To be clear, I’m using linear in the theoretical sense. This can include things like non-linear regressions, and likely does. Given what is public about his modeling techniques in his own writings, he appears to be using a number of complex statistical models. However, what he is not doing is utilising anything beyond this that would allow for dynamic mechanistic changes to influence the outcomes of his predictions. Many people fail to understand this in social psychology and over-extend their interpretation of complex statistical models like GLMs and SEMs. The fact is, these are not applicable to longitudinal data (which all forecasts inherently are). The SEMs that are applicable to longitudinal data assume the same mechanism to be changing the forecast uniformly, which, in human systems is not the case.
From a “good enough for government work” perspective, Silver is fine. But from an epistemological stance where we have to question the veracity of a model, he and others consistently fail because their underlying assumptions often have little or no correspondence to the real world target they are attempting to model. In the case of some models (like Turchin and the cultural evolutionists) some assumptions have even been falsified.