Model Structure

One network has been trained for each of the four main race classifications: Juvenile Non-Handicaps, Juvenile Handicaps, Non-Juvenile Non-Handicaps, and Non-Juvenile Handicaps. The data these networks use include all the basic information such as weight carried, days off the track etc, plus running style assessments, Peter May's unique speed figures and price movements.

Model Accuracy

The easiest way to show how accurate these models are is graphically.  The graph below presents the output of the network, in probability format, along the horizontal x-axis, and the actual win rate for these horses along the vertical y-axis. So the ideal model would be shown as a straight line running through the points were the x-value equals the y-value.  Such a model would have a 50% win rate, for example, for all horses assigned a 50% chance of success. As you can see from the graph the neural networks have achieved an impressive degree of accuracy and can quite confidently be expected to forecast the chance of success of each runner to a high standard of precision.