[xapata]

How does one accurately account for uncertainty?

If you're talking about Taleb's "black swan" style uncertainty, that's specifically the kind that cannot be taken into account.

If you're talking about events being correlated instead of independent, that's what set 538 apart from other poll-aggregators: Nate gave Trump a relatively high chance of winning, because a polling error in one state would likely indicate the same polling error in every other state.

[jtsnow]

As I understand it, FiveThirtyEight's models reflect current and past uncertainty, but do not properly take into account future uncertainty. This greatly undermine's their models for use as a prediction, a probability, or a forecast.

At one point, FiveThirtyEight was forecasting/predicting that Clinton had a 66% chance of winning the election. If there were infinite parallel universes that veered off from that moment in time, their implication is that she would win the election in two out of three times. Taleb is saying that is nonsense, that there is more variability than that.

Of course, Taleb's model wouldn't drive page views or get people excited. FiveThirtyEight gives voters a reason to check-in daily on how their candidate is doing in polls, but it is misleading as a forecasting tool.

[xapata]

Eh? Taleb was criticizing the variability of 538's estimates and said they should not have been making such drastic changes to their estimates. Taleb waffles between Bayesian and Frequentist and in this case he's being Bayesian.

He didn't criticize them for underestimating uncertainty, but for being too uncertain.