We make a unique contribution to momentum research by proposing a way to quantify momentum with performance indicators (i.e., features). We argue that due to measurable randomness in the NHL, sequential outcomes` dependence or independence may not be the best way to approach momentum. Instead, we quantify momentum using a small sample of a team`s recent games and a linear line of best-fit to determine the trend of a team`s performances before an upcoming game. We show that with the use of SVM and logistic regression these momentum- based features have more predictive power than traditional frequency-based features in a pre-game prediction model which only uses each team`s three most recent games to assess team quality. While a random forest favors the use of both feature sets combined. The predictive power of these momentum-based features suggests that momentum is a real phenomenon in the NHL and may have more effect on the outcome of games than suggested by previous research. In addition, we believe that how our momentum-based features were designed and compared to frequency-based features could form a framework for comparing the short-term effects of momentum on any individual sport or team.
© Copyright 2024 International Journal of Computer Science in Sport. Sciendo. All rights reserved.