To explain the charts oversimply, say three polls come out that show a swing to Trump of three points, one point and five points.
The HuffPo model will quickly take Hillary's chances from say 90% to 80% to 75% and then stay there until new data. Nate's will do something like go from 85% to 83% to 82% to 81% to 77% to 75% to 76% because it's churning far far more data.
Nate's model is super-conservative because he's trying to smooth everything out by introducing tons more regressions. It looks like it's moving all over the place when it's actually just trying to settle over time. That's why he has the "nowcast" and "polls-plus" and they will ideally converge. (They basically have at this point.) Most of the other models are doing "nowcasts" and there's no expectation that the HuffPo projection from two months ago should drift towards the one today. Ideally, the "polls-plus" from four months ago looks more like the one today than any nowcasts did.
Today Polls-Plus: 48.4% Clinton (290), 45.6% Trump (248), 4.5% Horseshit
Today Nowcast: 48.3% Clinton (290), 45.5% Trump (248), 4.6% Horseshit
Day Before GOP Convention "polls-plus": 47.5% Clinton (298), 44.7% Trump (240), 6.5% Horseshit
Day Before GOP Convention "nowcast": 45.9% Clinton (306), 42.5% Trump (232), 10.4% Horseshit
Clinton's peak in the nowcast was August 8th: 50.3% Clinton (388), 40.6% Trump (150), 7.8% Horseshit
On the same day, the polls-plus was projecting: 49.0% Clinton (324), 44.4% Trump (214), 5.3% Horseshit
Ultimately, Nate is making things more complicated and creating his own readers confusion because there's only one trial. But he's most likely working from a sports stats mentality of putting in past seasonal data to smooth out the flukes. This probably explained why he was so damned adamant about finding a model that explained every primary back to 1972 and fitting 2016 into it rather than looking at the data in front of him (and the completely different calendar, and sheer number of candidates and...) as a separate trial and dismissing everything except some aspects of 2008 and 2012.