Historically Anchored LLM Improves Public Opinion and Election Prediction Accuracy
ORIGINAL / Matching-LLM: Historically Anchored Synthetic Prediction of Public Opinion and Elections
This study introduces Matching-LLM, a weighted framework combining LLM-generated responses with historical human behavioral data to improve public opinion and election predictions; its novelty lies in calibrating LLM outputs with historical data and validating across multinational surveys and the 2024 U.S. election.
01 ABSTRACT
The study proposes Matching-LLM, a weighted framework that integrates LLM responses with historical human behavioral data to reconstruct survey responses and predict election outcomes. Results show it generally improves accuracy over LLM-only, including 11 of 12 comparisons on survey questions excluded from calibration, though it does not consistently outperform History-only. In the 2024 U.S. presidential election, Matching-LLM predicts 310 Republican electoral votes versus the actual 312, outperforming both single-source benchmarks; the authors conclude historical data can improve LLM-based predictions but task-specific validation is essential.
02 KEY FINDINGS
- Matching-LLM outperforms LLM-only in 11 of 12 survey questions excluded from calibration.
- It does not consistently beat the History-only benchmark, indicating historical data itself contributes substantially.
- For the 2024 U.S. election, it predicts 310 Republican electoral votes versus the actual 312, outperforming both single-source benchmarks.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV Q-FIN