Combining classification algorithms to predict the winner of each professional baseball game
Kyle Johnson
Blog post: https://kylejohnson363.github.io/sourcing_mlb_data
The ability to predict the future, even just slightly better than a coin flip, can be enourmously lucrative. Without having a crystal ball, the next best thing that we can do is harness the power of large datasets to find hidden patterns that can be used to give a slight edge in making large amounts of predictions. Baseball is perfectly suited for this because virtually everything that happens is quantifiable and repeats hundreds of times per game and each game is repeated thousands of times per year. The goal of this project is to use machine learning techniques to make predictions about Major League Basebal games in such a way that is better than the Vegas book makers. Being able to predict 70% of games correct is of no use if Vegas also predicted those same games correctly; in order to have a useful model, I must create one that consistently makes money when betting against Vegas bookmakers.
Please see the notebook titled "Summary_Start_Here" for a detailed road map through this project in order fully understand the process.
The data for this project was sourced from MLB Advanced Media's API, baseball-reference.com and sportsbookreviewonline.com and then pre-processed into a useful form. Four classification models were then created and optimized, which then used a voting procedure to make a final prediction.
The performance benchmark for this project is the predictions that the Vegas odds-makers create. If the created model can make money by betting against Vegas, then we know the model has added value. Below is a graph showing the relationship between the confidence that Vegas has in a prediction vs the percentage of time that prediction is correct. The orange and blue lines are quite correlated, meaning that Vegas is quite good at predicting games, which makes sense because they would be out of business very quickly otherwise.
The final model was able to outperform the Vegas odds-makers with statistical significance in both the accuracy of picks and the return on risk generated by placing bets on the games predicted.
Below is a visualization of a simulated betting account performance over the out-of-sample data starting with $1,000.
-I was able to create a model that predicts MLB games more accurately and more profitably than the Vegas odds in a statistically significant way. I did this by querying data from several online baseball databases and then optimizing several different classification models, before combining them to vote on the outcome of each game.
-Oddly enough, it seems that always betting with the Vegas odds is a profitable strategy but using the model created in this project is potentially almose twice as profitable. This tells us that Vegas is good at predicting MLB games but there are still inefficiencies that can be exploited.
-
Use more types of data (new and highly advanced statistics) and more games from previous seasons.
-
Optimize the number of days in the "recent" category of statistics.
-
Automate the process of gathering the necessary data for today's games and publishing a report of which games to bet on.
-
Create "minor predictions" such as runs to be scored or allowed and feed those predictions into the classification model.