The UK since 2001-2005 has apparently managed to get extremely accurate exit-polls:
Prior to 2001 the methods used had been different, and the accuracy of predictions based on exit polls was rather variable — sometimes good, sometimes poor. Perhaps the most famous failure was in 1992, when the forecasts made by both BBC and ITV based on separate exit polls were that the Conservatives would only just get more seats than Labour, resulting in a predicted hung parliament. It turned out that the Conservatives had actually done much better than that in the 1992 election, winning 65 more seats than Labour and achieving a parliamentary majority of 21 seats.
And it doesn't look like simply a matter of well-gerrymandered seats; the page only give data for 2005 percentage-wise, but it does look impressive
In contrast the US exit polling is still very troubled by the "red shift" (i.e. failure to predict Republicans winning, when they do), so troubled that it led to a partial dissolution of the their National Election Pool (NEP):
The Associated Press confirmed Friday it has joined Fox News in abandoning the so-called National Election Pool — the election surveying instrument that the news media, campaign operatives and political junkies have come to love and hate — marking the end of an era when one ubiquitous Election Day survey shaped the understanding of presidential and state election outcomes.
[...] the early exit-poll results proved misleading in the 2016 presidential election. [...] The early exit polls were similarly off in 2004 — in the same direction. Exit polls showed John Kerry ahead of then-President George W. Bush in a number of key states.
Fox News cited this when it announced it was leaving the pool earlier this year. “We’ve had concerns with Election Day exit polling for many years, and this year once again proved that they are problematic,” Jay Wallace, Fox’s executive vice president of news, said back in April.
The UK page I liked to does go over the main characteristics of the present-day UK exit-polling system, but it's not immediately apparent how this compares to the (NEP) US system. So, question: what are the main methodological ingredients in the UK exit-polling system missing from the US NEP system?
From a non-technical article on one of the architects of the UK exit-polling system:
The secret, he says, is measuring change. He goes to the same polling stations sampled in the previous election and measures the change in party support between the two exit polls, modelling this according to likely variables – the Conservatives were likely to do better in places where Ukip had done well previously, for example, and Labour in places with more graduates – eventually producing a set of equations for all the constituencies in England, Scotland and Wales. The final poll is based on a sum of probabilities across all parties of their chances of winning.
Is this the secret sauce missing the US exit polls, "measuring change"? I noted for instance that there has been no exit polling for the Brexit referendum, because there was no baseline to measure change from.
There's an academic-ish article on the 2017 UK poll as well. One interesting factoid:
The exit poll interview has just one question: for whom had the respondents just voted? They report this by filling in a mock ballot paper, unseen by the interviewers, and dropping it into a mock ballot box.
I guess this is intended to avoid any "shy factor". I haven't heard of this being done in the US, by the way, although evidence for this kind of effect skewing the exit polls in the US is apparently weak.
Also this paper gives the 2017 results which are not spotless (prediction-wise), but still pretty good: