I recall reading about a paper in SciAm or American Scientist a couple of decades ago, where they had trained a ML model to predict regional conflicts or civil wars. The main input was scarcity of food, mainly through price IIRC.

They trained it on historical data up to the 90s or so, and had it predict the "future" up to the time of the article. And as I recall it did very well. They even included some actual near-future predictions as well which also turned out pretty accurately as I recall.

Which I suppose isn't a huge surprise after all. People don't like to starve.