A new approach to momentum: a novel framework to evaluate momentum in sports

Abstract

Momentum often faces skepticism as a discredited phenomenon in sports. While some contradictory evidence has been found, we provide new insights by quantifying this phenomenon. Momentum literature often relies on proving the dependence or independence of sequential outcomes. However, we argue that this approach is not appropriate due to the large body of literature showing that sports are highly subject to randomness. If sports are subject to randomness, we should focus on what leads to winning rather than not winning itself. Here, we engineer momentum-based features that quantify a team’s linear trend of play in several underlying performance indicators and compare the predictive power of these features to more traditional frequency-based features when only using a small sample of recent games to assess team quality. We developed a complete data pipeline that allows us to compare the effects of momentum on multiple sports. We found evidence of momentum in the NHL and the five major European football/soccer leagues; however, we could not find evidence of momentum in the NBA. The differences between these sports indicate that momentum could be a sport-specific phenomenon. We also found that the combination of momentum-based and frequency-based feature sets, along with more powerful machine learning techniques such as random forest, led to very promising results. In the future, we believe that by combining these two feature sets with proper hyperparameter tuning and feature selection, better pre-game prediction models can be created that accurately capture both short-term and long-term quality of play.