Of course it did - his model assumes a correlation with various states, and that a polling error in one state will likely apply to another state. As the overall polling error became clear as states were called and the actual-versus-polls became known, the model adjusted for that. You're also seeing the effect of swing states being called - as states are called, they stop being "70%/30%" chances are start becoming "100%/0%" chances, and that flat-out eliminates certain possibilities.
Yes, it rapidly swung from 70% Clinton to 70% Trump - when the east coast votes were tallied and it was clear that Clinton was losing swing states. But only when actual, real data was coming it.
The actual difference between the final polls and the actual results was something like 2%, which is well within the margin of error. It turns out that the polls this year were actually more accurate than they were during the 2012 election.
His model was fairly accurate throughout the year - it showed a highly volatile and uncertain race that was slightly in Hillary's favor. It's starting to sound like the failing in the polls has more to do with the assumptions of who was going to vote - turnout this year was far lower than in 2016, probably because a lot of voters couldn't stand either choice.