[Note, October 13: The probability is currently displayed as a decimal, i.e. 0.4 means 40%. The reason for this is that the precision of the probability is no better than +/-10%. This is the case for other aggregators as well - we're just being explicit about it. - Sam]
As planned for a long time, we’re switching soon to a short-term forecast. As I wrote last month, the Meta-Margin has some predictive value for where it will be within five weeks. The election is in five weeks, so it’s time to start factoring in current conditions.
I’m traveling today, so there will be a slight delay in bringing it live. In the meantime, after reading your comments on yesterday’s thread, I have some comments of my own. [Update: after the break, I have now added a simpler summary.]
[I realize that what I originally wrote is fairly technical. Here's a simpler version:
To make a prediction, I have assumed that October will be like the June-September general campaign, a period when Democrats led for Senate control most of the time. However, Democrats took a small dive last week, raising the question of the assumption's validity. So I asked: what if October is like all of September? As it turns out, the math all comes out the same because Democrats have led for most of September as well. So there’s no easy way to make a different long-term prediction based on polls only. Besides, changing the math would make some people mad, so I shouldn't.
As the election gets closer, current polls will do a better job of predicting the outcome. As time passes before Election Day, I'll count those more and more heavily. If they show a Republican lead, eventually the prediction will flip.
Finally, I realized thanks to commenter Jay that polls can be a bit off in either direction, even on Election Eve. I am adding some math to accurately reflect that possibility.]
Changing the time window. There was a split of opinion about whether to readjust the time window for setting the long-term forecast. Currently it’s at June-now; the question was whether to switch to September-now. The argument against switching was: it’s cheating. The argument in favor was: if it’s likely to be more accurate, do it.
As much as some would like to make this a nerd-off, I think most people just want to know what is likely to happen in November. Think of weather forecasting. You don’t care what the forecasters think of each other. You just want to keep out of the rain.
At the moment, the two time windows currently give the same result, within 2 percentage points. The reason is that the June window was originally set not only by logic (Memorial Day to start the season), but also to emphasize the fluidity of the race. At the time, this meant finding a time window that favored Republicans as much as possible.
The two approaches won’t change future predictions either, mostly. The September window would get the probability moving toward current conditions a few days sooner, since it averages data over a shorter time period.
Note that in both cases, the short-term forecast will (of course) look progressively more like the current-conditions snapshot. On balance, I lean toward keeping things the same.
Unknown systematic error. Last week, commenter Jay suggested that I might be underestimating the possible discrepancy between Election-Eve polls and actual results. If you look at the discrepancies from past elections, which I listed yesterday, they showed a range of Republicans outperforming polls by 1.4% to Democrats outperforming polls by 2.9%. The distribution of discrepancies is coded in two parameters: sigma_systematic (which is symmetric in either direction) (systematic), and the long-tailed distributions that I favor using (blackswanparameter, currently set at a 1-d.f. t-distribution). My current thought is to increase the parameter systematic in the script from 0.7% to 1.2%.
Display. You didn’t bring it up, but the display is currently rounded off to the nearest 5%, since all forecasts (including ours) aren’t any more precise than that. That hid some small movements, and gave the false appearance that nothing was changing at all. The media market is set by big players like the NYT, and has established higher precision as a reporting standard. Seems like we might start rounding to the nearest 1%.
One final note: today’s forecast would hardly change, even if all of these changes were implemented.