As far as I have read of their blogs, they’re par for the course.
I mean it’s not like them attaining “The Founders” status (excuse me, Isaac Asimov), not even closed.
Yes, and the publish datasets of all their work to let others analyze them too. They're the most accurate of all the major sites.
Here's their github site [1]. It contains tons of datasets, past performance, analysis tools, everything you want to actually check if your claims are true or false, and most importantly, how accurate 538 has been over time.
Now try to do that analysis on, say, Fox News :)
Good luck.
Their seems to be a systemetic bias for them to be underconfident. That is to say, when they a result is likly (>50%) to happen, they underestimate how likely. Conversly, when they think an event is less likely (<50%), they overestimate its chances. Still, more likely events tended to occur more often than less likely events.
Correcting for this systematic error would mean they should have given Clinton an even higher chance.
> Trump is a significant underdog — he has a 13 percent chance of winning the election according to our polls-only model and a 23 percent chance according to polls-plus. But those probabilities aren’t that small. For comparison, you have a 17 percent chance of losing a “game” of Russian roulette.
A) Their more advanced model gave Trump a 23% chance, the 13% chance you mention is their simpler model.
B) Things with a 23%, 13%, and 7% chance happen all the time
C) You often have a better understanding of events after they happen relative to before.
If you want to show a predictor is unreliable, you need to look at a bunch of their predictions and see how they perform in aggregate. While also considering how correlated the results are (and, if applicable, predicted to be)