On the main form click "Train" to pull up the Profile Training Wizard. Click "Next profile range" to select the range of profiles that will share the same profile weights. Click "Load all races" to load all races within the range you have visible in the Min and Max text boxes.
When the races are loaded two buttons appear to the left of the "Menu item to optimize for" label. Click "Auto" to pass through all profile weight combinations until there is no change. Click "One Pass" to make one pass through all the profile weights. Click "Save Profiles" to name and save the trained profile as all profiles within the range between Min and Max.
You may manually train individual weights by clicking the button above the column of the profile weight you want to train.
Training profile weights can find a set of profiles that are profitable for individual tracks but it is not as easy as simply clicking the train button and then saving the profiles. Statisticians know that any model that intends to predict the future must be tested against a different set of data than was used to develop the model. That is what we are about. We are developing a model to predict the outcome of future horse races based upon data derived from a history of horse races.
We can test the model we develop by using alternate days as the training data then using the flip side alternate days to test the model. Use every other day in the training data then use the days that are not in the training data for the test. We do this by configuring two lists of race days. One is FLIST.DAT, the training list. The other is ALTLIST.DAT, the testing list.
As an example, we might test the theory that it is good to use all zeros as a seed and train A to Win for each distance. Using the last few days at LAD we get this result in the training data. Then we click Helpers; Help; flip.bat to flip FLIST.DAT to the alternate days and get this result in the testing data. Not nearly so good. The test data with the default of all 10's was somewhat better.
Another way to develop test data is to set up the "Two Week Model". This model uses the most recent week for training, and the previous week for testing. You can also use the last ten days for training and the previous ten days as a test. This method might be slightly better than alternate days since we use all of the most recent days as a training model.
The training objective should be to make the test results look as much like the training results as possible. The closer we come to this objective the better will be our future projections. Future results will almost always look more like the test results than the training results.
The test results should also be better than results using the default profile of all 10's. Otherwise we gain nothing by training and may as well use the default. The min and max range of the numbers of the handicapped data file are scaled to favor the default profile and the Training Wizard will select the default if it does not find anything that produces better results.