The author doesn't seem to understand how the 538 model works. The 538 model does take epistemic uncertainty into account; it just doesn't call it that. The m

The author doesn't seem to understand how the 538 model works. The 538 model does take epistemic uncertainty into account; it just doesn't call it that. The model calls it "likelihood of polls moving by X% between now and election day". In 2016 they offered a "now-cast" which didn't include that, and answer the question "based on where the polls are right now, how likely is any given outcome if the election were held today?" while the "classic" model answered the question "based on where the polls are right now and the fact that unanticipated things will happen between now and the election, how likely is any given outcome?". For 2018, they dropped the "now-cast", because people didn't understand what it was saying. True, the epistemic uncertainty is limited -- it's based on past elections, because hey, 538 works with actual data. One could argue that rather than saying "71.4% Clinton 28.6% Trump", a better model would have said "71.3% Clinton 28.4% Trump 0.2% Election is cancelled due to nuclear war / natural disaster / etc" but I don't think anyone sensible is interpreting the model as excluding such extreme outcomes.