It is interesting that some trading models are "built-up" from domain specific ideas, a bit like in physics, while other models a created out of data with little domain specific knowledge. Google now translates text with algorithms that have learned only from existing text and have a very abstract notion of similarity in sequence of words. Likewise, people use neural networks in financial models. With these models it is often not easy or even possible to explain "how" a result was found. When decisions need to made in milliseconds it is not possible to double check with humans before an action is taken. The game is then to make more money than to lose it while not even really understanding the why of the actions taken.