Browsing by Author "Remander, Sebastian"
Now showing 1 - 2 of 2
- Results Per Page
- Sort Options
- Comparison of classification methods in predicting NHL game event outcomes
Perustieteiden korkeakoulu | Master's thesis(2021-01-25) Remander, SebastianPrior research in predicting success in sports consider hockey to be one of the most challenging domains, due to relatively low prediction accuracies. Hence, research on utilizing hockey game predictions in sports betting is often only briefly discussed, tested with a simple fixed strategy, or left as future work. To the best of my knowledge, no other research has studied a system that combines National Hockey League game predictions with the Kelly Criterion betting strategy and simulates betting over different NHL seasons. In this thesis, that system is implemented and used to simulate betting for four common bet event types: Moneyline, Threeway, Over/Under and Asian handicap. The purpose is to predict NHL games with machine learning methods and to study how the predictions fare against the odds markets. Different predictor models, domain-specific feature extraction techniques and fractional Kelly Criterion strategies are studied and compared. Moreover, the performance of the system proposed is measured at each game during the season, rather than applying traditional machine learning train/test splits or k-fold cross-validations that are not well suitable for chronologically processed game data. The four latest NHL seasons are analyzed, which enables a comprehensive evaluation of the capability of the system. The highest accuracy is recorded by random forests for the Asian handicap predictions, while the best return on investment in the betting simulations is achieved using random forests for the Over/Under events. The largest profits are obtained with the most risk tolerant betting strategies; they also cause the largest and inevitable losses for event types with too inaccurate predictions. The risk averse strategies minimize losses and can even provide financial gains for events that are unprofitable with other strategies. However, risk averse strategies generate inferior profits when success is guaranteed. The results show significant differences between the event types studied, in both accuracy and betting simulations. On the other hand, accuracy variation is negligible for different predictor model configurations within individual event types. However, small differences in accuracy result in a relatively large variation in the betting simulation. The most significant impact in the betting simulation results originates from the wide selection of different fractional Kelly Criterion strategies studied. - Henkilön massan määrittäminen systeemikomponenttien avulla
Sähkötekniikan korkeakoulu | Bachelor's thesis(2015-08-27) Remander, Sebastian