I think the proper scientific approach to modeling is to first try your best to understand the phenomenon you are studying. In particular, if a statistical model appears warranted by observed randomness, then one should first try to understand, as much as possible, the mechanism producing the apparent randomness -- and that should naturally lead to an appropriate model.
Unfortunately, this does not seem to be the way modeling is usually done. It seems that usually statisticians just choose a model or a class of models among those they are familiar with and then try to fit it to the data. At best, they would afterwards do something like what is described in Section "Evaluating the quality of forecasts" of the Wikipedia article Autoregressive model:
The predictive performance of the autoregressive model can be assessed as soon as estimation has been done if cross-validation is used. [...] if the predictive quality deteriorates out-of-sample by "not very much" (which is not precisely definable), then the forecaster may be satisfied with the performance.
There is no example in that Wikipedia article of how or where an autoregressive model may naturally arise, and I cannot imagine such an example either. So, at this point, this model seems to me entirely contrived (I would be glad to learn that I am seriously mistaken about this).
As for the simplicity criterion, I think it is applicable when one wants to choose one of a number of models with all relevant merits similar to one another and differing mainly in the degree of simplicity.