Predicting MLB Betting Lines with Neural Networks

This project leverages TensorFlow's Keras API to compile, train, and use a neural network to predict betting lines in a baseball game. Training data is gathered from ESPN (the state, or inputs) and DraftKings (the prediction, the outputs).

Simply put - for any point in a theoretical game of baseball, this model predicts what the standard betting lines for the game should be, and thus, can predict a winner at any point before or during a game.

How to run the model

1. Download and unzip one of the pre-trained models in the download section below.
2. Replace the value of the nn_model_path variable in the ui.py with the path of the folder within the unzipped folder.
3. Install required dependencies:
   1. tensorflow: python -m pip install tensorflow
4. Run ui.py to run the model and GUI!

Model Inputs (the State)

The model considers the following 16 inputs when predicting betting lines, in this order:

• Away team record, as a percentage (i.e. 0.8 if the team is 8-2, meaning the team won 8 out of their 10 games total)
• Home team record
• Number of runs the away team has
• Number of runs the home team has
• Number of hits the away team has
• Number of hits the home team has
• Number of errors the away team has
• Number of errors the home team has
• The current inning - i.e. 1.0 for first inning, 2.0 for second, etc. And "0.0" would mean the game is yet to be started and the betting lines are a pre-game line.
• Top or bottom of inning? Top = 0.0, Bottom = 1.0
• Number of outs
• Number of balls in the batter's count
• Number of strikes in the batter's count
• Is there a runner on first base? No = 0.0, Yes = 1.0
• Is there a runner on second base? No = 0.0, Yes = 1.0
• Is there a runner on third base? No = 0.0, Yes = 1.0

Model Outputs (the Prediction)

The model predicts four distinct betting lines, in the following order. You can read more about what each of these lines mean here.

• The Run Line - the "point spread" between the two teams.
• The Total Line - the prediction for what the under/over would be for the combined number of runs in the game.
• The Away Team's Money Line
• The Home Team's Money Line Using the odds above, particularly the money lines, we can use these to calculate the implied win probability for either team.

Model Downloads

Name	Parameters	Description
model4	7,469	Trained on ~5,300 examples. Warning, trained on data that likely contained errors.
model5	378,274	Trained on 5,473 examples. Warning, trained on data that likely contained errors.
model8	378,274	Trained on 14,594 examples

Training Data Downloads

These are .jsonl files. Each line is self-contained JSON object with both the state (game scenario) and real-world observed betting line information.

Number of Examples	Size	Description
5,473		Warning, likely contains errors.
16,515	2 MB	Warning, likely contains errors.
14,594	1.7 MB
41,037	4.8 MB

In this Repo

This repo contains the following programs:

• A program for capturing training data from ESPN and DraftKings, written in .NET 7
• A python script to assemble, compile, train, and save a TensorFlow Keras neural network
• A python program to leverage a pre-trained model to allow you to predict for various scenarios, leveraging a pre-trained model

Future Areas of Improvement

• When a batter walks, ESPN will mark it with 4 balls in the count AND a man on second temporarily. If there are 4 balls and a man is on, count it as 0 balls.