My final project for a Machine Learning course exploring if daily fantasy salary data can be exploited to identify profitable opportunities for sports gambling. This was not a comprehensive approach to the problem, rather a first investigation into it.

The Jupyter notebooks highlight the process that I took to arrive at the final result of my project for a school project in Machine Learning at Georgia Tech. The numbers on the notebook files indicate the order in which they were prepared and should be looked at to evaluate the work. I did not publish the data that I used for this project as some of it was from a paid source, though I hope this can introduce some of my capabilities of data wrangling and analysis on an interesting subject.

I have also attached the paper that I submitted in the Fall 2022 semester, which has already been graded, therefore, no concerns with sharing it here. Since submitting the project, there are many things I would do differently or add to the project, but the biggest one is that I would have done much more exploratory data analysis to identify relationships in the data and better guide my model selection process.

In the paper, I discuss much more about the process and the approach I took, and share the following conclusions:

Overall, I am concluding that daily fantasy salaries are not enough, in conjunction with odds offers, for finding value in sports gambling markets using the identified methods and outlined gambling strategy. One interesting finding was that the LightGBM model consistently did find the salary data as the most important in its models. For all three objectives, wins/losses, spread results, and over/under results, four out of the top five features in importance were salaries of players from either team. This contrasts with XGBoost which always had the home and away team moneyline odds as some of the most important features. From this, I do believe that the daily fantasy salaries prove they provide some amount of value in assisting with identifying results, but the casinos are not “giving away” immense insight to untapped value by publishing these salaries.

There are a few ways I believe this could be expanded upon to potentially find value in these gambling offers. First, I would incorporate more data as potential training features. Advanced player metrics that are published could provide value to help identify the results more successfully. Second, I believe that ensembling some of these predictions could help. Perhaps connecting the results of the win/loss result predictions with the daily fantasy salaries could help the spread predictions. Third, a regression approach could be used with the spread and over/under results. For example, the models could predict a point total for each team given their players’ salaries, or a total for the game, then compare that to the offer from the casinos and make the gambling decision on that. Finally, there could be other decision criteria for the decision to bet or not bet on an offer. With the classification models providing predicted probabilities, the intuition is that if the predictions are higher than the implied probability, it is good value. This intuition does not hold up in general, and other strategies, with new selection criteria, need to be identified. For example, on moneyline bets, only betting on results with a predicted probability over 0.7, regardless of the odds – or on spread bets, if a game has a spread of more than 7 points, do not bet on it.