NHL predictions and rankings - December 31, 2019
The outcome of a hockey game, or any other team sport, depends on the relative strengths of the two teams and random events (luck). Observed outcomes inform about teams’ strengths which can be summarized in a statistical model.
I estimated a model of the NHL 2019-20 season with the objective of predicting standings at the end of season and the winner of the Stanley Cup, and rank teams by strength (power ranking). I give technical details at the bottom of this post.
Predictions for the 2019-20 season
The table below shows the latest predictions. Let’s take the Washington Capitals as an example to understand how to read the table. At the end of the season, the expected points for the Capitals is 110.3, giving them a 27.5% probability of winning their conference, a 10.5% probability of winning the President’s trophy (team with the most points) and a 99.7% probability of making the playoffs. The Capitals will make it to the quarter finals with a 60.7% probability, make it to the semi-finals with a 26.6% probability, to the finals with a 8.1% probability and winning the Stanley Cup with a 2.7% probability.
Table 1: Predictions for end of season standings in the NHL
Power ranking
The power ranking is based on simulated seasons where a team plays every other teams, including itself, at home and away. In the table below, the Blues are the top team with 114.42 points in a balance season, in which they would score on average 3.40 goals per game and allow on average 2.29 goals per game.
Table 2: Power ranking of NHL teams
Technical details
Below are a few details about the methods I use for the predictions and the power ranking for those interested. I do not go into all the technicalities but I can discuss them with you if you wish.
Predictions for the 2019-20 season
The model assumes that a team’s strength can be summarized by six parameters: 1) offensive rating at home; 2) defensive rating at home; 3) offensive rating away; 4) defensive rating away; 5) rating of a team in overtime at home and 6) rating of a team in overtime away. This means that the model contains 31*6=186 parameters. The model also assumes that recent games are more informative of teams’ strengths and as such more weights are given to recent games.
In practice, the final score of a game depends on how the game develops. That is, the number of goals scored by a team will affect the number of goals scored by the other team. My model does not consider this. The number of goals scored by the home team is independent of the number of goals scored by the away team. The model focuses on the final score and does not consider the order the goals are scored. Because the development of a game and its final score depends on the offensive and defensive abilities of the two teams, then my model will get the average scores right.
Once I have estimated the model with the most recent data, I use its outcome to predict the rest of the season, adding to it the points already accumulated. The model generates games’ scores. If the outcome of a game is a tie, then the model determines which team wins in overtime.
The outcome of any game is affected by the relative teams’ strength and stochasticity. To take this into account I simulate the rest of the season several times. I use the outcome of the simulations to calculate probabilities of each team making the playoffs, winning their conference, winning the President’s trophy and reaching the quarter finals, the semi finals, the finals and win the Stanley Cup.
Power ranking
The team who finishes at the top of standings is not necessarily the best team. Indeed, teams do not play the same schedule and some teams have a schedule that can be significantly easier than others. My model can help in ranking from best to worst.
I simulate the outcome of balanced seasons where a team play every other team and itself, at home and away. This makes a 62 games season from which I calculate the point accumulated by a team and then rescale it to 82 games. I make each team play a balanced season 1,500 times to calculate the expected number of points over 82 games.
I use the model also to rank teams by the strength of their offense and the strength of their defense. I then calculate the difference in goals (goals scored - goals allowed) and the ratio (goals scored / goals allowed) to better compare teams.