CES Hackathon Project
A reproducible election-forecasting pipeline that benchmarks theory-driven voter-level regression against simple survey aggregation, then improves performance via Pareto-optimal feature selection, calibration, and cross-year validation.
Read time
7 min read
Published date
Feb 9, 2026
Category
Machine Learning
Weekly Newsletter Update!
Stay up-to-date with the latest innovations, features, and tips in no-code website building!
Why this project (and why “poll averaging” is not the whole story)
Poll or survey aggregation is a strong baseline because it is transparent and resistant to overfitting. But aggregation also collapses the mechanism: it summarizes what respondents report, not how combinations of ideology, retrospective evaluations, trust, and issue salience jointly map into vote choice.
This project frames election forecasting as a comparative research design:
Baseline: simple aggregation of survey signals
Alternative: voter-level models (regression) that encode voting-behavior theory
Improvement layer: auditable, multi-objective feature selection that trades off accuracy and complexity
Central question: Can modeling individual voter reasoning produce better forecasts than simply averaging poll responses?
Data and scope
The pipeline is explicitly built around major public election-study infrastructures
ANES (US): 2012, 2016, 2020
CES (Canada): 2011, 2015, 2019, 2021, 2025
The Canadian setting is particularly demanding because forecasting is not just “two-party swing.”
The workflow supports multinomial vote choice (six-party classification: CPC/LPC/NDP/Bloc/Green/PPC) and treats survey weighting as first-class rather than an afterthought.
Methodology: a two-step comparative design
Step 1: Direct replication (theory-driven)
Step 1 follows Camatarri (2024) using manual feature selection motivated by voting-behavior theory. The repository explicitly enumerates feature families such as:
ideology (left-right self-placement)
retrospective economic evaluations (national economy, personal finances)
trust in government
issue salience (climate, housing, immigration)
demographics (age, gender, education, ethnicity)
province fixed effects
Models include:
Multinomial logistic regression (multi-party vote choice)
Bayesian logistic regression (MCMC: 4 chains, 2000 iterations)
Aggregation baseline: weighted mean threshold classification with survey weights, compared against simple poll/survey aggregation.
The point is not that theory-driven selection is “best,” but that it provides a transparent benchmark against which any automated selection must justify itself.
Step 2: PPV–Pareto enhancement (data-driven, constraint-aware)
Models include:
Multinomial logistic regression (multi-party vote choice)
Bayesian logistic regression (MCMC: 4 chains, 2000 iterations)
Aggregation baseline: weighted mean threshold classification with survey weights, compared against simple poll/survey aggregation.
The point is not that theory-driven selection is “best,” but that it provides a transparent benchmark against which any automated selection must justify itself.
Copyright © 2024 – All Right Reserved