I recently posted about a reply to a PNAS paper in which we show that, contrary to the claim in the original publication, the “true” model does by no means delivers worse predictions for chaotic population dynamics than time-series methods without any specific mechanism underlying, provided appropriate statistical estimation methods are used. To be clear, I use “true” as a shorthand for a mechanism very close to the data-generating process, fully acknowledging that there is likely no “true” model in the philosophical sense in the real world. The reply was now accepted by PNAS and should appear soon. In the meantime, you can already read the author’s response, which was also posted on the arXiv.
I might or might not comment on this response in more detail at a later point – I suppose the question Perretti et al. pose in their response, namely whether we should search for a mechanistically correct model in the first place, or whether we should go straight for “predictive, model-free methods” (basically ecology goes machine learning) is quite different from the claim made in the original paper. It is actually a question that could and has been extended to many other situations beyond the chaotic dynamics that were the original topic. However, I have so far not seen any consensus on this question, and maybe there can be none, as different people do science to different ends, so I am not sure whether it’s really worth diving into this discussion again (see also related posts here, here or here).
Just to say where I am standing though: I believe science is about more than making good prediction. Predictions get obsolete with new data and time, but mechanisms stay. So, call me a lost reductionist soul, but I’m not afraid to say that I don’t care which model predicts better, I’m still more interested in finding the “true” model (and I maintain that this model should also have the best fit to the data on the long run) 🙂