Even then, and as I wrote in another post, a custom trading NN might be working a strategy which is fine under normal market conditions whilst leading to massive losses if those conditions change (i.e. "picking nickels in front of a steamroller") and because of the black-box nature of how Neural Networks work and their tendency to end up with the outputs being very convoluted derivations of the inputs (I expect even more so in Markets, were the obvious strategies that humans can easilly spot have long been arbitraged away, so any patterns such an NN spots during training will be so convoluted as to not be detectable by most humans), nobody will spot the risky nature of that strategy until getting splattered.

Neural Networks working in predicting market movements are, unlike a predictive text keyboard or even an automated train driver, not operating in a straightforward mainly non-adversarial enviroment.