Not quite. The number was 313.0 - the number displayed there quite often included a fraction of an electoral vote, something clearly impossible to anyone with a cursory knowledge of the electoral process.
Come, now: I’m the one who first whipped the Wang out.
The OP’s argument seems to be built on a foundation of rank ignorance.
The value of a model is not only its accuracy, but its reliability as well. There is a chance that a monkey throwing a dart at a dartboard might hit the bullseye, but I’m not taking him on late night television until he does it repeatedly.
If you value less accurate predictions because they were made at a point less proximal in time than others, you must love you some Nostradamus.
Yahoo Signal’s prediction of 303 electoral votes from a point 9 months out from the election is fine, if it is fundamentally strong. But as SlackerInc notes, any model should be able to get it right on election eve.
So, when I look and see Yahoo Signal’s election eve prediction that Obama will win 303 electoral college votes, I start to suspect that their model is not as good. In fact, their model only gave a 4% chance that Obama would win somewhere between 330-339 EV.
While this model is close, it certainly is much less accurate than many others.
That’s yet another post where I said he called 332 the most likely outcome without saying “Silver predicted 332 EVs,” the words you keep demanding everybody needs to stop using.
I appreciate the thought, but it’s not necessary. I pointed out a bunch of problems with the way you were comparing the work of Silver and Wang and Yahoo Signal and you never responded to any of them except to bicker about the prominence of the numbers 313 and 332 on Silver’s blog, which is by far the least interesting and consequential part of the discussion. The analysis Yahoo Signal did is interesting and was very close to the right result, so if they continue to do forecasts like this it’ll be interesting to see how they compare to other people doing economy-based analysis and analysts who look at the polls. It doesn’t mean we should throw over one for the other or that people should stop looking at polls because those types of forecasting may not match up in the future. There is no one perfect model and we’re best served by looking at a bunch of them.
That’d be a point against them - even though Yahoo Signal also brags that they “won the election forecasting game” and said “Anyone can average a bunch of polls and call the election a week before it happens.” (Translation: we weren’t quite right and were less accurate than some others, but we were not quite right and less accurate first!) Ultimately Signal and Silver/Wang agreed on the basic points and all came very close to the right result, but I’m not as impressed by Signal’s bragging. They’re right that poll aggregation doesn’t always tell us what factors are influencing the election, but it’s not supposed to. Their accurate forecast might be one data point in support of their theory, but it doesn’t prove everything else is irrelevant.
I’m a big fan of Nate Silver, but I do not think than the mean/average/expected-value number that was posted prominently on their blog was useful. It’s not even labeled well (smart people being able to deduce that it’s a mean is not an excuse for not labeling as such). A mean is only a useful number if 1) there’s a big-center, small-tails distribution, or 2) many trials are going to be performed. Neither is the case here.
The single number I followed was the chance of winning. That’s the best way to capture the model results in a single number. I also looked regularly at the histogram–the distribution has the real meat of the prediction. It speaks poorly of the blog that you can’t click on the histogram to get an expanded view of it.
But the poor presentation of the results does not besmirch the accuracy or usefulness of the work.
Exactly. And I suspect this had more to do with the NYT graphics people than it did with Silver.
Incidentally, because I had nothing better to do and I’ve been meaning to do it anyway, I ran my own simulation using Silver’s state vote percentage predictions and margins of error.
For each run, I assumed that each state (congressional district in ME/NE) had errors which were independent of each other state, which simplifies things, but is a bad assumption. Most states are easy enough to call that this doesn’t matter however. I also assumed that the MoE given by Silver is twice the std. dev. rather than 1.96 times the SD, as this seems to fit his results better.
With these assumptions run in 1 million simulated elections, I obtained the following results:
Obama wins 90.5% of the time with an average of 312.6 electoral votes across all runs. The single most common result was 303 EV for Obama, occurring 28.2% of the time. This was followed by 332 EV for Obama, occurring 23.7% of the time.
This matches Silver’s numbers fairly closely. The biggest discrepancy was the difference in most common results. I believe that this can be explained by the fact that, using Silver’s rounded numbers, I had Florida entirely even. In this particular set of runs, Romney apparently won it more often than not. Additionally, there will be more winning maps for Obama in which he gets 303 EV and wins Florida than there will be winning maps in which he gets 332 EV and loses Florida.
Your language here kinda implies it:
And I thought this did, too:
Here’s an interesting article on the internal polls in the Romney camp.
The article doesn’t explain why they focused only on their internal polls, rather than take some sort of average of all the polls. Anybody who lacks the disposition to process outside polling information probably would also hire a pollster who tells them what they want to hear. Such habits and inclinations can be disastrous in a policy context.
To be fair, the model also gave a ~23% chance that Obama would obtain 330EV or better. It seems that they thought that if Obama would win Florida, he would likely win a lot of other places. I too suspect some miscalibration. But methinks their team is operating within the same reality-based territory as Nate. I welcome their contribution, though frankly I’m not overwhelmed by their work. Their Nov 13th self-puffery could have included some references to others in their field such as Ray Fair or a rather extensive list of political scientists. Still, worth a bookmark in 2016.
I’m surprised they claimed that Silver was primarily a poll aggregator. That was true in his late October model, but during the summer he included economic variables which he phased out over time. Sam Wang criticized him for ad hoccery.