going to work. Most of the “#fancystats” work that hockey analysts do, using shot attempt differentials and other measures, is valuable if you’re trying to determine how much of a team’s results are due to their underlying quality (as opposed to luck). When it comes to the playoffs, however, these statistics are less important. After 82 games, luck effects tend to average out, making it less difficult to assess how good teams are. More importantly, as I’ve found out in the course of picking over 40% of games incorrectly this season, #fancystats aren’t very useful for predicting outcomes over small samples of games. Which is a problem in the postseason, as seven-game series are the definition of small samples.
For me, the ideal starting point for a playoff model lies in the variable with the greatest utility in single-game prediction: home ice advantage. In the eight postseasons since the 2005 lockout, 55.8% of playoff games have been won by home teams. Assuming the 2-2-1-1-1 format the NHL uses, I used this number to work out a bunch of conditional probabilities that described both teams’ chances of winning a seven-game series. For example, if two perfectly matched and perfectly average teams met in a series, the team with home ice would have a 0.519 probability of winning the series. The probabilities of each series outcome, before any games are played, are described in the table below.
What’s more, the probabilities in this model can be updated as we move through the series. Under the same assumptions about a perfectly even matchup, for example, the probability of winning the series jumps to 0.658 if you win Game 1 (because the remainder of the series offers three home games and three road games for each side, the win probability is the same for either team with a 1-0 lead). If the team with home ice wins the first two games, their probability of winning jumps to 0.800. If their opponent wins the first two on the road, however, they have an 0.829 probability of taking the series, which makes sense (i.e., for that team, three of the next four games are at home). And so on.
Of course, few playoff series feature perfectly matched teams, so the model’s other component is a measure of how good each team is. I asked around via Twitter for suggestions on how to measure this, and got a number of responses. Some were less than serious (e.g., “fights/60 (tied)”), but the best idea came from Derek at Fear the Fin: a modified version of GF% that uses regressed versions of shooting percentage and save percentage, and league-average power-play and penalty-killing efficiency numbers. Specifically, each team’s even-strength Sh% and Sv% for 2013-14 were regressed to their three-season averages, and then applied to their total SF and SA at evens to estimate 5v5 GF and GA. League-average PP% and PK% were then applied to each team’s total PP and PK opportunities to estimate power-play goals scored and allowed. The goals for and against were totaled to create a GF% for each team. Ratios of the modified GF% values for each team were then multiplied by 0.558 to characterize each team’s probability of beating the other on home ice; calculation of each team’s probability of winning on the road followed.
Prior to each series, I’ll offer a preview and a pick based on these metrics, with the probabilities of each series outcome being updated as the games are played.