Predicting Voter Choice from Census Data Isaac Evavold
Motivation Election models built by hand using vote history, polling data Interested in viability of predictions made using census data and neural network Build a network that could make voting predictions about arbitrary groups of voters given aggregate statistics about group
Data Data required a lot of pruning and consolidation Collected from 2010 US Census Aggregated per state and per congressional district Pro: Approximately same population nationwide Con: Produces a limited dataset to train and evaluate model Data required a lot of pruning and consolidation
Methods Feature vector includes: District population Racial composition Median income Education level Each district was classified by whether it voted for the Democratic or Republican presidential candidate in 2016 Neural network used 4 hidden layers with 100-75-50-25 neurons
Results & Conclusions Network was able to make correctly predict vote ~64% of the time More experiments with different feature vectors or more sophisticated data could yield better predictions This approach is still limited by using census data which cannot account for current events, only demographic voting trends