Evaluating machine-learning algorithms and strategies in sports betting context

Abstract

This thesis studies various different machine learning algorithms, and their performance in a sports betting context. In addition to this, some strategies for choosing the bets to take are evaluated. As a context, player propositions, or player props are the market studied. This is done because of the lack of studies, increased popularity of said market, and potentially higher edge on the bettor’s side.

The testing is starts with getting the data and preparing it for the algorithms. This includes filtering, creation of new variables and feature selection. Three regressionbased algorithms, along with three classification algorithms are chosen. Of them, multiple iterations are created and tested with different datasets. The main objective is to find a system that would consistently beat the sportsbooks by gaining profits in a simulation that is run in a realistic setting, thus showing the inefficiency of the market. In addition to that, the best bet picking strategy and algorithmic techniques are looked into.

The objective set was reached: using correct strategy, multiple iterations were able to make noteworthy profits in the simulation, using 2023-24 season as the test. Not only were they very profitable, but also stable, which reduces the risk of bankruptcy. Other findings included the superior performance of regression-based algorithms compared to the classifiers, and the essentiality of finding a good strategy for picking the bets. For the strategy, one based on maximal difference from the sportsbooks’ prediction was used.