Predicting the NHL playoffs with Poisson regression

Abstract: Using historical data from the past two seasons of the National Hockey League, three different prediction models based on Poisson regression are developed. The aim is to determine whether taking into account the recent form of a team as well as data from how they have previously performed against their opponent can help make better predictions of how many goals they will score against this opponent and thereby calculate the likelihood of each outcome. The three models are evaluated using different measures, for example comparing the odds yielded by the models against the odds of bookmakers. Different ways to account for recent form are discussed. The paper concludes that using recent form and head-to-head data will indeed improve predictions.