Title

Using Recurrent Neural Networks to Model Political Outcomes

Author(s)

Ethan FultonFollow

School Name

South Carolina Governor's School for Science & Mathematics

Grade Level

12th Grade

Presentation Topic

Computer Science

Presentation Type

Mentored

Abstract

In an increasingly data-driven world, political scientists and statisticians are searching for new methods of predicting election outcomes. Traditionally, these forecasts have relied on polling as well as standard regression models. As machine learning becomes an increasingly robust field, it is natural to consider extending its applications to this area. Recurrent neural networks are an effective tool for time-series forecasting, and this study evaluated their applicability in political forecasting. Given datasets with and without 2012 election data, a recurrent neural network produced models that were, respectively, 97% and 92% accurate at the county level. The Keras and TensorFlow libraries were effective in facilitating a machine learning model for this purpose. Changing factors within the neural network such as the test set size, the number of hidden layers, and the optimizer had a measurable but not drastic effect on the overall accuracy. Demographic factors within the model that increased accuracy the most were the racial statistics, percent in professional occupations, and percent driving alone to work. The model was most accurate in Southern states but no other trends appeared, with North Carolina being the most accurate large state and Illinois the least. The accuracy of this model showcases that neural networks have promising applications in the political realm, and can be applied to future modeling at a sufficiently granular level.

Location

Furman Hall 109

Start Date

3-28-2020 10:15 AM

Presentation Format

Oral Only

Group Project

No

COinS
 
Mar 28th, 10:15 AM

Using Recurrent Neural Networks to Model Political Outcomes

Furman Hall 109

In an increasingly data-driven world, political scientists and statisticians are searching for new methods of predicting election outcomes. Traditionally, these forecasts have relied on polling as well as standard regression models. As machine learning becomes an increasingly robust field, it is natural to consider extending its applications to this area. Recurrent neural networks are an effective tool for time-series forecasting, and this study evaluated their applicability in political forecasting. Given datasets with and without 2012 election data, a recurrent neural network produced models that were, respectively, 97% and 92% accurate at the county level. The Keras and TensorFlow libraries were effective in facilitating a machine learning model for this purpose. Changing factors within the neural network such as the test set size, the number of hidden layers, and the optimizer had a measurable but not drastic effect on the overall accuracy. Demographic factors within the model that increased accuracy the most were the racial statistics, percent in professional occupations, and percent driving alone to work. The model was most accurate in Southern states but no other trends appeared, with North Carolina being the most accurate large state and Illinois the least. The accuracy of this model showcases that neural networks have promising applications in the political realm, and can be applied to future modeling at a sufficiently granular level.