Controlled experiments must be performed in order to discover what works best for a given classification or regression predictive modeling task. 1

...

A modern alternative is to consider the selection of data preparation, learning algorithm, and algorithm hyperparameters one large global optimization problem. 1