MLB Sports Betting Predictor

Maching learning algorithm used to gain an advantage against sports betting books

This repo is designed for one to be able to gather all of the necessary statistics and train a neural network to come up with a more educated guess on the winners of MLB regular season games, for the purpose of profits against the sportsbooks.

Libraries / outside data used

• Past odds gathered from database on website: https://sportsbookreviewsonline.com/scoresoddsarchives/mlb/mlboddsarchives.htm

• MLB Data API used to gather player names and hand-dominance (using requests and json python libraries)

• Python wrapper of MLB API created by Todd Roberts to gather player statistics; link: https://github.com/toddrob99/MLB-StatsAPI/wiki (pip install MLB-StatsAPI)

• The neural network relies on Tensorflow with Keras, so visit the Google Tensorflow website on how to get the dependencies necessary to create your own model

• Need to add your own API keys to the keys/keys.py file. Name the Twilio key "TWILIO_KEY" and auth token "TWILIO_AUTH"

To run the program

1. Run the main.py program to gather statistics necessary

Navigate to directory of repo and run:

$ python main.py

which will take a few hours to complete. Program gathers all player statistics from seasons 2014 - 2019 (with the exception of 2016 due to data error in used database) to give neural network five seasons of data to train with.

Data gets saved in .pickle files so user does not need to run long program each time it needs to gather data. Make sure to delete pickle files each time you run main.py program so you don't accidentally overlap data in selected years.

2. Run the magic.py program to train your neural network

$ python magic.py

This program will prompt you to pick a loss function to train your neural network.

• win_loss will train network to optimize for predicting moneyline bets effectively
• spreads_loss will train network to optimize for predicting spread bets effectively
• ou_loss will train network to optimize for predicting over/under bets effectively

The choice is up to you based on your betting style, however moneyline bets have proven to be most lucrative with this model (boasts around 62% accuracy).

A .hdf5 file is created after training, which allows you to not have to train model each time you wish to predict.

3. Run the precition_creation/current_games.py program in its directory every day to gather current season statistics (use cron to run current_games.sh bash script in bash script directory each day)

$ python current_games.py

This is necessary so the model can use relevant statistics to base its predictions for future current-season games. Make sure it runs every day of a regular season.

3. Run prediction_creation/predict.py in its directory on future games to get outcome prediction

$ python predict.py

Currently I have this running hourly during the day to check if lineups have been created for games, then I have it set to text my number (configure Twilio settings in send_sms_skeleton.py file) if any predictions are above my confidence threshold, to maximize winning probability and profitability, which can be changed at the top of the predict.py file.

The model uses individualized statistics for the idea of better accuracy. The statistics used to train and predict are as follows:

Away stats first, followed by home stats:

• Current team win percentage
• Teams win percentage over last 10 games
• P-Expectation of team over the last 10 games, with the idea of tracking the teams hot/cold recency factor
• Batter stats, including batting average of lineup, on base, slugging, OPS, average home runs/strikeouts per game, etc.
• Same stats but only taken over last 10 games of lineup
• Pitching stats, beginning with starting pitcher ofllowed by bullpen, including strikouts per game, ERA, WHIP, Ryanicity, earned run ratio, etc.

The model will spit out the name of the teams playing each other in the format:

AWAY TEAM vs HOME TEAM
ml: AWAY ODDS, HOME ODDS
spread: AWAY ODDS, HOME ODDS
o/u: OVER ODDS, UNDER ODDS

The model only shows the odds of the favorite for each bet (ml, spread, o/u), since the odds for the other team is simply the negation of the negative odds. We only care about winners, so the winning predictions are the only ones shown.

Backtesting

You can go into the tests/ directory to make sure the data backtests properly using

python3 backtesting.py [CONFIDENCE_VALUE] [AMOUNT_PER_BET]

where confidence value equals how sure you want the models to be before you bet (displayed in American odds format), and amount bet is how much you would want to bet.

Let's make some money!!