Locking in Election Predictions

George Turner in reply to George Turner says:

Hrm… Do I know an integer from a real, a number from a number that represents a measurement that had an error range? I think I do.

Your reliability data was gathered at great trouble and expense, with lots and lots of materials tests to back it up, and then built on a huge amount of maintenance data.

What do you do for a protyoype that’s flown six times and had major design reworks after each? You inspect everything you can easily get to, every flight, and tear down anything that the engineers were worried about. If only they had Nate Silver to tell them that he’d modeled the aircraft based on in-flight photographs of the six test hops, and in 78.6% of the simulations, no stress cracking was observed in the first 2,000 hours of operation.

Perhaps this is why airlines don’t hire baseball statisticians to run their maintenance programs.Report

George Turner in reply to Shazbot3 says:

That’s just one of many sources of uncertainty. For example, in one of the latest polls (Gallup, I think) Romney and Obama were tied, but in a poll that had +11 Democrat sampling, whereas other polls have found R’s and D’s evenly matched. That’s a pretty big error bar in the data.

Keep in mind, if the polls were always accurate we wouldn’t have upsets.

Or, put more simply, you and I have a bet to see who can get the most friends to come over for a party. We call all our usual friends up and note how many say they’ll be there. We can go look at our previous four or five parties held over the past twenty years and make some estimates as to who will win. But since we can both do that, and we really want to win the bet, we start doing all sorts of crazy things, like driving around to our friend’s houses and trying to drag them with us. We work the phones to try to get our friends to bring some of their friends. We come up with clever tweets to try and cause a cascade of party goers to show up. We do what’s called “the ground game.”

There’s only one real certainty involved in this bet, that the geek sitting on his iPad building a probability spreadsheet is not going to get laid.Report

George Turner in reply to Shazbot3 says:

Because Silver’s Monte Carlo simulations are computational intensive, and if they had run on computers back in the day, those same computers would still be running the LBJ Goldwater race?

Wrong answer. We used them in the 1940’s to design bombs. Any of Nate Silver’s techniques could be run on an 4.77 MHz IBM PC/XT just fine. You could run them on an ENIAC, but you couldn’t get the card’s punched fast enough to get the program done.

A few minutes ago Obama and Romney both appeared for a half-time interview on a sportscast I was watching. Needless to say, it was like watching the head of the Olympics debate Britney Spears about sports. It was so bad, and will be so fresh in the minds of sports fans, that it will probably generate a 60 point shift in the election, putting Romeny at +110 and Obama and -10, vacating Obama’s 2008 victory over McCain and giving us a do-over of the last four years, depending on the ref’s decision. If that happens, we get four years of McCain and then four years of Romney.

The statistical models don’t include breathtakingly stupid decisions taken on election-night eve, and all should really include the caveat “Assuming Obama doesn’t totally embarrass himself on national TV at halftime the night before the election, in front of every sports fan in the nation.”Report

George Turner in reply to Shazbot3 says:

Oh, and I wrote my own Forth interpretter in assembly language on an 8080 (which got me my first university programming job). Random numbers aren’t a problem, especially when you’ve got a copy of Knuth handy.

Randomly shifting state voting percentagess around a centroid of state polling data and counting how often a candidate wins wouldn’t tax a TRS-80. Heck, you should’ve seen what the horse industry was doing back then with the jockey club data, much less the video capture and gate analysis.

That’s why I’m saying that if the method was inherently robust, any university person, even an undergrad, could’ve been using it at least since the early 1980’s. Since everyone knew about such methods, the most likely reason they didn’t use them is that either they’re pretty obviously not robust, or even more likely, if you try them on any electoral data sets from that era they probably fail miserably, leading to their immediate rejection. This raises an interesting question. If Nate Silver only goes back X number of elections, and if nobody in the early PC era used these methods (back when the programmers were trying out their new found mathematical powers on everything from horse racing to stock picks), perhaps that eras elections do create staggering failures with the method, which is why Nate Silver doesn’t try to use them.

If so, that might bear a similarity to the modeling failure that led to the collapse of the housing market, where the innovative (and stupid) method they were using looked great on all the input data they had, but their short sample of input data had never included a downturn in the housing market because the last downturn was too far in the past.Report