November 2, 2020: Final Forecast
The last eight polls in the RCP average for Pennsylvania have five showing Biden with a lead, and three showing Trump with a lead. The previous seventy polls showed a lead for Biden in all but three polls. This is the last minute tightening that is a huge positive signal for the Trump campaign. The RCP average for Pennsylvania has shrunk to Biden +2.5. In 2016, Trump beat the polls by 2.6%. Trump has been hitting the Keystone State hard over the past week, with large turnout for his rallies, and a heavy advertising blitz focusing on what a Biden/Harris administration would do to the energy sector. ElectoralMap.net is predicting a Trump win in Pennsylvania, with the strong possibility that an extended ballot counting period of days or weeks flips it to Joe Biden.
Although polls have tightened in Minnesota, Michigan, and Wisconsin, Biden still has a larger margin than the polling error in 2016 - so those states are staying in the Dem column in the forecast. Expect the races to be close, though. Likewise Arizona, Georgia, and North Carolina are breaking in Trump's favor in last minute polling - states that were probably Trump's all along.
We may have a fairly good barometer for this election: Florida and North Carolina returns could come in reasonably early, and if either of those states goes for Biden he will almost certainly win, if he takes both then it's basically over for Trump. This brings us to the topic of Florida. The latest polling data hints that Florida could be Trump's biggest vulnerability. The polls are not tightening in Florida in the same way they are in Pennsylvania. This is really bad news for Trump, as he's still about 0.7 points away from getting inside 2016's polling error. That RCP figure, however, doesn't weight the Trafalgar poll (October 29, Trump with a 3.1% lead) the same way ElectoralMap.net does - a much heavier weighting for 2016 accuracy. Over on the PredictIt prediction market, Trump has had a massive surge in Florida which started around October 19th. Although this site doesn't weight PredictIt heavily, the surge is big enough to move the Sunshine State into Trump's column. However, Florida could very well be every bit the showdown as it was in 2000.
No, the polls were not mostly right in 2016.
October 15, 2020 There's a common refrain you've heard since the 2016 election, that the polls that year were actually mostly right. It's not true, and if you pay close attention to what is being written, even by reputable sources, you can pick apart the bad analysis. Some deflect with stories of down-ballot accuracy. Some insist the polls were accurate when taken, and blame the error on late-deciders. Most trying to make the case that the polls were good conflate state-by-state polling with nationwide polling, and once they've blurred the lines, you'll see the two used almost interchangeably. Although there was national polling that predicted the outcome of the popular vote quite nicely, it's irrelevant. The state-by-state polling was a disaster - and that's all that counts.
The United States does not hold a nationwide vote for President. Rather, the outcome of the Presidential race is determined by the Electoral College, which is determined by the individual elections in each state and D.C.. In 49 out 51 of these elections it's winner-take-all. This changes the strategy for an overall win, and it changes the way we try to predict that win. You can win the national popular vote and still lose the election, so we must forecast most of the individual states correctly. A good forecast for the national popular vote makes for an interesting piece of trivia, and nothing more. To correctly predict who will walk away with the Presidency, you need to have a good idea of how the vote will land state-by-state. Bad polling in just a handful of states can ruin the entire forecast. Missing key states is missing the election. Bad polling in a few key states is bad polling in toto. The pollsters know this, and the forecasters who rely on poll numbers know this too, but very few will admit how bad it was in 2016. The site you're reading right now missed 2016. FiveThirtyEight did too, by substantially more electoral votes. They, however, are hesitant to call the polling bad.
Not only did 2016's state-by-state polling miss critical states, it missed them all in favor of Clinton. In fact, virtually all of the closely watched swing states had polling error in favor of Clinton. The error in Wisconsin was absolutely terrible, being off by over 7%. Iowa was just as bad at 6.5% Pennsylvania, Michigan and Ohio all had error in Clinton's favor between 3% and 5%. North Carolina clocked in just under 3% error. These numbers are not just outside a reasonable margin of error, they are horrendously outside - and horrendous in nature. One of the primary benefits of poll-aggregation, not just for the overall election but for each individual state's polls, is that biases and errors tend to cancel each other out. However, when every single miss is in one direction, the aggregation is going to be garbage as well.
|State||2016 RCP avg||2016 Actual||Error/Bias||2020 RCP avg|
|Wisconsin||Clinton +6.5||Trump +0.8||Clinton +7.3||Biden +6.3|
|Iowa||Trump +3.0||Trump +9.5||Clinton +6.5||Biden +1.2|
|Ohio||Trump +3.5||Trump +8.0||Clinton +4.5||Biden +0.6|
|Michigan||Clinton +3.4||Trump +0.2||Clinton +3.6||Biden +7.2|
|Pennsylvania||Clinton +1.9||Trump +1.1||Clinton +3.0||Biden +6.5|
|North Carolina||Trump +1.0||Trump +3.7||Clinton +2.7||Biden +3.3|
|Florida||Trump +0.2||Trump +1.1||Clinton +0.9||Biden +2.7|
|Georgia||Trump +4.8||Trump +5.2||Clinton +0.4||Biden +0.4|
|New Hampshire||Clinton +0.6||Clinton +0.3||Clinton +0.3||Biden +11.0|
|Arizona||Trump +4.0||Trump +3.5||Trump +0.5||Biden +3.5|