When I first started designing algorithms to trade stocks, I was fascinated with all things related to artifical intelligence and truly believed that this was the way forward in developing an 'all seeing' formula for success.
My early attempts at algo-trading were notable for the complexity of the models I built and, in due course, the realisation that I was merely adjusting variables and parameters in order to 'fit' my model more closely to a back-tested curve. It was only after reading Ernie Chan's great book Quantitative Trading did I realise that it is more important to work on simplifying one's model that trying to make it all singing and all dancing. Strangely, the hardest thing about model development is to simplify it. You know you are getting close to a good model when adding one more variable or parameter moves it away from the optimum. I have found through 'gut feel' rather than empirical evidence, that anymore than 4 or 5 variables and the same number of parameters makes a model less than optimal.
A simple model will allow it to work in more varied market environments, with more resilience and for a longer time. When you are reading about the complexity of the Wall Street quant models, always remember that simplicity is the friend of the amateur.