On Oct. 4, Nate Silver “forecast” (that is the label of his graph) that Barack Obama had an 87.1% chance of winning the November election. Ten days later, he said that Obama’s chance was 62.9%. Princeton’s Sam Wang offered somewhat higher percentages at both times, but in the same pattern.
The odds changed by 25 points, way outside any reasonable margin of error. Does that invalidate the early-October forecasts? If Obama wins in November, will that show that the forecast was more correct in early October than it is now? Or will we know with hindsight that Obama actually had a 100% chance of winning, because he did win? An election doesn’t seem random in the same way that a die-roll is random. Once the result occurs, it seems that it had to. If there is an element of sheer randomness in an election, like the effects of rain on turnout, that element is small.
One excuse for the forecasters is that they could not foresee the President’s poor debate performance and its substantial impact on public opinion. I didn’t think he performed well, but, as Kevin Drum shows, the slide in the polls began before the debate and continued smoothly through the debate–suggesting that it was not the reason for the decline. So the forecasters failed to predict a trend (not a single, random-seeming event like a flubbed debate). Does that make them bad forecasters?
A forecast can move in either direction substantially and unpredictably, because who knows what crazy events may occur before Election Day? Mitt Romney could suddenly start yelling imprecations at the 47% during tonight’s debate. But if a forecast moves for a foreseeable reason, such as reversion to the mean, then the forecasters should have predicted it, and their earlier predictions were errors.
Thus, a forecast is a falsifiable hypothesis, but we don’t test it by waiting for the ultimate election result, which will either be a 100% Obama victory or a 100% Romney victory. We rather ask whether the forecast changes in a theoretically foreseeable fashion before Election Day. The only changes that are acceptable are (1) those caused by large and truly random events, and (2) a gradual movement toward certainty, which reflects the diminishing time left for random events to occur. By that standard, the predictions made in early October have now been falsified, and we’ll see how the current ones hold up.