Improving NHL Player Ability Ratings with Hazard Function Models for Goal Scoring and Preventio

• Activity Number: 64
• Type: Topic Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Sports
• Author(s): Andrew C. Thomas*+ and Stephen Ma and Samuel Ventura
• Companies: Carnegie Mellon University and Carnegie Mellon University and Carnegie Mellon University
• Keywords: baseball ; simulation ; bernoulli

Abstract

The public availability of data on NHL games now includes information on which players were on the ice for each team at any point during the game, and can naturally be divided into ``shifts''. In terms of goals scored during these shifts, the data are considerably sparse; the vast majority of shifts, on the order of one minute long, contain a goal by either team, making standard regression tools difficult to implement in assessing the contribution of each player to any goal scored for either team, including the weighting problem provided by shifts of different lengths. Additionally, the very nature of shift changes, in which player changes mean that the chances of scoring a goal are reduced, negate the use of a standard Poisson process. We turn these disadvantages into strengths by modelling the scoring rates for each team as semi-Markov processes, whose hazard functions depend on the players on the ice for each team. We show that this method not only produces more stable estimates of player ability (including goaltending skill), it allows us to identify possible interaction effects between pairs of players who may be better suited to playing with each other