Great read. I’m not a statistician, but all the excitement about 538, even in 2008, had smelled fishy to me, specially the “49/50" business. In 2016 i stopped paying attention when I realized their bias in choosing which polls to include in the Meta model and the fact that bring statisticians and not physicists they didn’t calibrate their data against ground truth by talking to people, which would have led them to speculate that people were lying on the polls and admitted to it when they felt safe.
My question: are what you are calling epistemic and aleatory variance the respective sources of the bias and variance terms in machine learning? Or not quite?
Oh and nice choice of example to illustrate the difference, since alea means game of dice as I remember from Asterix.