The Washington PostDemocracy Dies in Darkness

How well could we have forecast the midterm election without polls?

Placeholder while article actions load

One issue that animated some conversation during the 2014 campaign was not just whether the election could be forecasted accurately — it was — but how we should do it.  Three of the forecasting models at media organizations — ours at Election Lab, The Upshot’s and 538’s — relied on a combination of “fundamentals” and polls initially, but increasingly gave weight to the polls as Election Day approached.  For this reason, these three forecasts largely dovetailed with the forecasts based only on polling data (Pollster, Daily Kos and the Princeton Election Consortium).

The fundamentals in our model included economic growth, presidential approval, whether it was a midterm or presidential election, whether the incumbent was running, the previous vote received by any incumbent running, the previous political experience of the candidates and the balance of fundraising.

So imagine that we had predicted the election purely with where the fundamentals stood on Election Eve, while ignoring the polling entirely?  How well would we have done? On balance, not badly.

Here is a graph that compared a fundamentals forecast to the actual outcome in the Senate races.

Unsurprisingly, there is a strong correlation (r=0.93).  The fundamentals would have “called” 3 races incorrectly, which are highlighted in red.  Of course, adding polls improves the forecast — increasing the correlation, reducing errors, etc.  A purely fundamentals-based forecast is also likely to contain substantial uncertainty.  But the fundamentals by themselves still do reasonably well.

One interesting thing: in several races where the polls really underestimated the Republican candidate’s vote share — Kentucky, Georgia, South Dakota, West Virginia– the fundamentals did significantly better.  So, although the polls do trend toward the fundamentals during the campaign, they may not converge to the outcome forecasted by the fundamentals.

Of course, the outcome forecasted by the fundamentals wasn’t always an improvement on the polls, as attested by states like Arkansas, Michigan and Virginia.  And given that poll errors are themselves fairly unpredictable election-to-election, it would be difficult to know ahead of time whether and where the fundamentals might prove more accurate than polls. There’s more to learn about this, clearly.

In the meantime, the simple point here is that in 2014, the basic features of the political landscape could predict many things about this election.  I’d say that 2014 was largely a victory for the fundamentals.