Show HN: Predicting the winner of the 2016 US presidential election
I have set up a website for presidential election predictions (http://prespredict.com), and it is now up and running for the 2016 election. It uses daily state poll data and calculates the probability that each candidate will win. It also estimates the most likely number of electoral votes each candidate will get.
I originally did this back in 2008, and then applied this to the 2012 election. The methodology was also retroactively applied to the 2004 election using available state polls. In all three past elections, the calculation correctly predicted the winner. It also had a pretty good estimate of the number of electoral votes that each candidate ended up getting.
The approach and the website are not as fancy as 538.com but the results so far are pretty good. The methodology is explained here: http://prespredict.com/methodology.html.
It's a bit early and there are several caveats, including the accuracy of polls in the current political environment, the independence assumption in the model, the choice of the winning probability curve, etc. Since the predictions for the last three elections have been pretty accurate, I decided to leave the calculation as is, despite the presence of the caveats.
The website is automatically updated daily at 8am PST, so you can check back periodically. Cool model.
Did you look at empirical data to derive the shape of the curve? The output is very different from 538's. Any intuition for what part of the model is behind that? I did look at some empirical data to derive the shape of the curve, though the derivation is not as thorough as it can be. If I get time to make the derivation more formal, I'll post in in the methodology section. One difference I have with 538 is that, I believe, they have some model of the state results, maybe with some correlation, and then simulate the election many times to come up with the PDF of results. On the other hand, I assume the state results are independent (conditioned on the latest polling) and so I am able to derive a closed form calculation of the PDF, with no simulations. So the independence assumption makes my calculation easier, though may not be as accurate as a model with correlation between states. However, the model has done quite well for the past three elections, so maybe the independence assumption is not that bad.