Do you have any references for the approach that you recommended? You're saying that differentiation (not even just numerical differentiation) is "not something you want to apply to real-world data" ... This is surprising since differentiation obviously is used in the real-world in some very critical applications.