A predictive framework that measures how voters evaluate candidates’ future leadership and policy expertise accurately predicted the outcome of a close 2024 US presidential election. This approach provides an alternative to traditional models that rely on past economic performance and provides election strategists with specific guidance for shaping public perceptions. The study was published in the journal Research and Politics.
Election prediction has a long history in political science and often relies on a concept known as retroactive voting. This theory posits that the public acts as an auditor for incumbent political parties, punishing or rewarding candidates based on recent economic indicators or job approval ratings. These models calculate a candidate’s likelihood by looking back at the outgoing administration’s performance.
However, retrospective prediction becomes complicated in open-seat elections in which the incumbent president does not participate in the vote. This is exactly the scenario in the 2024 election after President Joe Biden withdrew from the race. Without a direct track record for a new candidate, voters tend to shift their focus away from their judgments about the past.
Instead, many voters rely on future voting, a process that evaluates candidates based on their expected future performance. The researchers found that these forward-looking assessments were a key driver of voter behavior in open seat elections. Voters are paying attention to what policies and leadership styles the new candidates will bring to power.
To capture these forward-looking dynamics, Andreas Grefe, a researcher at the Macromedia University of Applied Sciences in Munich, Germany, developed a predictive tool known as the Problem and Leader Model. Graefe designed a framework to address common limitations in mainstream election forecasting. Popular modern models such as standard poll aggregators are often updated daily to provide a real-time snapshot of the race.
These aggregators help campaigns identify where to allocate financial resources, but provide little strategic advice on a campaign’s message. They tell observers who is currently winning in standard preference surveys, but they do not explain which policy areas or personality traits are driving voters’ preferences. Graefe sought to create a tool that tracks specific considerations underlying voters’ choices.
The problem-leader model focuses entirely on two variables: problem-handling ability and leadership perceptions. To calculate problem handling scores, the model requires three conditions to be met. Voters must recognize the issue, recognize it as important, and trust one candidate over the other to manage the issue.
To determine the relative importance of various issues, the model draws on responses from a regular Gallup poll that asks Americans to name the country’s most pressing issues. These issues are categorized as economic, foreign policy, and other domestic issues. The model then weighs how much importance voters assign to each category.
Once the relative weight of an issue is established, the model analyzes survey data about which candidates voters trust to handle the issue. For his 2024 analysis, Greif used polling data compiled by election analysis website FiveThirtyEight. This dataset contains 586 questions about the competencies in question, drawn from 87 independent surveys conducted between October 2023 and November 2024.
The model calculates daily averages of voters’ confidence in the incumbent party’s candidates on all these issues. It uses a mathematical technique called exponential smoothing, which gives more weight to recent polls while retaining some influence from older data. This prevents the model from overreacting to small short-term fluctuations in public opinion.
The second component of the model measures perceptions of leadership. The index is based on polls that ask respondents simple and direct questions about which candidate they believe is a stronger leader. The 2024 Tracking included 22 unique studies conducted between February and October that focused explicitly on this one dimension of leadership strength.
To turn these daily scores into actual election predictions, Greif relied on historical data spanning 13 U.S. presidential elections from 1972 to 2020. By analyzing past voting patterns, statistical models determine exactly how much weight to assign to partisanship, issue handling, and leadership scores at different stages of a campaign. This model assesses how these variables interact with the incumbent party’s final share in the two-party popular vote.
Historical analysis has revealed significant changes in voter behavior as election day approaches. After 100 days, much of the intent is simply tied to underlying partisan loyalty. By the final hours of a campaign, basic partisan influence drops by more than half. Instead, candidate-specific assessments of policy expertise and leadership carry considerable weight in predictions.
Applying the model to the 2024 election campaign, evolving voter perceptions were tracked in real time. Just 100 days before the election, Vice President Kamala Harris maintained a slight advantage over former President Donald Trump in terms of overall issue proficiency. In contrast, she started at a huge disadvantage in terms of leadership perceptions, losing 20 points to Trump in late July.
Over the next few months, Harris steadily closed the perception gap among leaders. By the eve of the election, she had reduced Mr. Trump’s lead in this area to less than 5 points. Despite this late momentum, Mr. Trump maintained a perceived leadership advantage that offset Ms. Harris’s slight advantage on policy issues.
The final forecast, generated by the model on the eve of the election, predicted that Mr. Trump would receive 50.2 percent of the two-party popular vote and Ms. Harris would receive 49.8 percent, a near tie. This cautious forecast contrasted with previous polling averages that generally showed Ms. Harris maintaining a narrow lead. In the end, Trump won the national popular vote by about 1.5 percentage points.
The model’s final prediction underestimated Trump’s final vote share by just 0.5 percentage points. Over the entire 100-day follow-up period, the model had an average error of only 0.65 percentage points. This level of accuracy is consistent with its performance in the 2012, 2016, and 2020 election cycles, indicating its reliability as an out-of-sample forecasting tool.
In addition to providing accurate predictions, Graefe says the model provides specific strategic direction for campaigns. Recognizing that voter perceptions determine elections, candidates can actively try to improve their reputations on important policies. You can also try to direct media coverage to topics that already enjoy a reputational advantage.
Because perceptions of leadership are heavily influenced by relatively fixed characteristics such as professional background and personal attitudes, political parties can use these indicators during primaries to make more informed choices. Identifying candidates who naturally exhibit the leadership qualities expected by a broader electorate can give parties an early structural advantage.
The main limitation of the current model is the restriction on national popular voting. This model is unable to generate Electoral College predictions due to the lack of detailed state-level polling on specific issue handling or leadership questions. In the US system, the popular vote does not determine the presidential winner, making state-by-state projections highly desirable.
Extending this methodology to individual states is a promising direction for future research. A state-level approach could capture regional differences in what voters value most. For example, immigration policy may be more important to voters in border states, while economic manufacturing issues may dominate in industrial areas.
Access to localized data allows researchers to adjust the sensitivity of their models to these geographic differences. Such enhancements provide campaigns with localized intelligence to tailor ads and stump speeches. Until such polls become widely available, national models will remain a useful tool for understanding the underlying expectations that shape voting behavior.
The study, “Predicted Voting and Issue and Leader Models: Predicting the 2024 U.S. Presidential Election,” was authored by Andreas Graefe.
