With regard to Moody’s Analytics, I’m happy to hear they have a model. First time I’ve heard of it, and I’ve been in the business for decades. I suspect that they back-forecasted the past, as they calibrated the parameters of the model; rather than first developed a model and then used it, election after election, to forecast the unknown future.
Ray C. Fair of Yale has a peer-review economics-based model and he’s forecasting a big, 10-point victory for the Republicans.
https://fairmodel.econ.yale.edu/vote2016/index2.htm
Fair first published his model in 1978, based on data through 1876; and, he has updated it, in a transparent way, each election since.
If you’re wondering, Fair forecast a close win for Mitt in 2012, and instead Obama had a close win. This is what he said afterwords:
“It looks like the two-party vote share for Obama will be about 51.3 percent. The last prediction of the vote equation below (dated October 26, 2012) was 49.0 percent, which is an error of 2.3 percentage points. The standard error of the equation is 2.5, and so the actual error is about equal to the standard error. An error this small means that when I reestimate the vote equation with the 2012 election added, the coefficient estimates are not likely to change much. There is nothing in the 2012 results that cast doubt on the specification of the equation.”
Those of you who tracked the election know that Mitt was ahead ever so slightly going into the 3rd debate. Then, Mitt fell into the prevent defense and lost that debate. And, then there was Superstorm Sandy and a big hug from the Governor of New Jersey.
While Fair says, correctly, his model wasn’t off by much, it’s a lot easier to accept a forecast that is slightly off when your candidate is the winner, than when your candidate is a loser.
Interesting! Thank you for the link and info.
I’ve now gone to the source, and find - exactly as I had expected - the model was calibrated on elections of the past so many years.
It has yet to generate one out-of-sample forecast that was subsequently sustained or contradicted.
Basically, with 8 elections upon which to calibrate a model, it’s a guarantee that calibrating a model with 8 parameters will give an exact fit to the past. But, there would be no measure of forecasting precision for the future. Paradoxically, the better the fit to the past (with a large number of parameters), the lower the out-of-sample forecasting ability.
In the report, I count income, gas prices, housing prices, voter fatigue, presidential approval, and past vote share. That’s 6. Relative to eight elections, that’s a large number of parameters. I am sorry, but the Moody’s Analytics people either don’t know what they’re talking about, or do know and think their readers don’t.
To correct my prior post, Ray C. Fair of Yale first developed his model in 1978 with data through 1976 (not 1876).