Keys to the White House 2020
UPDATE: This blog is a copy of a blog I authored in 2016 right before the crazy outcome of the Clinton - Trump election. Lichtman has release his outcome for 2020; and he predicts a win for Biden due to 7 of the 13 keys falling his way. Let's see if he's right this time!
My blog from 2016:
In 1981, a historian named Allan Lichtman had dinner with a geophysicist, Vladimir Keilis-Borok. During that dinner the two discussed the idea of using concepts from earthquake research to predict the outcome of elections. After that discussion, Mr. Lichtman, devised his "13 Keys to the White House" or a way to use pattern detection to predict if a party will retain or lose control of the U.S. presidency in the upcoming election. So far his simple algorithm is flawless - predicting the winner of every election since 1984. His 13 keys were also able to predict Trump's victory in 2016 despite a host of political pollsters and pundits predicting a Hillary Clinton win.
Statistics Got it Wrong
Mr Lichtman's against the grain prediction highlighted something for me. My beloved statistics got it wrong. You see, most major news outlets and pollsters gave Hillary strong chances to win the election. They were wrong. This guy was right. Why?
I believe it's because he used machine learning (ML) techniques - perhaps without even knowing it. Data reduction (13 questions), pattern detection and prediction algorithms ignore why... They don't care about the news cycle, the email scandals, the accusations... These models only care about predicting and focus on the data that predicts.
Can Machine Learning Get It Right?
The reason I'm writing this post is not to congratulate Mr. Lichtman. I actually don't believe he knows what he has stumbled on to. This short post is simply a reminder to myself to think outside of business and science applications for ML. Statistics and polling have dominated election news and are big business for big media. Unfortunately, they are merely observations. To accurately predict, you'll need concepts from machine learning - because stats don't really matter.
(Apologies to 538 and Nate Silver - I still love your site.)