What we're doing at the FHI is acting like a machine learning problem: splitting the data into a training and a testing set, checking as much as we want on the training set, formulating the hypotheses, then testing them on the testing set.