There are a lot of reasons for that.. It's not really a NN problem, its how they're used, what data they're fed, avoiding overoptimization, etc, ensuring that your results 'make sense' (with walkforwards, etc)
In my own NN research, I get much better results trying to make my nets as stupid...