The customers I build models for insist on interpretability and robustness, which deep learning just doesn't give them right now. Actually just got a conference paper out of a modification to a classical method, which was kinda fun.