This year's Senate polls were more picked-over and modeled than ever before — and their final performance was not impressive. Yes, overall the polls pointed toward a Republican Senate takeover, and correctly predicted the victor of nearly every competitive race. (They missed the Republican victory in North Carolina, and most averages incorrectly gave Greg Orman a slight edge in Kansas.) But the margins they were showing were pretty far off the mark — very far, in several instances:
In the chart, I've showed how far off FiveThirtyEight's projected margins were, for simplicity. But in nearly all of these races, FiveThirtyEight's estimates were within one to two points of simple poll averages like RCP's and the other models that projected margins. Just before the election, everyone was looking at the same polling.
There's no evident pattern to where the polls were the worst. They were way off in deep-red states like Kansas, where independent Orman didn't even come close to winning despite leading most recent polls, and Arkansas, where Sen. Mark Pryor ended up losing by 17 percentage points.
Yet polls similarly underestimated Republican performance in purple states like Virginia —which was far more close than practically anyone expected, though Democrat Mark Warner seems to have barely managed to hang on. The same held true in Iowa, where polls predicted Ernst was ahead by about two. (This means that the famously accurate Des Moines Register poll, which showed Ernst ahead by seven points, has beaten the averages yet again.)
As more votes are tallied it's possible these margins will change somewhat. But these results present new ammunition to the analysts who place less trust in polling averages. When it comes to predicting margins, those averages gave us no indication that so many sweeping Republican victories were about to occur.