Predicting NBA player bets using machine learning

Abstract

Sports betting is known as the activity of attempting to predict the outcome of a sports event and placing a wager on said outcome. With the rise of the Internet and e-commerce, a new type of betting has been popularised, online sports betting. Bookmakers create their odds using machine learning and mathematical models that use different data to use the odds that will maximise their profits. Although bookmakers likely have extremely accurate models for predicting sports event outcomes [1], creating a model that would outperform theirs is theoretically possible. Thus, if a model can outperform a bookmakers model, then it can predict sports events outcomes at a higher accuracy, meaning that if a bettor had to consistently place bets from this model over a lengthened period of time they will make a profit. In this FYP, we propose to create a model that can rival that of the bookmakers. We chose to focus on a specific type of bet in the National Basketball Association (NBA) known as player props. The nature of these bets is to bet on a specific player’s performance as opposed to a team winning or losing a game. To accomplish this task we needed a dataset large enough to train a machine learning model. Unfortunately, there are no publicly available datasets, prompting us to create our own using web scraping. We also needed a dataset which included player performances. Luckily an NBA stats Application Programming Interface (API) is publicly available which aided us to gather the data we needed for training. We then decided to use three models which can use sequential data since they will be trained on the previous performances of a player, these being: recurrent neural network (RNN), long short-term memory network (LSTM), and a transformer. Once each model was trained and hyperparameter tuned, we used each of them with different betting strategies to see what could give a bettor an advantage when betting. The final results showed that, overall, the transformer was the most accurate, followed by the LSTM, and lastly the RNN. Although the transformer was the strongest, it was still not capable of toppling the accuracies of the bookmakers’ models and was not able to be profitable when betting. However, when using a metric called peak, which measured the highest balance amount during a betting simulation, the transformer reached heights of seven times the initial balance. Indicating that it is possible for it to go on a streak of hits which is profitable.