If you believe the odds to win are right at the track you will see sometimes that the payouts to place and show are really high in comparison, sometimes almost as high to win but you get 2x or 3x the chance to get a payout.
You are not looking for a good horse, you are looking for a good bet and you will sit out many races until a good opportunity comes up. Maybe 2 or 3 races a day have a bet with odds so favorable to overcome the track’s rake.
Most people who work with machine learning and AI follow the herd using certain methods of evaluation that don't necessary lead to working systems. I'm more familiar with applications to business, text analysis, and a bit about finance such as algo trading.
This is one of the missing links
https://scikit-learn.org/stable/modules/calibration.html
if you have a prediction that says "this is in class A" that's not very useful in itself. If your predictor is calibrated and says "there is a 55% chance that this is in class A" you probably don't want to take action that prediction but if it says "there is a 97% chance this is in class A" you will take action.
IBM Watson won at Jeopardy because it was calibrated and could make a rational decision of whether or when it should hit the button.
From an algo trading point of view the other thing you need to turn predictions into actions is
https://en.wikipedia.org/wiki/Kelly_criterion
I bet there is some similar way to turn predictions into actions based on control theory. Speaking of betting there is a practical application of this in this book
https://www.amazon.com/Dr-Beat-Racetrack-William-Ziemba/dp/0...
which is a gambling system that really makes money. The book is not very mathematical, Dr. Ziemba has written a lot about hedge funds, algo trading, etc. and probably in his body of work there is something that covers the same ground and is more mathematically rigorous.
The right way to think about it is "finding a good bet". A few times a day you find a situation where a place or show bet is terribly mispriced compared to win. The tote board gives highly accurate odds to win but sometimes you see place or show paying almost the same as win except you have two or three chances to win instead of one. In a case like that if you believe the win odds are fair, place or show is a slam dunk and you can really make money that way.
See https://www.amazon.com/Dr-Beat-Racetrack-William-Ziemba/dp/0...
It's essential if you want to:
* make money by counting cards at Blackjack (the odds are a function of how many 10 cards are left in the deck)
* make money at the racetrack with a system like this https://www.amazon.com/Dr-Beat-Racetrack-William-Ziemba/dp/0...
* turn a predictive model for financial prices into a profitable trading system
In the case where the bet loses money you can interpret Kelly as either "the only way to win is not to play" or "bet it all on Red exactly once and walk away " depending on how you take the limit.
The most important concept to understand is
https://en.wikipedia.org/wiki/Kelly_criterion
That is, you make money trading if you have a predictive model and use Kelly to turn that into bets.
See this book
https://www.amazon.com/Dr-Beat-Racetrack-William-Ziemba/dp/1...
for a strategy that still works and an example of "predictive model + Kelly". (e.g. that strategy still works because anybody who masters it and would scale it up goes on to Wall Street and plays for bigger stakes)
