We're not expert handicappers. However, over the years of giving technical support, we've picked up a few things. So, here's what we know:
It's probably unrealistic to hope to build a network which picks the winners for all nine races each day. Our successful customers specialize on a particular type of race, e.g. two-year-olds, harness, quarterhorses, etc.
You'll need history data for a couple hundred races which are in your chosen class. This data can be typed in from old racing forms, but you'll probably prefer to buy it on-line. Data services advertise in your racing form. One such company is BloodStock.
The basic idea is to show BrainMaker data for each horse which will help predict it's performance. First, eliminate irrelevent data: the horses name, for example. Exactly what is relavent is up to you, but here's some stuff other customers have used:
last practice time
last practice time / track record for this class
finish position for last three to five races
seconds behind first place for last three to five races
last race position at first post, at second post, at third post
last race speed at first post, at second post, at third post
Jockey's overall win/loss record
Jockey's win/loss record in last ten weeks
Stable's overall win/loss record
Stable's win/loss record in last ten weeks
finish position on last bad track
finish position on last good track
These are not so much the "right" data to show the network, but rather suggestions of particular types of data. Here's some things to keep in mind:
Horses have psychology. Some are very fast, but cannot lead. Some practice poorly, but are very competitive and are always contenders. Also, like people, horses have good days and bad days. It's worth telling BrainMaker how the horse looked at the gate if you can get that information for the historic races.
Some horses are more powerful and will tend to do better in short races or on bad tracks (Carl Lewis looks like a body builder, and has incredible acceleration and speed for 100 yards). Some horses have better aerobic conditioning and will tend to have better finishes in longer races (your typical marathon winner can barely do a pushup, and is really not so great at 100 yard dashes). Speeds at the 1/4, 1/2, and 3/4 marks will reflect these differences: a more powerful horse will be heavier and leave the gate quickly, build speed quickly, but have little left to hold off a late challenge in a longer race. A better conditioned horse will perhaps be lighter and not as impressive out of the gate or at the first post, but will build speed throughout the entire race and finish very strongly, allowing it to make a strong challenge at the end of a longer race. These considerations are worth remembering when comparing 5/8, 3/4, and mile races.
Jockeys, trainers, and owners count. Some combinations of these people seem to consistantly get 105% out of their horses, and some seem to consistantly get 95% out of their horses.
It's extremely difficult for BrainMaker to learn to predict the finish order of all nine horses. Besides, you don't care about this: only the first three count. So whenever your data indicates a position, 1,2,3,4 is probably enough, where "4" means "worse than 3rd". More generally, don't let the slow horses swamp your data: if you have one horse running 10 seconds off the pace, this makes all the other horses running 1/2 - 1 second off the pace look the same. It's better to thrown out the slowpoke, or edit his time to something less pitiful, like 3 seconds off the pace.
Whenever available, seconds behind or lengths behind the leader is better than position. Running in 2nd place 12 legnths behind the lead horse is much worse than running in 6th place 1 length behind the leader.