Can Ai Predict the Presidential Election?
While AI has great promise across many fields, it may be overkill for predicting the outcome of a U.S. Presidential election.
With roughly 9 “swing states” and the remaining states’ outcomes essentially pre-determined, there aren’t too many “plausible” outcomes to consider.
Forecasters such as 538 try to use polling to predict how these swing states will vote. A crude model would treat each swing state as equally likely to vote for a Democrat or a Republican, and such a model would have to predict 9 “bits” of information or choose one of 2^9 = 512 outcomes.
However, each state isn’t equally likely to swing in either direction, and polling uncertainties between states are correlated, which means that some outcomes are far less plausible than other outcomes, and thus the effective amount of information or the number of bits that need to be determined is less than 9. A cautious estimate might be 7 or 8 bits.
AI and Machine Learning benefit from additional data and many outcomes to train and test against, but the infrequency of presidential elections means that validating such methods with direct data is limited.
While methods that use transfer learning from the more frequent house and senate elections, may in theory give more precise presidential election odds, to what ends does more accurately distinguish a 72% win probability versus a 74% win probability materially benefit a forecaster for an event that occurs once every 4 years?
The whole election has maybe 8 bits of entropy. One could just randomly choose swing states with a coin flip and call it a model. But to be more accurate what would we need?
We can probably define it as a binary classification problem.
States are pretty correlated, and swing states are as well.
So the net amount of information entropy is far less than 50 bits.
One can basically ignore all the non-swing states.
We get around 9 ‘swingish’ states. Drop one bit of entropy for state to state correlations to be conservative. In reality probably way more correlation?
2^8 = 256 “reasonable” outcomes
The entropy of the very presidential election -- the effective dimensionality (the sum of singular values of the PCA, as it's often taken) is most likely far below 8.
Edited by Samson Qian, Qilin Guo Thomas Braun, Yunfei Chen, Dima Korolev, Alexander Fleiss, Calvin Ma, Helen Wu, Hongyi Liang, Wenxin Mu & James Rhinelander