Preprint / Version 1

Predicting Election Outcomes from Facial Images of Candidates Using an Unbiased Machine Learning Model


  • Raymond Webber Academy
  • Jadelyn Tran Webber Academy
  • Tyler Giallanza


election prediction, machine learning, classifier, transfer learning, voter bias, physiognomy


In the past, studies on psychology have shown that humans are capable of creating instantaneous judgments of a stranger’s personality and characteristics, just based on a picture of their face. The most successful study to our knowledge has reached a 72.4% accuracy in predicting election outcomes. There have been machine learning studies that tried to replicate this success, but to our knowledge, some kind of human input and bias has always been present. We wanted to create a bias-free, independent machine learning model that only used the image of political candidates to predict their success. With no other information than a candidate’s face, we were able to achieve a 70.43% accuracy predicting election results. Not only does the different approaches in our experiment give a quantitative way to compare different types of human thinking, but it can be used as a benchmark for future research that further investigates the relationship between facial traits, human judgments, and machine learning.

References or Bibliography

Arora, N. (2021). Age, Gender and Ethnicity (Face Data) csv. Kaggle.

Ballew, C. C., & Todorov, A. (2007, November 13). Tipping elements in the Earth's climate system | PNAS. Predicting political elections from rapid and unreflective face judgments. Retrieved January 2, 2023, from

Google Colaboratory. (2017). computer software. Retrieved January 2, 2023.

Joo, J., Steen, F. F., & Zhu, S.-C. (2015, December). Automated Facial Trait Judgment and Election Outcome Prediction: Social Dimensions of Face. International Conference on Computer Vision.

Olivola, C. Y., & Todorov, A. (2010, January 23). Elected in 100 milliseconds: Appearance-Based Trait Inferences and Voting.

Olivola, C., & Todorov, A. (2009, May 5). The Look of a Winner. Retrieved January 2, 2023, from

PyTorch. (n.d.). Version (1.13.1). From Research To Production. Retrieved January 2, 2023, from

scikit-learn. (n.d.). Version (1.2.0). scikit-learn Machine Learning in Python. Retrieved January 2, 2023, from

Serengil, S. I. (2022, May 10). DeepfaceVersion (0.0.75). deepface 0.0.75. PyPI. Retrieved January 2, 2023, from

Todorov, A. (2018, May 14). Can we read a person's character from facial images? Scientific American Blog Network. Retrieved January 2, 2023, from

Todorov, A., Mandisodza, A. N., Goren, A., & Hall, C. C. (2005, June 10). Inferences of Competence from Faces Predict Election Outcomes. Washington D.C.; American Association for the Advancement of Science.

Ventura, C., Masip, D., & Lapedriza, A. (n.d.). Interpreting CNN Models for Apparent Personality Trait Regression. Institute of Electrical and Electronics Engineers.