Forecasting Methods for Horseracing

Author: Peter May
Publisher: High Stakes
Buy this book at Amazon

<< Previous Page   Next Page >>

Forecasting Methods For Horseracing

 

 

 

Figure 1.2: Classification patterns

Each letter occupies a subset of the squares on the 10x10 grid. Therefore it would be possible to use each square (or pixel) as an input to the system, with a used pixel represented by a one and a blank pixel represented by a zero. Naturally, the system would need to be capable of handling 100 inputs, given the size of the grid. While it would be possible for a neural network, trained on the sample patterns, to classify each correctly, such a system would not, however, be guaranteed to classify new (different) patterns very well. In other words it might not generalise well. Generalisation is naturally of great importance in classification and forecasting systems where the aim is to generate statisticalmodels of the data rather than systems whichmerely memorise historical patterns. As an example, a new pattern could consist of a very small' W located in the lower left quarter of the grid. To the human eye the letter is clearly a' W, however, to the system the arrangement could be meaningless. One possible solution is to extract features from the training data to use in the model. For this example these features could refer to the angle the lines of the letters intersect some defined axes. A (near) horizontal line would then indicate the letter 'L'. A second feature may relate the length of the lines to each other generating a set of ratios which could form another input to the model. Given this information the model is more likely to generalise well to previously unseen patterns.

 

17