> this statistics data-to-model only works on large datasets.
> deals with very small, very heterogeneous datasets.
The population itself is a very heterogeneous "dataset", that becomes (perhaps) apprehensible after you interrogate it with respect to the statistical question you have in mind. Let's not forget that the neatness of the data is constructed from the narrowness of the question asked of it, and even still the data is expected (in the colloquial sense) to take on a minimally-informative, max-entropy distribution, except in respect to the statistic of interest. The critical questions levied against historiography also apply here: perhaps the question you are collecting data for is not related to the population the way you think it is.
> why historical models and grand unified theories are so often bunk.
A lot of history practiced today are self-aware of just how much narrative has to be invented to "make sense" of historical events, and so historians value diverse perspectives. One understands that the same events may not hold the same meaning across communities, and so seeks to record and compare how they are perceived across communities or individuals; that perception in that community's narrative is at least subjectively authentic for that community or individual.
To the extent that the historical goal is to predict political/cultural/economic future, perhaps political/cultural theorists/economists are better positioned.