Asking Who Will Win: Citizen forecasting

In brief

Taking the most common answer to who will win, district by district, beats the average individual guess, and in some countries it beats the vote-intention polls as well. Two things shrink the gain. In close races the expectation question carries almost no signal, so the places that decide elections are the places it tells you least. Accuracy is also higher among people who backed the eventual winner, which means part of any headline figure reflects which party was simply bigger rather than what voters there knew.

How to use this

Take the majority answer inside each district as your forecast, not the average of individual guesses, and check that enough people in each district answered for that majority to mean anything. Before trusting a headline accuracy figure, ask whether the sample was tilted towards one side: pick a battleground district and see whether expectation still separates the candidates. If it does not, pair the forecast with something else. Do not assume a method validated abroad will hold where you are; commission it locally and publish the method. Resist hand-picking your most engaged supporters as forecasters, since the group-level evidence for that is unresolved.

What the story is about

Ask people who they think will win, rather than how they will vote, and you can build a forecast district by district. In 2021, Leininger et al. (2026) ran a survey for the German federal election. They asked respondents which candidate would win in their own constituency, what share each candidate would take there, and what share each party would take nationally. The survey went to people who had signed up to answer questions online, which the authors describe as "a non-representative sample from an online-access panel". Inside each constituency they took the most common answer and called that the winner. The majority answer was right in 224 of 299 constituencies. That same survey overestimated CDU/CSU seats, and it correctly predicted that the FDP and other small parties would win no constituencies outright.

One country would be just one result but the same pattern turned up elsewhere. In Canada, across six national and provincial contests from 2011 to 2022, about 60.4% of individual respondents named their district's winner, while taking the majority answer in each district was right 78.8% of the time, a gain of over 18 points (Mongrain et al., 2025). The German survey showed the same direction, with individual respondents naming their own constituency's winner correctly 46.8% of the time and the district majority answer gaining about 28 points on that (Leininger et al., 2026). Leininger et al. (2026) also summarise earlier records from the United States and Britain, though those come from their review of other researchers' work rather than a fresh test. None of this means the ordinary vote-intention question has stopped working. Ko et al. (2025) found that intended-vote measures predict the aggregate popular vote reasonably well, and that the post-election reported vote is off by 2.23 percentage points on average. Their caution is the one this article needs: "forecasting the popular vote may not always reveal the actual winner".

Whether the expectation question beats the vote-intention polls is a separate question, and the record is mixed. The strongest version of the claim came first. Graefe (2014) found that across the last 100 days of the seven US presidential elections from 1988 to 2012, expectation-based vote-share error averaged 51% lower than that of polls published the same day, and that a majority of respondents named the winner in 193 of 217 surveys from 1932 to 2012. Later studies narrow that. Murr (2025), looking at 55 surveys across Mexican presidential elections from 2000 to 2024, finds that citizens predicted worse than chance in 2000 and now call recent elections correctly, on a par with vote-intention polls on the winner, while still worse on vote percentage. Leininger et al. (2026) read the overall record as favouring citizen forecasts at least in mature democracies, a narrower claim than Graefe's. What the crowd knows beyond the polls a respondent could already have seen remains unmeasured.

Supporters of a party tend to expect their own party to win, a pattern documented throughout this research. Scholars agree on the pattern and disagree about what it means. Barnfield (2025) argues that the usual reading, in which these voters are simply irrational, is wrong. He calls the pattern electoral hope, and describes it as "rational in a practical, if not necessarily theoretical, sense". Curl and Rocheleau (2025), on a nationally representative 2024 US sample, report a strong tendency for people to take their own preference as the majority view: "participants overwhelmingly believed that the majority would vote for their preferred candidate".

The aggregate advantage has two soft spots. The first is close races. Huber and Tucker (2023) used 2020 US data and found that across states, expectations tracked the actual results quite closely. Restrict the analysis to battleground states, and the signal largely disappears: on the 0-to-1 scale where 1 means expecting Biden to win the respondent's state, the average expectation in close Biden-won states and close Trump-won states differed by 0.035, so the two sets of close states came out indistinguishable. The places where elections are decided are the places where the expectation question tells you least. The second soft spot is who is doing the expecting. Mongrain et al. (2025) found that having voted for the winning candidate raised the chance of a correct district forecast by about 20 points at the low end (2022 Quebec) and 51 points at the high end (the 2015 Canadian federal election). Thompson-Collart and Mongrain (2025), working on nearly 60,000 respondents across 22 presidential elections in five Central American countries from 1999 to 2022, found that "supporters of the eventual winner are significantly more likely to forecast the correct outcome". Bigger margins of victory also go with more correct forecasts: a noticeably larger margin of victory raised the odds by 7 to 15 points (Mongrain et al., 2025).

Taking the majority answer across many respondents in the same district beats the average individual guess, and in some countries it beats the vote-intention polls too. This forecast is the pooled majority answer inside a district. Two things shrink that gain. In close races the expectation question carries almost no signal, and accuracy is higher among supporters of the side that ends up winning. Part of any headline accuracy figure, then, reflects which party was simply bigger in that district rather than what the voters there knew.

So what

If you commission one of these surveys, treat the district-level majority answer as your number, and treat the topline of individual guesses as a weaker companion to it. That number only means something when enough people in the same district answer. Check whether it moved because respondents knew something, or because the sample leaned towards one side. The method is weakest exactly where elections are decided, so in a close race pair it with something else.

For political parties

The tempting shortcut is to survey only your most politically engaged supporters and trust their judgement. That is not a proven fix. Thompson-Collart et al. (2024) recommend delegating the forecast to better-informed respondents, citing earlier work that this raises competence. Mongrain et al. (2025), on 279,003 respondents across six Canadian contests, found the opposite at the group level. Education and political interest do raise a person's chance of being right, but delegating on those variables "does not necessarily lead to improvements in the accuracy of aggregate-level predictions", and the authors found no evidence that sociological or informational diversity helped group-level accuracy. Two peer-reviewed papers, both published since 2022, disagree. Until that is settled, hand-picking respondents to flatter your own side buys you a mirror rather than a forecast.

For government

The method's record is geographically uneven, so a government should not assume it has been validated at home. The UK has a long history of citizen forecasting (Murr, 2011), but no citizen forecast of the 2024 UK general election has been published. Given how much results vary between countries and how much they weaken in close races, a finding from another system is a hypothesis to test rather than a tool to rely on. Commissioning one in your own country, and publishing the method alongside the result, is how that gap gets closed. The alternative is a forecast that looks authoritative and has never been checked where you are governing.

References

Barnfield, M. (2025) 'Electoral hope', Political Studies, 74(3), pp. 1163–1180. Available at: https://doi.org/10.1177/00323217251359362 (Accessed: 9 September 2026).

Curl, L.S. and Rocheleau, C.A. (2025) 'Polarized perceptions: how anchoring shaped voter expectations to the 2024 presidential election', Discover Psychology, 5(1), 43. Available at: https://doi.org/10.1007/s44202-025-00365-0 (Accessed: 9 September 2026).

Graefe, A. (2014) 'Accuracy of vote expectation surveys in forecasting elections', Public Opinion Quarterly, 78(S1), pp. 204–232. Available at: https://doi.org/10.1093/poq/nfu008 (Accessed: 9 September 2026).

Huber, G.A. and Tucker, P.D. (2023) 'What to expect when you're electing: citizen forecasts in the 2020 election', Political Science Research and Methods, 12(3), pp. 624–632. Available at: https://doi.org/10.1017/psrm.2022.61 (Accessed: 9 September 2026).

Ko, H., Jackson, N., Osborn, T. and Lewis-Beck, M.S. (2025) 'Forecasting presidential elections: accuracy of ANES voter intentions', International Journal of Forecasting, 41(1), pp. 66–75. Available at: https://doi.org/10.1016/j.ijforecast.2024.03.003 (Accessed: 9 September 2026).

Leininger, A., Murr, A.E., Stötzer, L. and Kayser, M.A. (2026) 'Citizen forecasting in a mixed electoral system', International Journal of Forecasting, 42(1), pp. 203–215. Available at: https://doi.org/10.1016/j.ijforecast.2025.03.007 (Accessed: 9 September 2026).

Mongrain, P., Fréchet, N., Thompson Collart, B. and Dufresne, Y. (2025) 'Working the crowd: citizen forecasting, sophistication and diversity in Canadian federal and provincial elections', Canadian Journal of Political Science, 58(1), pp. 68–94. Available at: https://doi.org/10.1017/s0008423924000465 (Accessed: 9 September 2026).

Murr, A.E. (2011) 'Wisdom of crowds? A decentralised election forecasting model that uses citizens' local expectations', Electoral Studies, 30(4), pp. 771–783. Available at: https://doi.org/10.1016/j.electstud.2011.07.005 (Accessed: 9 September 2026).

Murr, A.E. (2025) 'Predicciones ciudadanas de las elecciones presidenciales mexicanas, 2000-2024' [Citizen forecasts of Mexican presidential elections, 2000-2024], Política y Gobierno, 32(1). Available at: http://www.politicaygobierno.cide.edu/index.php/pyg/article/view/1751 (Accessed: 9 September 2026).

Thompson-Collart, B. and Mongrain, P. (2025) 'Citizens' electoral expectations in imperfect democracies: insights from five Central American countries', Revista Latinoamericana de Opinión Pública, 14, e32143. Available at: https://doi.org/10.14201/rlop.32143 (Accessed: 9 September 2026).

Thompson-Collart, B., Cadieux, H., Ouellet, C. and Dufresne, Y. (2024) 'Lessons learned: citizen forecasting, candidate resignations, and the 2024 US presidential election', PS: Political Science & Politics, 58(2), pp. 312–317. Available at: https://doi.org/10.1017/s1049096524000969 (Accessed: 9 September 2026).