This repository contains the source code and PDF for the MSc Data Science group thesis “Forecasting Election Outcomes in Contemporary Democracies”, completed at the University of Birmingham. The project was awarded a Distinction for its ambition, relevance, and technical execution, demonstrating effective use of machine learning methods such as PCA for dimensionality reduction and clustering for stratified modelling.
The study found that national social and economic conditions are strong predictors of election outcomes; when conditions worsen, voters are more likely to seek a change in government.
The main challenge was data sparsity. For example, in the UK, elections occur roughly every four years, with reliable data only available for the past 100–200 years. This limits training data to ~25–50 observations, which was insufficient for complex modelling.
To address this, data from multiple countries was aggregated. However, this introduced variance in political contexts and data quality (many countries have only recorded elections since the 1990s). To mitigate this, countries were clustered, and separate models were trained for each group.
- Connor Boyd-Lyon
- Yarkın Yorulmaz
- Evelina Ivanova
- Matthew Randall
- Yeni Jung
- Zehn-Ul-Abideen Sharif
This repository serves as an archive and is not actively maintained. A cleanup is scheduled for October 2025. Questions are welcome, but future updates are not guaranteed.
MIT licensed. Please cite as follows:
@mastersthesis{electionforecasting2025,
author = {Connor Boyd-Lyon and Evelina Ivanova and Matthew Randall and Yarkın Yorulmaz and Yeni Jung and Zehn-Ul-Abideen Sharif},
title = {Forecasting Election Outcomes in Contemporary Democracies},
school = {University of Birmingham},
year = {2025},
type = {Group Thesis},
note = {MSc Data Science group project}
url = {https://github.com/connorlyon10/ElectionOutcomes}
}