VOTING WITH YOUR TWEET:
An experiment in political forecasting
As of 15 October, we showed Democrats winning about 33% (123 districts) of the 369 districts for which we generated vote share predictions. The win/loss predictions ran more or less the same. The figure at right shows an overall trend of declining Democratic Party fortunes. Since the end of September, our projected seat counts for the Democrats have trended steadily, if modestly, down.
For comparison, the New York Times currently projects that the Democrats will likely win 183 of 435 seats, with another 24 up for grabs. If we give half of the tossup seats to the Democrats, that’s a win rate of about 45%.
Why the difference? First, our forecast contains only the districts for which we have data to make a prediction.* That sample has fewer districts with Democratic incumbents than the actual House of Representatives (36% compared to 44.6% in entire pool of House seats). Since political scientists generally agree that an incumbent is more likely to win re-election, this means that the set of districts we predict for is biased towards races where Democrats are more likely to lose.
We can do a thought experiment to try and compensate for the imbalance in our sample. We could assume that the incumbent party always wins in the districts where we have no data. If we do that, the numbers even out somewhat: we predict that about 148 Democrats will win, and about 284 Republicans. But this still works out to a win rate for the Democrats of about 35%.
Thus we still predict fewer Democratic victories than more established forecasts. As we note in the FAQ, this may indicate that our algorithms contain some biases of their own. That said, we are in line with other predictions in one sense: more or less everyone thinks that the Republicans will probably keep the House. But we predict a far larger Republican majority than most mainstream forecasts.
But we emphasize that this is still early in the campaign and this experiment in forecasting. And because we don’t generate predictions for every district, we cannot say precisely what the makeup of the House would be if the election were held today. We’ll have to wait for the election for the final verdict on how our methods do compared with other, more proven forecasts.
* We don’t have data on districts when a race is uncontested, one of the two major candidates in the race is neither Republican nor Democrat or we only have Twitter data for one candidate.