In 2004, I introduced a model I built to predict presidential elections. More specifically, it predicts the outcome of the electoral college, which is all that matters, after all. So far, the model has worked out pretty damn well: in 2004, it hit the electoral count on the nose, and in 2008 2012 it got the winner right, missing the actual count by only 21 votes both times, or roughly the value of an Arizona swinging from one candidate to the other.
The model is built on "prediction markets," where real-world betting happens on each state's outcome. Betting is continuous, and from it we may infer each state's real time probability. I then multiply this probability times the number of electoral votes available in that state for the probablity-weighted outcome. It is understood that one cannot win part of a delegation, it's all-or-none (excepting Maine and Nebraska). Still, for predictive purposes, you don't want to hand 100% of, say, California's electoral votes to Hillary just because she has a 95% chance of winning those votes. You give her 95%. It's effectively a small hedge against something dramatically unexpected happening.
It's time to roll out the 2016 model. Right now, it has Hillary winning a somewhat tight election, 299 to 239. In 2012, Obama won 332 to 206.
State | Electors | Percentage | Repub margin | current betting | weighted expected |
of Electors | 2016 | market odds | electoral votes | ||
Alabama | 9 | 1.67% | 23 | 100 | 9.00 |
Alaska | 3 | 0.56% | 15 | 100 | 3.00 |
Arizona | 11 | 1.86% | 9 | 70 | 7.70 |
Arkansas | 6 | 1.12% | 24 | 100 | 6.00 |
California | 55 | 10.22% | -33 | 6.5 | 3.58 |
Colorado | 9 | 1.67% | -7 | 21 | 1.89 |
Connecticut | 7 | 1.30% | -17 | 0 | 0.00 |
Delaware | 3 | 0.56% | -19 | 0 | 0.00 |
Dist. of Columbia | 3 | 0.56% | -87 | 0 | 0.00 |
Florida | 29 | 5.02% | -1 | 40 | 11.60 |
Georgia | 16 | 2.79% | 8 | 78 | 12.48 |
Hawaii | 4 | 0.74% | -43 | 0 | 0.00 |
Idaho | 4 | 0.74% | 21 | 100 | 4.00 |
Illinois | 20 | 3.90% | -17 | 7 | 1.40 |
Indiana | 11 | 2.04% | 10 | 85 | 9.35 |
Iowa | 6 | 1.30% | -6 | 43 | 2.58 |
Kansas | 6 | 1.12% | 20 | 100 | 6.00 |
Kentucky | 8 | 1.49% | 23 | 100 | 8.00 |
Louisiana | 8 | 1.67% | 17 | 100 | 8.00 |
Maine | 4 | 0.74% | -15 | 0 | 0.00 |
Maryland | 10 | 1.86% | -26 | 4.5 | 0.45 |
Massachusetts | 11 | 2.23% | -24 | 7 | 0.77 |
Michigan | 16 | 3.16% | -9 | 22.5 | 3.60 |
Minnesota | 10 | 1.86% | -8 | 17 | 1.70 |
Mississippi | 6 | 1.12% | 11 | 100 | 6.00 |
Missouri | 10 | 2.04% | 10 | 81 | 8.10 |
Montana | 3 | 0.56% | 13 | 100 | 3.00 |
Nebraska | 5 | 0.93% | 21 | 100 | 5.00 |
Nevada | 6 | 0.93% | -6 | 34 | 2.04 |
New Hampshire | 4 | 0.74% | -6 | 33 | 1.32 |
New Jersey | 14 | 2.79% | -17 | 9 | 1.26 |
New Mexico | 5 | 0.93% | -10 | 0 | 0.00 |
New York | 29 | 5.76% | -28 | 8.5 | 2.47 |
North Carolina | 15 | 2.79% | -2 | 51 | 7.65 |
North Dakota | 3 | 0.56% | 20 | 100 | 3.00 |
Ohio | 18 | 3.72% | -3 | 42 | 7.56 |
Oklahoma | 7 | 1.30% | 34 | 100 | 7.00 |
Oregon | 7 | 1.30% | -8 | 0 | 0.00 |
Pennsylvania | 20 | 3.90% | -6 | 34 | 6.80 |
Rhode Island | 4 | 0.74% | -28 | 0 | 0.00 |
South Carolina | 9 | 1.49% | 11 | 100 | 9.00 |
South Dakota | 3 | 0.56% | 18 | 100 | 3.00 |
Tennessee | 11 | 2.04% | 20 | 95.5 | 10.51 |
Texas | 38 | 6.32% | 16 | 91.5 | 34.77 |
Utah | 6 | 0.93% | 48 | 84.5 | 5.07 |
Vermont | 3 | 0.56% | -36 | 0 | 0.00 |
Virginia | 13 | 2.42% | -4 | 22.5 | 2.93 |
Washington | 12 | 2.04% | -15 | 8 | 0.96 |
West Virginia | 5 | 0.93% | 26 | 100 | 5.00 |
Wisconsin | 10 | 1.86% | -6 | 21 | 2.10 |
Wyoming | 3 | 0.56% | 40 | 100 | 3.00 |
Totals | 538 | 100.00% | |||
red = swing state (odds between 30 and 70) | |||||
270 needed to win | Republican | 238.6 | |||
Democrat | 299.4 |
There are caveats.
First, since the election is four months out, the data is pretty thin (i.e. the state-by-state markets are very thinly traded right now - some haven't even traded at all, in which case I relied on the closing data from 2016).
Second, when I cited the model's accuracy, what I was really referring to was the last snapshot before each election. During the months-long run-up, there will be plenty of fluctuation, but it will still tend to be the best indicator at any point in time. Obviously, as circumstances change, so does the forecast.
Why does this work better than polls? Because it's real people betting real money, not people answering polls on the phone where their responses could be shaded for any number of reasons - particularly this year!
I will start to graph all this and post more frequent updates as the election draws nearer.
Naked Dollar readers with good memories will also remember I have another model that shows directional movement for both campaigns. I'll get to that shortly.
Really interest, Scott! Look forward to following.
ReplyDelete