AIR U.S. Presidential Election Algorithm Falsified
by Eric Schulman and Daniel Debowy
The Annals of Improbable Research United States Presidential Election Algorithm (Debowy and Schulman, AIR Online, 20 October 2003) was developed based on the experience of the major party candidates for president and vice president in each of the 54 United States presidential elections between 1789 and 2000. It correctly predicted the outcome of the 2004, 2008, and 2012 United States presidential elections, but it did not correctly predict the outcome of the 2016 United States presidential election. The algorithm predicted that the Democratic ticket of Hillary D. R. Clinton and Timothy M. Kaine would defeat the Republican ticket of Donald J. Trump and Michael R. Pence on November 8, 2016, but this did not happen.
The concept of falsifiability is important to the progress of science. A theory needs to be testable in order to determine whether its predictions agree with reality. If it can never be falsified (for example, if one could always modify the theory to make it correct after the fact), then it is not a good theory. The theory that the outcome of United States presidential elections could be predicted based on the experience of the major party candidates for president and vice president was a testable theory. You can never prove a scientific theory beyond all doubt, but you should be able to test the theory to gain confidence in its correctness.
We obtained another data point every four years since 2004 and each success increased our confidence in the theory. Three successes out of three trials showed that it was possible the algorithm would always work, but it did not exclude the possibility that the successes were caused by random chance (those data suggested with 95% confidence that the algorithm would correctly predict the results between 37% and 100% of the time). Had the algorithm correctly predicted the 2016 and 2020 elections, then we could have concluded with more than 95% confidence that it was better than random chance.
The outcome of the 2016 United States presidential election showed that the algorithm does not always work. Three successes out of four trials allows us to conclude with 95% confidence that the algorithm will correctly predict between 20% and 99% of United States presidential elections. As long as the winner always comes from one of two parties, merely flipping a coin would correctly predict the outcome of a United States presidential election 50% of the time. If we can’t be confident that the algorithm correctly predicts outcomes more than 50% of the time, it is not useful.
Therefore, we have concluded that the Annals of Improbable Research United States Presidential Election Algorithm is not a useful method for predicting the outcome of United States presidential elections.