By Kevin Minkus (@kevinminkus)
As in our 2016 iteration, playoff probabilities come from a combination of where teams are now in the tables, what their remaining schedule is, and how good our model thinks they are. The remaining games of the 2017 season were simulated 10,000 times based on win-loss-draw predictions for each game. The probabilities and averages given below are calculated from those simulations.
You'll notice that we're missing a Supporter's Shield column this year - that's because in all 10,000 of our simulations Toronto won it. To reiterate just how great Toronto's season is going, on the final weekend of the season last year we still had a 35.7% chance that Colorado would win the Shield.
A bit more in-depth explanation follows:
To predict game outcomes, Dixon and Coles’ method is followed. First, the number of goals each team is expected to score is calculated from a Poisson GLM. That model, built on data from only the 2017 season (for now), takes into account the strength of a team’s offense, the strength of their opponent’s defense, and home field advantage. Those goal predictions are used to get the probability of each possible scoreline. A correction is made to the probabilities to account for the lack of independence between each team’s goals. Adding up across scorelines then gives the probability of each possible result- win, loss, or draw.
The model itself probably needs some tweaking. Just in terms of simple out of sample accuracy, it performs about as well as always predicting the home team will win. Dixon and Coles suggest further improvements that can be made, and those will probably be looked into for future iterations of the model. For now, though, this more straightforward version is perfectly fine, especially for an application as broad as playoff probabilities.
The code and data for the project can be found here - https://github.com/kcm30/2017-playoff-probabilities