Predictive Horse Race Handicapping Using Neural Networks

1
Predictive Horse Race Handicapping Using Neural
Networks

• Andrew Schurr
• Advisor Ralph Morelli

2
Goal

• This project proposes to use a trained neural
  network to predict the results of horse races. It
  will allow a person to bet at the track with the
  same level of skill as a professional handicapper.

3
Betting on Horse Racing

• On the surface, betting on horses is similar to
  other games of chance
  
• However, luck is usually not the dominant factor
• Closer to poker than roulette skillful analysis
  of variables gives the gambler an edge
  
• Fact better horses will usually win
• but how do you determine which horses are
  better?

4
Handicapping

• Handicapping is the art of selecting the best
  possible horse to bet on
  
• Handicappers look at available race data for each
  horse, such as
  
• Past racing times
• Post position
• Favorite track conditions (muddy, dry, ect.)

5
Handicapping

• Intuitive, data-driven process
• One race involves hundreds of pieces of data that
  need to be carefully weighed
  
• Handicappers use careful data analysis along with
  intuition to pick winners

6
Can a computer do it better?

• Computers are good at crunching lots of data
• But
• How do we mimic the intuitive process used by
  professional handicappers?
  

7
With Neural Networks!

• Neural networks simulate the structure of the
  human brain
  
• Made up of a set of interconnected neurons
• Neurons fire when stimulated
• But only when stimulation is above a certain
  threshold
  

8
Neural Networks Layout

• Data enters through input layer.
• Hidden layer processes the information
• Output layer outputs the result

9
Neural Networks

• Neural networks can learn
• Firing thresholds of the neurons adapt based on
  how often and how strongly they are stimulated
  
• After sufficient training, the network learns
  what output is expected from a given set of data

10
Example The XOR Problem

• Neural network is given a set of binary inputs
  0-0, 0-1, 1-0, or 1-1
  
• Initial output will be random (untrained)
• Training
• Input 0-0 or 1-1, output 0
• Input 0-1 or 1-0, output 1
• Hidden layer weights are adjusted through
  training
  
• Once trained, the input 1-0 will result in an
  output of 1

11
Uses for Neural Networks

• Good at predicting future events based on past
  data.
  
• Example
• Weather predictions
• Stock market fluctuations

12
Neural Networks for Handicapping

• Why use neural networks?
• Handicapping involves many variables
• Neural networks excel at representing complex
  relationships between variables!
  
• Neural networks require a large, accurate set of
  data for training.
  
• Data from past horse races are readily available
  in computerized format!

13
Defining the Input and Output

• Too many variables in horse racing to use them
  all
  
• Ten key variables will be selected as inputs
• based on research into current handicapping
  theory
  
• The output will be the probability of a given
  horse winning the race.

14
Internal Structure

• In neural networks effectiveness depends on
  structure
  
• How many layers? How many neurons?
• How do we determine an effective structure?

15
With Genetic Algorithms!

• Create a set of candidate network structures
• Random number of neurons, layers, weights
• Test each structure for effectiveness
• Keep strong performers, eliminate weak ones
• Cross-pollinate winners
• Mix neuron layers, weights
• Test new variants
• Continue until effective structure is found

16
Training

• The network is trained using existing race data
  from previous years
  
• Several thousand cycles
• Eventually, the network will learn what
  combination of variables makes a winning horse

17
Will It Work?

• Handicapping is difficult
• In the long term, almost everyone loses money at
  the track
  
• Simple neural networks have not been able to
  break even
  
• However
• A more sophisticated network may be able to do at
  least as well as a human handicapper

18
The Finished Project

• Standalone Java application
• Will have a fully-trained network
• Will use the Joone Neural Network Library,
  developed by Paolo Marrone.
  
• Features
• A file-loading system for easy entry of current
  horse input data.
  
• A GUI for set-up and interpretation of predicted
  results for each horse race

19
Project Timeline First Semester

• End of September Early October Research will
  be conducted on neural network variants,
  training, and construction, as well as on the
  current science and theory of horse race
  handicapping. Potentially important racing
  variables will be selected and pruned into a
  usable set.
• Mid-October Early November A large set of
  previous races will be located for use in
  training the network. Research into the class
  structure and usage of Joone will be conducted. A
  simple prototype of a neural network will be
  constructed using Joone.
• Mid-November December The previous race data
  will be formatted, and the ability to load it
  into the neural net program will be added. The
  prototype will be expanded to include the actual
  race data and several of the key variables.

20
Project Timeline Second Semester

• January February Research into using genetic
  algorithms will be conducted, and the
  capabilities of Joonegap will be explored.
  Preliminary use of genetic algorithms will be
  used to improve the structure of the network.
• March The full set of race variables will be
  integrated into the network. Genetic algorithms
  will be used to optimize the full network
  structure. Training on the full data set will be
  conducted until optimum efficiency is reached.
• April May The user interface, race history
  file loading system, and application package will
  be created and polished. The network will be
  tested on current race data to determine its
  ultimate effectiveness.

21
Sources

• Neural Networks for Fun and Profit, Bret Halford,
  http//csel.cs.colorado.edu/cs3202/papers/Bret_Ha
  lford.html
  
• An introduction to neural networks, Andrew Blais
  and David Mertz, IBM developerWorks,
  http//www-106.ibm.com/developerworks/library/l-ne
  ural/
• Expert Prediction, Symbolic Learning, and Neural
  Networks An Experiment on Greyhound Racing, H.
  Chen, P. Buntin, L. She, S. Sutjahjo, C. Sommer,
  D. Neely, http//ai.bpa.arizona.edu/papers/dog93/d
  og93.html
• Using Machine Learning To Predict the results Of
  sporting matches, Michael Baulch,
  http//innovexpo.itee.uq.edu.au/2001/projects/s348
  234/thesis.pdf