Most people didn’t predict Trump’s win. I didn’t, the Polls didn’t, the financial markets didn’t (stocks, foreign exchange, etc.), the betting markets didn’t, and most experts didn’t. However, there is a group that did predict a trump win using statistics that has consistently worked well at predicting presidential elections: econometrics models. This literature was started by Yale University economist Ray Fair back in the 1970s. These models largely predict who will win based upon growth in GDP. It turns out that when GDP has been growing well, the incumbent president is much more likely to be reelected. Other models that have used this methodology include Vox.com, Time-for-change, Lewis-Beck & Tien, and Trial-heat.
These models predicted that a generic Republican should have won the popular vote and that Donald Trump significantly underperformed the fundamental political landscape that he was given by fate. Even though he won, the econometrics models still judge him to be a somewhat weak candidate because his vote share was considerably smaller than what they predicted. Unfortunately, the models use GDP rather than median income. Median income would be the best data for predicting the median voter, but the models use GDP instead because it is much easier to get timely GDP data than median income data and the latter data is noisier due to cheaper, smaller sample sizes.
Indeed, given how badly GDP correlates with the wellbeing of the median voter, it is amazing that it helps predict elections at all (and 538’s Nate Silver argues that it has little value). Consumption would also be a better measure of economic wellbeing in theory, but it might not matter much for the political economy models in practice because GDP is a fairly good proxy for US household consumption in the short-run. (The two are highly correlated in the short-run because consumption is the biggest component of GDP and is causally interdependent with the other components.) But consumption should produce better political predictions than gross GDP.
Nate Silver (538) agrees with me. His 538 election model uses personal consumption expenditures and real personal income rather than GDP, but 538’s prediction model gradually reduces the weight placed on economic variables as the date gets closer to Election Day. On Election Day, they completely discarded the economic fundamentals information and just use polling data. That helps explain why their prediction was worse than the other econometrics models that stuck with the fundamentals and ignored the polls. It wasn’t a problem with using consumption data. The problem was that they gradually discarded information about it from their model and increasingly just used polls which turned out to be way off.
One reason nobody paid much attention to the econometrics models is that they usually give similar predictions as the polls give. This election they contradicted the polls and most people trusted the polling data more than the economic fundamentals which they ignored. Even many of the people creating the economic models tended to discount their predictions this time around. For example, Vox’s Andrew Prokop said:
Abramowitz publicly disavowed his own forecast’s projections, arguing that it applied for “mainstream” candidates and not for Trump. “Donald Trump is far from a mainstream candidate,” he told Vox’s Dylan Matthews in an email. That’s not to beat on Abramowitz here — I also was aware of what these models showed and yet rarely wrote about them, since I shared the opinion that they were probably going to be “off” this year because Trump was just so strange.
In the end, it turned out that Trump was indeed less popular than the economic models predicted, but not by much. I suspect that if we had good enough data about median income to use that in our econometric predictions we would see that Trump underperformed the economic conditions even more than was indicated by the models based on GDP. We just don’t have good enough data to know.
For example, although the Census Bureau announced the biggest increase in median income in recorded history just before the election, their announcement was about data that was already a year old, so it was too old to be useful for political predictions based on what we know from the GDP models. Median income is so neglected by the government that there is no official measure that is up-to-date, but Sentier Research posts the most up-to-date measures and their data shows median income was down in 2016.
Plus, even if we used more up-to-date data, some of the apparent increase was just a statistical illusion created by changing work hours and a change in the survey methodology. The new survey increased their measure of household median income by 3.2% so that could account for most of the apparent increase. The true change in median income was highly uncertain given the ongoing noise in the data caused by small sample sizes and the methodological change. You can see the noise in the data in the jumpiness of the red line above. Most of that is an illusion too. We need better data!