validity of hypotheses unnecessary when cross-validation is used? I had recently a big discussion with my collegues regarding the need for hypotheses validation in the context of cross validation for model comparison.  
Here is the case: I'm using a repeated cross validation procedure (50 repeats, 2 folds) to compare prediction capabilities (MAE & MBE) of a dozen types of models (including linear and machine learning). I was told that I still have to validate hypotheses (normality of residuals, heteroscedasticity, colinearity, etc.) even in case a linear model comes first. 
My opinion is rather that whenever a model hits first in prediction capability, we can disregard any violation of assumptions of linear models. 
Thanks for your help, and if possible an link to an article backing your opinion (i couldn't find any)
 A: I can only offer an opinion rather than something backed up by publications. Firstly, to some extent I agree with a "whatever works works" philosophy. If a model does well in predicting in practice, then it seems like a good model for the particular application. Secondly, one needs to be a bit careful with relying on that argument too much. E.g. when some major assumptions are violated and you extrapolate even slightly beyond the kind of data you have worked with before things can go really, really wrong. Like when you can approximate the probability of death yes/no very well for a narrow range of some predictor with a linear model, but then at slightly higher values the model comes up with a probability <0 or >1 of death. Some assumptions are possibly more vulnerable to this sort of thing than others. 
E.g. using an outcome distribution that does not correspond to the actual outcome range (e.g. normal distribution for 0/1 binary outcomes, normal for 0,1,2,3,4,... counts, binary 0/1 when there is in fact an underlying time-to-event process etc.), some kind of functional form for continuous predictors when you predict to the edge (or beyond) of the previous data range and Weibull shape parameters that are estimated on a certain time range used to extrapolate beyond that duration to name a few things are things I might worry about. You can probably find plenty of publications and examples for that (e.g. Challenger O-rings). Personally, I would perhaps not be so worried about normality of residuals and colinearity for prediction tasks, if the model performs well. Heteroscedasticity could of course be an issue when you extrapolate from previous data.
