The purpose of rating systems is to have an accurate and reliable way to describe the strength of competitors in sports and games. Elo model is the base for the various real-life applications of these systems and interest to forecast sport events has also shed academic interest towards this area. Although academic research has covered various of different sports, Elo model hasn`t been researched on an academic level in ice hockey context up to this date. This thesis presents and evaluates several applications of Elo rating models in National Hockey League context in order to evaluate their predictive power to forecast full time match results. The objective of the research was to find the most accurate application of Elo to rate NHL teams and to evaluate their predictive power against some benchmark methods. Emphasis of this paper is more on evaluating predictive power of Elo-based ratings with different error measurement, rather than using economic measures in order to determine if Elo models could be profitable in the betting market. Different applications of Elo rating models from previous academic literature are considered and evaluated to provide necessary scope. These include the choice of normal distribution against logistic distribution and using modified coefficient to better reflect the actual performance. Different error measurements are assessed to ensure that the predictive power is measured with wide enough perspective. Publicly available NHL data from season 2005-2006 up to season 2016-2017 were used in applying different models. Data were modified and validated in Excel before importing to R, which was used to build models and analyse results. In addition to Elo models, two naïve prediction methods, market odd probabilities and an accumulative rating system, adapted from International Ice Hockey Federation, were used as benchmarks. Results indicated that adding goal difference variable to adjust the rate of change in ratings did not provide enhanced results in NHL context when compared against the original Elo model. Normal distribution was speculated to be better in modelling performances in National Hockey League when compared against the logistic distribution. However, paired t-test on these models indicated that this difference was not statistically significant. Comparison results gave further evidence on Elo`s capability to model performances in sports. Elo models were able to outperform naïve predictions with statistical significance, which is further proof for the consensus amongst previous academic papers. However, results couldn`t provide statistically significant evidence to prove Elo-based probabilities to be inferior against market odds-based models. However, odds-based model was the best performing amongst models and benchmarks, indicating similar results to previous research on other sports. Overall, results indicate that properly constructed Elo system can be used in modeling performances in ice hockey and that Elo-based probabilities can be used as a considerable benchmark for future research on match result forecasting.
© Copyright 2019 Julkaistu Tekijä Aalto university. Kaikki oikeudet pidätetään. / Kääntänyt https://www.DeepL.com/Translator