复旦大学-重磅：2020年美国总统选举预测（英文）-2020.11-6页_368kb

2020 United States Presidential Election Forecast Based on ABM Simulation

Core Content

This document presents the results of an agent-based modeling (ABM) simulation conducted by the Center for Complex Decision Analysis (CCDA) at Fudan University, Shanghai, China, on November 1, 2020, two days before the U.S. presidential election. The simulation focuses on the relative vote shares in six key states and aims to predict the electoral college outcome using a combination of ABM and polling data.

Key States and Predicted Results

Main Points

• Methodology: The research team used agent-based modeling (ABM), a computational simulation method that models individual voter behavior and interactions, to forecast the election. The ABM approach does not rely on public opinion polls.

• Model Variants: Two main model groups were used:

  • Group A incorporates employment rates across different sectors (agriculture, manufacturing, others).
  • Group B uses ethnic background as a key variable instead of employment data.

• Scientific Focus: The team emphasized that their forecasts are a scientific exercise, with no intention of influencing actual election outcomes.

• Vote Share Prediction: The models predict relative vote shares (i.e., the share of votes obtained by a candidate compared to the total share of both major parties). This provides a more nuanced understanding than just predicting the winner.

• Electoral College Forecast: By combining ABM results with polling data, the team also predicted the overall outcome of the presidential election. According to the simulation, Biden & Harris has an 89.6% to 92.4% chance of winning, with an average of 91.1%, while Trump & Pence has an 7.1% to 9.6% chance, averaging 8.3%.

Key Information

• Project Background: This is part of a yearlong project by Prof. Shiping Tang's team at Fudan University. They have previously accurately predicted election outcomes in Taiwan (2016, 2020) and U.S. Senate elections (2018).

• Simulation Details: The team conducted three simulations in April, July, and September 2020, with the September results being released publicly. They chose six states for the 2020 election due to budget constraints.

• Data Sources: The team acknowledged the Intellisia Institute for helping with data collection. The ABM method combines individual-level voter data with structural-level economic and social data.

Technical Approach

• The team integrates long-term steady effects and short-term fluctuation effects in their models.
• They use Monte Carlo simulations to explore various possible scenarios and determine the likelihood of a candidate crossing the 270 electoral college vote threshold.
• The models are designed to be standardized, rigorous, and computationally sound.

Conclusion

This ABM-based forecast represents a novel scientific approach to election prediction, offering detailed insights into vote shares and electoral college probabilities without relying on opinion polls. The results highlight the potential of ABM in political science and demonstrate the team's ongoing commitment to accurate and reliable election forecasting.

Appendix References

• Appendix-I: Describes the method of combining ABM results with polling data to predict the overall outcome of the election.
• Appendix-II: Provides links to earlier forecasting efforts, including the 2016 Taiwan presidential election, 2018 U.S. Senate elections, and 2018 Taiwan local elections.