Slide 1

An ANN Approach to
Predicting NHL Game
Outcomes
Erik Everson
ECE 539
12/10/01
Problem Description
• NHL Background
– Began in 1917 with 5 teams
– Today one of the most popular sports
in North America, with 30 teams
• Lack of accurate prediction system
– Most systems adapted from other sports
– Unique aspects to hockey
Data Acquisition
• Data culled from www.hockeynut.com,
wrote PERL script to process html files
• Used results of 3 seasons for training
data (1997-2000)
• Testing data set from 2000-2001
regular season
Data Processing
• Basic MLP inputs
– Wins, losses, goals for, goals against
• Additional MLP inputs
– Home or away game
– Tie games
– Winning or losing in overtime vs.
regular play
Data Processing (continued)
• Three possible outputs
– Win, loss, or tie
– Did not factor any overtime information into
output
• Schedule abnormalities
– Teams play 82-game season over 7 months,
games occur on all days of week
– Four NHL teams introduced since the 19971998 regular season
Results Pending