Brad DeLong: Dylan Byers of the Politico Is Back! And Tweeting! And Meat's Back on the Menu, Boys!: Nate Silver Really Is a National Treasure Weblogging: For many years, decades really, I wondered why Napoleon was considered a military genius. He didn't do so well in the invasions with which I was familiar. Then I read a summary of the Penninsular campaign (battles in Italy with French forced commanded by Napoleon still a general the Republic). Monotonous. Just the same thing one time after another -- The Austrians raise an army twice as grand as Napoleons and send it South where it it utterly routed, rinse repeat.
Silver is not highly thought of because he managed to predict that Obama would win the general election in 2008.
He is respected because he called primary after primary, totally torching the pundits.
His claim was that support for Obama and Clinton was changing fairly slowly and very different by Democraphic category so the way to predict was to use old polls and polls from other states to estimate relatively precise cross tabs and then use the fitted probabilities times census data. This was a new approach and it worked much better than the old approaches.
The sacred journalistic principle that only the most recent events are interesting explains how someone could discuss the Silver question without knowing these facts.
Now I am not saying his NY Times job and current status is like capturing Moscow or anything. My forecasting rule is to avoid making forecasts (although I did better on the 2010 Senate election than Silver).
On Silver v RCP, I don't understand how RCP decides which polls to include in their average. I think transparency is very important.
On Silver v TPM, i love TPM but their state by state forecasts look like a joke. They allow one extremely recent poll to totally take over a state (Foster/McCallum and Michigan). The TPM machine might forecast worse than a pundit (when last checked had Ohio and Colorado leans Obama and Michigan and Wisconsin tossups).
The claim that simple smoothing works fine can't be correct. Different sites have extremely different state by state numbers. The choices are between algorithmic weighting (Silver) and "editorial judgment" (others) and falling for nonsense polls (others).
Now an important feature of optimal forecasts is that they are martingales -- the product of the probability now P1 than the probability at time t that X will happen is equal to the probability now that X will happen. The relative stability of Silver's Nov 6 forecast is evidence that it is efficient or biased. We will no more on that question November 7th. I just surfed over to TPM where I read "Poll: Obama Re-Emerges As Favorite In Wisconsin". Silver's model predicted that this would happen.
Deciding whose forecasts to trust in September based on whose early November forecasts were close to the outcome is idiocy. To just September forecasts one must confront them with the outcomes. To do this reliably, one needs data on dozens of elections. But to pretend that there is no problem deciding who to trust in September based only on performance in November is to be a fool.