Were there any studies which checked the accuracy of neural network predictions of greyhound racing results, compared to a human expert? Would it achieve a better payoff?
In greyhound racing (or horse racing), there is no definite underlying pattern that can be associated with the outcome of the race. There are far too many variables to record and code as features. Most of which cannot be accessed by the public. This includes eating, sleeping, and training patterns. Furthermore, there are variables that cannot be readily quantified, such as the trainer's techniques, training effort, health history, and genes. A mere history of racing results and age won't be that helpful.
A neural network can only be as good as the features that are used to represent the instances. If the features don't capture the necessary characteristics of the instances that are associated with the problem, then the learner cannot generalize to predict the real world outcome.
I would strongly recommend that you check the book "The Perfect Bet" by Adam Kucharski. It does not mention technical methods such as neural networks but it gives a good history (and very nice stories) on what people had done on that field. It gives you the notion that in order to get achieve a better payoff, your goal is not actually making a better prediction but choosing the better options by considering what other players are doing. If you ask why it is not just the better prediction, the answer is that there is always a balance of risk and payoff and since you cannot find a 100% guaranteed way of prediction, you will have to balance risk and payoff in order to gain in the long run. In addition, although theory suggests that you can make a good prediction by gathering all variables, this is not really possible in practice. Thus, human assistance and observing what others are doing is used as a way to improve mathematical predictions and possible payoffs.