重磅！复旦大学唐世平团队基于模型对2020美国总统选举进行预测，结果是....

用微信扫码二维码

分享至好友和朋友圈

2020年美国总统选举:

基于ABM的仿真模拟预测

（中文版）

2020年11月1日

复旦大学复杂决策分析中心

中国·上海

北京时间2020年11月1日中午12点（美国总统大选前两天），唐世平教授担任主任的复旦大学“复杂决策分析中心”研究团队正式发布最新的基于计算机模拟的选举预测结果，该结果预测了将于2020年11月3日举行的第59届美国总统大选中共和党候选人和民主党候选人在六个州的“相对得票率”。

特别说明：“相对得票率”等于某一个党的候选人的真实得票率除以民主党与共和党两位候选人真实得票率之和。

实际上，我们团队在2020年9月初就已经通过模拟得出了这些预测结果，且分别于今年的4月、7月和9月先后根据不同时期的数据和信息共计模拟了三次。我们之所以没有在较早时间公布这些预测结果是为了避免影响实际的投票过程。

我们的预测是基于ABM模型（Agent-based Modeling）的仿真模拟，完全不依赖民意调查，因此该方法在民调数据相对缺乏的州依然适用。我们从2015年初就开始研发和测试这种方法。我们先前的各项预测表现也相当准确，详见本报告末尾的相关链接（附录-2）。

今年，我们基于ABM的美国大选预测主要集中在6个州（迫于研究预算的限制）。必须强调的是，该预测项目是一项纯粹的科学活动，团队无意以任何方式影响美国的实际选举过程和结果。显然，普通大众更关心的是整个大选的结果。因此，通过将我们的方法与民调预测模型相结合，最终我们将预测范围扩展到美国全国选举团票数的分布情况，以及双方的胜选概率（附录-1）。感谢“海国图智研究院”对本研究在数据收集期间所提供的部分协助。

预测结果汇总表

注：A组模型与B组模型的一个重要区别是，除了共享的变量之外，前者还包含了各行业（农业、制造业或其他）的就业比例变量，而后者则包含族裔背景变量。

1.密歇根州

两组预测模型均预测拜登-哈里斯将赢得密歇根州的选举，平均而言获得 54.54% 或55.57%的选票。第一组模型（A组模型）预测拜登和哈里斯的相对得票率在53.18%（最低值）到55.91%（最高值）之间。第二组模型（B组模型）预测拜登和哈里斯的相对得票率在 54.18%（最低值）到56.95%（最高值）之间。

2.俄亥俄州

两组预测模型均预测特朗普-彭斯将赢得俄亥俄州的选举，平均而言获得 50.75%或 50.89%的选票。第一组模型（A组模型）预测特朗普和彭斯的相对得票率在49.52%（最低值）到 51.98%（最高值）之间。第二组模型（B组模型）预测特朗普和彭斯的相对得票率在 49.66%（最低值）到 52.12%（最高值）之间。

3. 宾夕法尼亚州

两组预测模型均预测拜登-哈里斯将赢得宾夕法尼亚州的选举，平均而言获得 52.04%或52.44%的选票。第一组模型（A组模型）预测拜登和哈里斯的相对得票率在 50.74%（最低值）到 53.34%（最高值）之间。第二组模型（B组模型）预测拜登和哈里斯的相对得票率在 51.13%（最低值）到53.75 %（最高值）之间。

4.印第安纳州

两组预测模型均预测特朗普-彭斯将赢得印第安纳州的选举，平均而言获得 51.65%或53.64%的选票。第一组模型（A组模型）预测特朗普和彭斯的相对得票率在50.44%（最低值）到52.85%（最高值）之间。第二组模型（B组模型）预测特朗普和彭斯的相对得票率在 52.48%（最低值）到 54.80%（最高值）之间。

5.西弗吉尼亚州

西弗吉尼亚州，由于该州的族裔同质化程度比较高，我们的模拟只能得到一组模型（A组模型）。

我们的预测模型预测，特朗普和彭斯将赢得西弗吉尼亚的选举，平均而言获得61.69%的选票。模型（A组模型）预测特朗普和彭斯的相对得票率在60.73 %（最低值）到 62.65%（最高值）之间。

6.密苏里州

两组预测模型均预测特朗普-彭斯将赢得密苏里州的选举，平均而言获得55.60 %或55.39%的选票。第一组模型（A组模型）预测特朗普和彭斯的相对得票率在54.49%（最低值）到 56.71%（最高值）之间。第二组模型（B组模型）预测特朗普和彭斯的相对得票率在54.27%（最低值）到 56.50%（最高值）之间。

上述预测是唐世平教授团队在上海复旦大学进行的为期一年的项目研究成果。它是继团队对2016年“台湾地区领导人选举”、2018年美国西维吉尼亚州和密苏里州参议院中期选举、2018年台湾地区台北、新北市和桃园市长选举以及2020年“台湾地区领导人选举”的精准预测之后的又一成果。更多信息，请访问我们的官方网站：http://www.ccda.fudan.edu.cn/

在模拟的过程中，团队力求做到：（1）在政治学选举理论指导下进行变量选择和模型构建；（2）结合选民和候选人的个体层面数据与经济社会发展的结构层面数据；（3）结合长期稳定效应和短期波动效应，试图通过标准化的模型、严谨的程序和算法来更准确地预测选举。通过引入ABM仿真模拟，本研究是一次探索新的选举预测科学方法的有益尝试。

唐世平及其同事所开发的预测方法避免了对民调数据的过度依赖。换言之，我们基于ABM的选举预测不需要任何民意调查的数据。因此，我们的预测技术能够特别在民调数据稀缺或者民调数据质量欠佳的国家或地区，依然能够实现精准预测。更重要的是，我们的方法不仅能预测哪位候选人胜选，而且可以基于基准模型给出每位候选人的得票率和得票区间。

本团队现在正在撰写学术研究论文，未来将提供更多关于该预测方法的技术细节。

附录1：预测2020年美国总统选举的总体结果

由于普通民众对总统大选总体结果的兴趣远大于候选人在具体州的得票率，因此我们也将发布对大选总体结果的扩展预测。

这里必须强调的是，在这部分的预测中，除了上述6个州基于ABM的预测结果之外，我们对其它各州的预测运用了民调数据（未来，在研究预算允许的情况下，我们将全部用ABM结果替代）。然后，我们要做的是利用蒙特卡洛模拟方法来模拟每位候选人获得超过270张选举人票（即获胜）的情形和几率。我们得出以下预测结果：拜登-哈里斯赢得选举的几率在89.6%-92.4%之间（均值为91.1%），特朗普-彭斯赢得总统选举的几率在7.1%-9.6%之间（均值为8.3%）。当然，尽管我们预测特朗普连任将是小概率的事件，这不意味着他重新当选的概率是零。

图1：美国2020大选预测模拟结果

您可以从以下链接获得本新闻发布稿的PDF文件：

中文版本：

http://www.ccda.fudan.edu.cn/uploads/2020/10/311544084521.pdf

英文版本：

http://www.ccda.fudan.edu.cn/uploads/2020/10/311932521510.pdf

附录2：团队预测工作的以往发布情况

（1）2016年01月05日（中文版）和2016年01月10日（英文版），也就是当年“台湾地区领导人选举”投票日之前，唐世平团队发布了对这次选举结果的预测。这是我们运用新方法进行选举预测的第一次尝试。请点击如下链接了解详情。

中文版本：

http://www.ccda.fudan.edu.cn/index.php?c=article&id=75

英文版本：

https://china.ucsd.edu/_files/01112016_Taiwan-election.pdf

https://chinafocus.ucsd.edu/2016/01/10/taiwan-election-results-predicted-in-computer-simulation/

（2）2018年11月04日，唐世平团队在美国参议院选举投票之前，发布了对两个州（西弗吉尼亚州和密苏里州）的选举结果预测。这是我们运用新方法进行选举预测的第二次尝试。请点击如下链接了解详情。

中文版本：

http://blog.sina.com.cn/s/blog_744a73490102y7bd.html

http://www.ccda.fudan.edu.cn/index.php?c=article&id=80

英文版本：

http://www.ccda.fudan.edu.cn/index.php?c=article&id=81

https://www.linkedin.com/pulse/united-states-senate-elections-west-virginia-missouri-tang-dr/?trackingId=shAzPfa2GP9KRu%2BbwvnIyQ%3D%3D

预测结果与实际结果的比较：

中文比较版本：

https://sirpa.fudan.edu.cn/info/1079/2741.htm

英文比较版本：

https://www.linkedin.com/pulse/prelimary-assessment-our-forecasting-excellent-perhaps-tang-dr/

（3）2018年11月22日，唐世平团队在选举投票日之前发布了台湾地区地方选举（台北、新北、桃园）的预测结果。这是我们运用新方法进行选举预测的第三次尝试。请点击如下链接了解详情。

中文版本：

http://www.ccda.fudan.edu.cn/index.php?c=article&id=83

英文版本：

http://www.ccda.fudan.edu.cn/index.php?c=article&id=84

https://www.linkedin.com/pulse/taiwan-local-elections-taipei-new-taoyuan-forecasting-tang-dr/?trackingId=VmpIAUnefnghFHhzA9ODeg%3D%3D

预测结果与实际结果的比较：

中文版本：

http://www.ccda.fudan.edu.cn/index.php?c=article&id=85

（4）2020年1月9日，唐世平团队在选举投票日之前发布了对2020年1月11日举行的“台湾地区领导人选举”的预测结果。这是我们运用新方法进行选举预测的第四次尝试。英文版本链接如下：

https://www.linkedin.com/pulse/2020-taiwan-elections-forecasted-computer-simulation-shiping-tang-dr/

对该次预测的初步评估：

中文版本：

http://www.ccda.fudan.edu.cn/index.php?c=article&id=96

英文版本：

http://www.ccda.fudan.edu.cn/index.php?c=article&id=97

2020 United States Presidential Election:

Forecasting Based on ABM Simulation

（English Edition）

November 1, 2020

Center for Complex Decision Analysis

Fudan University, Shanghai, China

At 12pm (Beijing time) on November 1, 2020 (two days before the presidential election in United States), a research team at the Center for Complex Decision Analysis (directed by Prof. Shiping Tang), of Fudan University, Shanghai, China, releases their latest computer simulation-based predictions for the "relative shares of votes" in six states in the upcoming 59th American presidential election that is scheduled on November 3, 2020.

By "relative shares of votes", we mean the share of votes by one party’s candidate divided by the total share of votes obtained by the two major parties (i.e., the Democratic Party and the Republic Party).

For these six states, our simulation has produced these forecasted results in early Sept. In fact, we have simulated three times: April, July, Sept. We have refrained from releasing our forecasting results to avoid affecting actual voting.

Our forecasting is based on agent-based modeling (ABM), with zero reliance on opinion polls. We have been developing this method since early 2015. For the releases of our earlier forecasting efforts, which have also been fairly accurate, please see the links at the end of this report (Appendix-II).

For 2020 U.S. election, our ABM-based forecasting primarily focuses on six states (due to the budget limit). It must be emphasized that our forecasting is a purely scientific exercise. The team has no intention of influencing actual elections in the U.S. by any means.

Obviously, the general public is more interested in the overall outcome of the election. We therefore also extend our forecasting to predict the final electoral college votes by combining our methods with polls projection (Appendix-I).

We appreciate the assistance "Intellisia Institute" provided in collecting part of the data.

Results of Our Forecasting

Note: A key difference between group A model and group B model is that except for the shared variables, the formers contain employment rates among different sectors (agriculture, manufacturing, or others) yet the latter replaces these variables with "ethnic background" variable.

1.Michigan

Our two sets of forecasting models predict that Biden&Harris will win the election with 54.54% or 55.57% of the popular votes in Michigan, on average. The first set of models (group A models) predict Biden&Harris will receive anywhere between 53.18% (lowest) and 55.91% (highest) of the total votes. The second set of models (group B models) predict that Biden&Harris will receive anywhere between 54.18% (lowest) and 56.95% (highest) of the total votes.

2. Ohio

Our two sets of forecasting models predict that Trump&Pence will win the election with 50.75% or 50.89% of the popular votes in Ohio, on average. The first set of models (group A models) predict Trump&Pence will receive anywhere between 49.52% (lowest) and 51.98% (highest) of the total votes. The second set of models (group B models) predict that Trump&Pence will receive anywhere between 49.66% (lowest) and 52.12% (highest) of the total votes.

3.Pennsylvania

Our two sets of forecasting models predict that Biden&Harris will win the election with 52.04% or 52.44% of the popular votes in Pennsylvania, on average. The first set of models (group A models) predict Biden&Harris will receive anywhere between 50.74% (lowest) and 53.34% (highest) of the total votes. The second set of models (group B models) predict that Biden&Harris will receive anywhere between 51.13% (lowest) and 53.75%(highest) of the total votes.

4. Indiana

Our two sets of forecasting models predict that Trump&Pence will win the election with 51.65% or 53.64% of the popular votes in Indiana, on average. The first set of models (group A models) predict Trump&Pence will receive anywhere between 50.44% (lowest) and 52.85% (highest) of the total votes. The second set of models (group B models) predict that Trump&Pence will receive anywhere between 52.48% (lowest) and 54.80% (highest) of the total votes.

5. West Virginia

For West Virginia, our simulation can only obtain one set of models (group A) due to the fact that ethnic diversity is low in the state. Our one set of forecasting models predicts that Trump&Pence will win the election with 61.69% of the popular votes in West Virginia, on average. The models predict Trump&Pence will receive anywhere between 60.73% (lowest) and 62.65% (highest) of the total votes.

6. Missouri

Our two sets of forecasting models predict that Trump&Pence will win the election with 55.60% or 55.39% of the popular votes, on average. The first set of models (group A models) predict Trump&Pence will receive anywhere between 54.49% (lowest) and 56.71% (highest) of the total votes. The second set of models (group B models) predict that Trump&Pence will receive anywhere between 54.27% (lowest) and 56.50% (highest) of the total votes.

Summary

Our forecasting is the outcome of a year long project conducted by Tang’s team at Fudan University in Shanghai. It follows our team’s accurate forecasting of the Taiwanese presidential election in 2016, the mid-term US Senate elections in West Virginia and Missouri in 2018, the mayoral elections of Taipei, New Taipei and Taoyuan in Taiwan in 2018, and the Taiwan 2020 general election. For more information, please visit our official website:

http://www.ccda.fudan.edu.cn

In the process of simulations, the team strives to (1) conduct variable selection and model construction under the guidance of election theories in political science; (2) combine individual-level data of voters and candidates with structural-level data of economic and social development; (3) combine long-term steady effects and short-term fluctuation effects, in an attempt to forecast the election more accurately through standardized models, rigorous procedures, and computational algorithm. By introducing ABM simulation, this research represents a meaningful attempt in exploring a new scientific method of election forecasting.

The method developed by Tang and his colleagues circumvents the need of relying on polling data. In other words, our ABM-based forecasts do not have any input from public opinion polls.As a result, our ABM-based method can still deliver (potentially) accurate predictions, even in places where poll data is scarce or poor in quality. More importantly, our method does not merely predict which candidate is going to win or lose, but their share of votes within a predictive interval dictated by their baseline models.

We are now drafting a full paper that will provide more technical details about our approach.

Appendix-I: Forecasting the Overall Election Outcome

Because the general public is far more interested in the overall outcome of the presidential election than the percentage of votes obtained by candidates in specific states, we also release our extended forecasting for the overall outcome.

What needs to be emphasized here is that for this part of forecasting, we can only rely on polls obtained by polling houses or agencies, except for the ABM-based forecasted results in the six states we have shown above. What we do is then to use Monte-Carlo simulation to simulate the possible scenarios for one candidate to cross the threshold of 270 electoral college votes, which means an electoral triumph.

Our simulations produced these following results: Biden&Harris has a probability of 89.6% to 92.4% (91.1% on average) for winning the presidential election, and Trump&Pence has a probability of 7.1% to 9.6% (8.3% on average) for winning the presidential election. Even though our forecasting shows that Trump's re-election probability is very low, this does not mean that the probability of Trump being re-elected is zero.

Picture I: Our US 2020 Presidential Election Forecasting Result.

You can download PDF version of this report from here:

English Version:

http://www.ccda.fudan.edu.cn/uploads/2020/10/311932521510.pdf

Chinese Version:

http://www.ccda.fudan.edu.cn/uploads/2020/10/311544084521.pdf

Appendix-II: Earlier Releases of Our Previous Forecasting Efforts

(1) Four years ago, on Jan. 05, 2016 (Chinese version) and Jan. 10, 2016 (English version), that is, before the election in Taiwan, Tang's team released our forecasting of this election. That prediction was our first attempt of election forecasting with our new approach. For details, please see:

Chinese Version:

http://www.ccda.fudan.edu.cn/index.php?c=article&id=75

English Version:

https://china.ucsd.edu/_files/01112016_Taiwan-election.pdf;

https://chinafocus.ucsd.edu/2016/01/10/taiwan-election-results-predicted-in-computer-simulation/

(2) On Nov. 04, 2018, our team released our forecasting of US Senate Elections in two states (West Virginia and Missouri) before the actual voting of the two elections. That forecasting was our second attempt of election forecasting with our new approach. For details, please see:

Chinese Version:

http://blog.sina.com.cn/s/blog_744a73490102y7bd.html

http://www.ccda.fudan.edu.cn/index.php?c=article&id=80

English Version:

http://www.ccda.fudan.edu.cn/index.php?c=article&id=81

https://www.linkedin.com/pulse/united-states-senate-elections-west-virginia-missouri-tang-dr/?trackingId=shAzPfa2GP9KRu%2BbwvnIyQ%3D%3D

Results compared:

Chinese Version:

https://sirpa.fudan.edu.cn/info/1079/2741.html

Brief English version:

https://www.linkedin.com/pulse/prelimary-assessment-our-forecasting-excellent-perhaps-tang-dr/

(3) On Nov. 22, 2018, our team released the forecasting of Taiwan local elections (Taipei, New Taipei, and Taoyuan) before the elections were started that year. That prediction was our third attempt of election forecasting with our new approach. For details, please see:

Chinese Version:

http://www.ccda.fudan.edu.cn/index.php?c=article&id=83

English Version:

http://www.ccda.fudan.edu.cn/index.php?c=article&id=84

https://www.linkedin.com/pulse/taiwan-local-elections-taipei-new-taoyuan-forecasting-tang-dr/?trackingId=VmpIAUnefnghFHhzA9ODeg%3D%3D

Results compared:

Chinese Version:

http://www.ccda.fudan.edu.cn/index.php?c=article&id=85

(4) On January 9, 2020, our team released the forecasted results for the Taiwan general election on January 11, 2020. That prediction was our fourth attempt of election forecasting with our new approach. For details, please see:

English Version:

https://www.linkedin.com/pulse/2020-taiwan-elections-forecasted-computer-simulation-shiping-tang-dr/

For our initial assessment of this forecasting experiment, see:

English Version:

http://www.ccda.fudan.edu.cn/index.php?c=article&id=97

Chinese Version:

http://www.ccda.fudan.edu.cn/index.php?c=article&id=96

审核、编辑：大可

版权声明：本文授权自“复旦复杂分析决策中心”，如若转载请与原作者联系。

特别声明：以上内容(如有图片或视频亦包括在内)为自媒体平台“网易号”用户上传并发布，本平台仅提供信息存储服务。

Notice: The content above (including the pictures and videos if any) is uploaded and posted by a user of NetEase Hao, which is a social media platform and only provides information storage services.