Political Science Election Forecasting
Mary Stegmaier, Helmut Norpoth
  • LAST REVIEWED: 15 March 2017
  • LAST MODIFIED: 27 June 2017
  • DOI: 10.1093/obo/9780199756223-0023


Election forecasting appeals to a basic human urge to peek into the future. Ever since elections were invented to choose leaders, humans have been tempted to find ways that would tell them with some degree of certainty who would win an election. The highly quantitative nature of elections aids them in such an endeavor. Few phenomena of interest to a social scientist lend themselves so readily to forecasting than electoral contests. In a two-party competition they produce a clear winner, and in multiparty settings they produce numerical shares for each of the contenders. With the advent of statistical techniques, electoral data have become increasingly easy to handle. It is no surprise, then, that election forecasting has become a big business, for polling firms, news organizations, and betting markets as well as academic students of politics. There are three major types of election forecasting. Perhaps the best known to the general public involves polls of the voting public. The growth of polling is supplying a near-endless stream of data on candidate and party support that can be marshaled for predictions of electoral outcomes. During a national election campaign in the United States and many other countries, a new poll is reported every day showing the current state of the “horse race.” Though not strictly a forecast, the poll result, or an average of the latest polls, is widely seen as the best guess of who is going to win and by how much. On election night, as the nation eagerly awaits the result, news organizations rely on exit polls to project the winner of a particular electoral contest. Alongside predictions of outcomes from samples of voters (or polls taken beforehand), academic scholars have constructed models of voting behavior to forecast the outcomes of elections. These forecasts are derived from theories and empirical evidence about what matters to voters when they make electoral choices. The forecast models typically rely on a few predictors in highly aggregated form, with an emphasis on phenomena that change in the short-run, such as the state of the economy, so as to offer maximum leverage for predicting the result of a specific election. Finally, betting markets provide forecasts of election outcomes based on the buying and selling of candidate futures with real money. These markets have witnessed a resurgence with the advent of the Internet, but their operations face legal obstacles in the United States.

General Overviews

The author of Bean 1948 may have been the first to publish a book using the phrase, “how to predict elections” in the title. Bean is best remembered for an approach that searches for the locality (state, county) whose vote matches most closely the national vote division. Such “bellwethers,” he believed, would provide a highly efficient way of forecasting the overall outcome, with little lead time, if all people vote on the same day. With an article published in the 1970s, Fair founded the econometric school of forecasting presidential elections, using mainly aggregate economic measures along with political variables and extending the time horizon back to the 1916 election. This approach is detailed in Fair 2002. In addition, Fair has kept his model up to date and lets users calculate their own forecasts on his website. Lewis-Beck and Rice 1992 is the first book-length overview of the various approaches to election forecasting. It establishes clear statistical criteria for building forecast models with proven determinants of the vote choice, such as economic conditions and presidential approval. The book is also valuable for applying the forecast model to congressional elections, as well as elections outside the United States. Jones 2002 extends the scope of forecasting to trial heats (polls), exit polls, expert judgments, cycles, and the nomination process. Norpoth 2014 demonstrates the predictive utility of electoral cycles with an analysis of nearly 200 years of US presidential elections. Holbrook 2010 covers forecasts through betting markets and state-level models. The template for using all the states in presidential election forecasting was first presented in Rosenstone 1983. Such a model would be able to forecast not only the national popular vote, but also the Electoral College vote, which ultimately elects a president. The introduction to an issue on a symposium, Lewis-Beck and Stegmaier 2014 presents a recent assessment of the various approaches to election forecasting.

  • Bean, Louis H. How to Predict Elections. New York: Knopf, 1948.

    Save Citation »Export Citation »E-mail Citation »

    A pioneering work in election forecasting, using past election results to identify states that are most typical for the national outcome of presidential elections (“bellwethers”). Maine’s early vote at that time put that state in a special role (“As Maine goes, so goes the nation”).

    Find this resource:

  • Fair, Ray C. Predicting Elections and Other Things. Stanford, CA: Stanford University Press, 2002.

    Save Citation »Export Citation »E-mail Citation »

    A pioneer in election forecasting presents his model, which heavily relies on measures of economic performance and covers elections from 1916 on. Also discusses the theoretical and methodological foundations of forecasting elections as well as other things. Fair’s website lets users calculate their own forecasts of upcoming elections.

    Find this resource:

  • Holbrook, Thomas. “Forecasting US Presidential Elections.” In The Oxford Handbook of American Elections and Political Behavior. Edited by Jan Leighley, 346–371. Oxford: Oxford University Press, 2010.

    DOI: 10.1093/oxfordhb/9780199235476.001.0001Save Citation »Export Citation »E-mail Citation »

    An assessment of the state of election forecasting as of 2010. Examines and compares the accuracy of presidential election forecasts from 1996 to 2004, with special attention to the lessons learned from the 2000 election; also considers alternatives such as political markets and state-level forecasting.

    Find this resource:

  • Jones, Randall J. Who Will Be in the White House? Predicting Presidential Elections. New York: Longman, 2002.

    Save Citation »Export Citation »E-mail Citation »

    A comprehensive and highly readable overview of the various approaches to election forecasting, from bellwethers to trial heats, presidential approval, and economic models, to name a few, with their respective forecast scenarios for the 2004 presidential election.

    Find this resource:

  • Lewis-Beck, Michael S., and Tom Rice. Forecasting Elections. Washington, DC: CQ, 1992.

    Save Citation »Export Citation »E-mail Citation »

    A classic book-length work of election forecasting. It presents an overview of approaches to election forecasting, focusing on the theory and statistical procedures of forecasting while covering models of presidential elections, but also including House, Senate, and state contests as well as elections abroad (France).

    Find this resource:

  • Lewis-Beck, Michael S., and Mary Stegmaier. “US Presidential Election Forecasting: Introduction.” PS: Political Science and Politics 47.2 (April 2014): 284–288.

    Save Citation »Export Citation »E-mail Citation »

    An introduction for a symposium in this issue on forecasting presidential elections, with coverage of approaches using fundamentals, polls, prediction markets, and cycles.

    Find this resource:

  • Norpoth, Helmut. “The Electoral Cycle.” PS: Political Science and Politics 47.2 (April 2014): 332–335.

    Save Citation »Export Citation »E-mail Citation »

    An autoregressive model for presidential elections from 1828 to 2012 shows the predictive power of the two previous outcomes, allowing for an early forecast. The White House party is favored after one term while change is more likely after two terms.

    Find this resource:

  • Rosenstone, Steven J. Forecasting Presidential Elections. New Haven, CT: Yale University Press, 1983.

    Save Citation »Export Citation »E-mail Citation »

    This book presented the first state-level model to predict the vote in presidential elections, using a pooled cross-sectional design, and covering elections from 1948 to 1972. Key predictors are issues, state of the economy, and incumbency. The model was used to forecast both the 1976 and 1980 presidential elections.

    Find this resource:


Articles on election forecasting can be found in a variety of academic journals. PS: Political Science & Politics, a publication of the American Political Science Association, is recognized for publishing the leading model-based forecasts for US presidential and midterm elections, usually in the issue before the election takes place. The International Journal of Forecasting publishes empirically based articles on forecasting across disciplines, including a strong presence of political forecasts. Recently, the journal has also devoted space to special sections on election forecasting in Europe and in countries where predictive modeling is a new venture. Other leading academic publications where scientific election forecasts can be found include Electoral Studies, Public Opinion Quarterly, and the Journal of Elections, Public Opinion and Parties. In addition to academic journals, the PollyVote website, maintained by a group of academic forecasters, offers up-to-date information on US election predictions. The site tracks predictions based on polls, models, Iowa Electronic Markets, index forecasts, and expert judgments. Forecasters post their updated predictions on this site and discuss innovations in election forecasting.

Polling Forecasts

These days, thanks to the proliferation of polling and the explosion of the Internet, anyone looking for a forecast of a major American election has to turn no further than the daily posting of a variety of “horse-race” polls on readily available websites. The best known site is Nate Silver’s FiveThirtyEight, which translates the weighted poll averages into the likelihood of a candidate winning an election. It includes Internet polls along with telephone polls. Its vote projection is based on a statistical algorithm to weight recent polls as a means to provide a forecast. This site also features a lively blog where polling issues are aired. HuffPost Pollster provides a similar service. Long before the advent of scientific polling, “straw polls” like the one conducted by the Literary Digest in the early 20th century provided forecasts of elections. Squire 1988 illuminates the biases of such polls with an analysis of the failure the Literary Digest poll in the 1936 election. The failure of more scientific polls, like the Gallup Poll in the 1948 election, led to an inquiry by social scientists, Mosteller, et al. 1949, that pointed out several problems of polls with a carefully drawn sample of the electorate in predicting electoral outcomes. Foremost among them were the fact that some voters changed their mind late in the campaign after the polls were conducted and the allocation of undecided respondents. Erikson, et al. 2004 calls attention to a further problem; namely, the difficulty to predict which respondents will actually turn out to vote. Given that barely half of eligible voters cast votes in a US presidential election, this is not a trivial problem. Erikson and Wlezien 2012 tracks all the trial-heat polls throughout presidential campaigns from 1952 to 2008 in search of the best moment to predict the outcome. On election night, exit polls provide the major tool for news organizations to project winners and losers. The method looks simple. An exit poll takes a sample of actual voters and then asks them how they just voted. Yet wrong calls in recent elections have called attention to shortcomings of exit polls as forecasting tools. Mason, et al. 2001 points to the problem with absentee voting, and Traugott, et al. 2005 notes response bias, in both cases favoring the Democratic side.

  • Erikson, Robert S., Costas Panagopoulos, and Christopher Wlezien. “Likely (and Unlikely) Voters and the Assessment of Campaign Dynamics.” Public Opinion Quarterly 68.4 (2004): 588–601.

    DOI: 10.1093/poq/nfh041Save Citation »Export Citation »E-mail Citation »

    A study of tracking polls of candidate preferences finds that likely voter preferences change far more than those of registered voters. The authors suggest that the reported variation in candidate preference among likely voters is not so much the result of shifts in preference as of changes in the likely voter pool.

    Find this resource:

  • Erikson, Robert S., and Christopher Wlezien. The Timeline of Presidential Elections. Chicago: University of Chicago Press, 2012.

    DOI: 10.7208/chicago/9780226922164.001.0001Save Citation »Export Citation »E-mail Citation »

    An analysis of trial-heat polls in presidential elections (1952–2008) shows that these polls get ever more accurate during the election year but that the post-convention campaign adds little value in predictive accuracy.

    Find this resource:

  • FiveThirtyEight.

    Save Citation »Export Citation »E-mail Citation »

    A popular site, founded and operated by Nate Silver, offers poll-based forecasts along with probabilities of winning in the presidential race as well as state contests. Links to each poll allow for an easy way to find out about the sampling procedures and question wording. FiveThirtyEight also rates the quality of each poll.

    Find this resource:

  • HuffPost Pollster.

    Save Citation »Export Citation »E-mail Citation »

    A site founded by Mark Blumenthal that presents results of a wide variety of pre-election polls, forecasts based on a weighted average of recent polls, and a blog with contributions from polling experts. Both FiveThirtyEight and this site predicted that Hillary Clinton had a better than 70 percent chance of winning in 2016, both popular and electoral votes.

    Find this resource:

  • Mason, Linda, Kathleen Frankovic, and Kathleen Hall Jamieson. CBS News Coverage of Election Night 2000: Investigation, Analysis, Recommendations. New York: CBS News, 2001.

    Save Citation »Export Citation »E-mail Citation »

    An investigation of what went wrong with the exit poll projection of a Gore victory in Florida in the 2000 election. Of the four factors examined, a bad estimate of the absentee vote was seen as the most critical one, compared with sampling error, past-race comparisons, and timing of the projection.

    Find this resource:

  • Mosteller, Frederick, Herbert Hyman, Philip J. McCarthy, Eli S. Marks, and David B. Truman. The Pre-election Polls of 1948: Report to the Committee on Analysis of Pre-election Polls and Forecasts. New York: Social Science Research Council, 1949.

    Save Citation »Export Citation »E-mail Citation »

    This report by statisticians and social scientists pointed out key problems with pre-election polls in 1948, all of which showed Dewey defeating Truman. Problems included ending interviewing too early, pro-Republican sample bias, allocation of undecided voters, and difficulty distinguishing voters and nonvoters. Text available online.

    Find this resource:

  • Squire, Peverill. “Why the 1936 Literary Digest Poll Failed.” Public Opinion Quarterly 52.1 (1988): 125–133.

    DOI: 10.1086/269085Save Citation »Export Citation »E-mail Citation »

    Using a 1937 Gallup poll that probed participation in the 1936 Literary Digest poll, the analysis shows that both the Digest sample and the response had serious biases that led to the erroneous forecast of a defeat for Franklin D. Roosevelt in a landslide. The response bias, however, was the more serious problem.

    Find this resource:

  • Traugott, Michael, Benjamin Highton, and Henry E. Brady. A Review of Recent Controversies Concerning the 2004 Presidential Election Exit Polls. New York: Social Science Research Council, 2005.

    Save Citation »Export Citation »E-mail Citation »

    Examines the Democratic bias of the 2004 presidential exit polls, with leaks of a Kerry lead on Election Day leading to rumors that he had won. The factors contributing to the Democratic bias include weighting schemes and lower inclination of Bush voters to respond to exit polls.

    Find this resource:

Model Forecasts

Trained in electoral analysis and statistical methods, political scientists have increasingly constructed vote models to forecast the outcome of an election. Fair 2002 (cited under General Overviews) blazed the trail for such model forecasts. The next four sections cite forecasting works applied to (1) the popular vote in presidential elections; (2) the division in the Electoral College and the US Congress; (3) major democracies outside the United States where such forecasting has taken hold, such as Britain and France; and (4) other democracies that have witnessed significant model forecasting.

Presidential Elections

In recent election years the journal PS: Political Science & Politics (cited under Journals) has served as a venue for publishing US presidential election forecasts ahead of the Election Day, along with some explanation of each model and its predictors. The October 2016 (Vol. 496, no. 4) issue is the latest in a series of pre-election forecast postings. Most of the forecasters have used the same or a modified version of their 2016 model in previous elections; references to previous applications can be found in the 2016 articles. All of the models except one (Norpoth 2016) employ economic performance, in one measure or another; some also rely on approval ratings of the incumbent president, even when he is not on the ballot (Abramowitz 2016, Lewis-Beck and Tien 2016, Holbrook 2016) while others add trial heats (Campbell 2016). Only one model (Norpoth 2016) uses primary elections to gauge the strength of the major-party nominees. Several models take into account the diminishing value of incumbency for the party holding the White House, either by including the length of time the presidential party has been in office (Abramowitz 2016, Lockerbie 2016, Holbrook 2016) by discounting other variables (Campbell 2016), or through an autoregressive cycle (Norpoth 2016). While most of the models are estimated with data covering elections since 1952, one reaches further back to 1912 (Norpoth 2016). The lead time of the models’ forecasts varies from a minimum of sixty days before Election Day of 2012 (Campbell 2016) to 246 days (Norpoth 2016). All except Abramowitz 2016 and Norpoth 2016 picked Hillary Clinton as the winner.

Electoral College and Congressional Elections

For some types of elections, forecasting the national vote division may not be enough to tell who will win. In the end, it is the vote of the Electoral College, not the popular vote, that decides US presidential elections. Likewise, in legislative elections, how many seats a party controls is more consequential than what share of the vote it obtained in an election. Following a trail blazed by the state-level model of Rosenstone 1983 (cited under General Overviews), forecasters of presidential elections have made strides in predicting the popular vote of the various states in order to pick the Electoral College winner. This, of course, requires collecting data for the predictors at the state level. Some may be easy to come by, such as a state’s presidential vote in past elections or a state’s level of unemployment, while others may not even be available for many states, such as a president’s approval rating in a given state. Klarner 2012 uses a number of economic variables and the vote in past elections at the state level along with presidential approval and trial heats at the national level to predict the electoral vote. Jerôme and Jerôme-Speziari 2016 conducts a pooled time-series analysis with economic and political variables. Forecast models of seat losses or gains of the White House party in midterm congressional elections build on the template of presidential vote models. Economic predictors, such as GDP growth, and political predictors, such as approval of the incumbent president, are popular, one or both being used in Abramowitz 2010, Campbell 2010, Cuzán 2010, and Lewis-Beck and Tien 2010. Beyond those variables, the “generic ballot,” which refers to vote preferences in polls, akin to trial heats in presidential contests, enters the predictor set of Abramowitz 2010 and Bafumi, et al. 2010. Also considered in some models is the impact of seats held. Campbell 2010 pays special attention to toss-up seats, which hold the balance for the ultimate outcome. For the particular election covered by the articles listed here from 2010, three forecasts favored (correctly, as it turned out) the Republicans to control the House, and two favored the Democrats.

  • Abramowitz, Alan I. “How Large a Wave? Using the Generic Ballot to Forecast the 2010 Midterm Elections.” PS: Political Science & Politics 43.4 (October 2010): 631–632.

    DOI: 10.1017/S1049096510001058Save Citation »Export Citation »E-mail Citation »

    The Abramowitz model relies on the “generic ballot” (the partisan preferences for congressional candidates in polls), the number of seats previously held, the approval in opinion polls of the incumbent president, and a midterm election dummy variable.

    Find this resource:

  • Bafumi, Joseph, Robert S. Erikson, and Christopher Wlezien. “Forecasting House Seats from Generic Congressional Polls: The 2010 Midterm Election.” PS: Political Science & Politics 43.4 (October 2010): 633–636.

    DOI: 10.1017/S104909651000106XSave Citation »Export Citation »E-mail Citation »

    This model offers two sets of forecasts. The first predicts the national House vote using the generic ballot (preference poll) and the party of the president; the second relies on the predicted national vote, along with an adjustment for open-seat contests, to forecast the winners of the 435 House races.

    Find this resource:

  • Campbell, James E. “The Seats in Trouble Forecast of the 2010 Elections in the U.S. House.” PS: Political Science & Politics 43.4 (October 2010): 627–630.

    DOI: 10.1017/S1049096510001095Save Citation »Export Citation »E-mail Citation »

    The Campbell model relies on an indicator of “seats in trouble” based on the Cook Political Report, seats previously held, and the approval in opinion polls of the incumbent president.

    Find this resource:

  • Cuzán, Alfred G. “Will the Republicans Retake the House in 2010?” PS: Political Science & Politics 43.4 (October 2010): 639–641.

    Save Citation »Export Citation »E-mail Citation »

    The Cuzán model relies on seats previously won, presidential incumbency, a midterm dummy variable, economic growth and inflation, along with an adjustment for the 1932 and 1948 elections.

    Find this resource:

  • Jerôme, Bruno, and Véronique Jerôme-Speziari. “State-Level Forecasts for the 2016 US Presidential Elections: Political Economy Model Predicts Hillary Clinton Victory.” PS: Political Science and Politics 49.4 (October 2016): 680–686.

    DOI: 10.1017/S1049096512000972Save Citation »Export Citation »E-mail Citation »

    This forecast model uses a pooled time-series design with national variables (presidential approval and unemployment) as well as state-level data for a variety of economic and political variables. Forecasts are made for every state, predicting a resounding Electoral College victory for Clinton.

    Find this resource:

  • Klarner, Carl. “State-Level Forecasts of the 2012 Federal and Gubernatorial Elections.” PS: Political Science and Politics 45.4 (October 2012): 655–662.

    DOI: 10.1017/S1049096512000960Save Citation »Export Citation »E-mail Citation »

    The major predictors are national as well as state-level economic variables, lagged vote, presidential approval, and two-term penalty. Klarner predicts both the popular and the Electoral College vote for president as well as House and Senate elections.

    Find this resource:

  • Lewis-Beck, Michael S., and Charles Tien. “The Referendum Model: A 2010 Congressional Forecast.” PS: Political Science & Politics 43.4 (October 2010): 637–638.

    DOI: 10.1017/S1049096510001071Save Citation »Export Citation »E-mail Citation »

    The “Referendum Model” relies on change in real disposable income, the approval in opinion polls of the incumbent president, and a midterm election dummy variable.

    Find this resource:

British and French Elections

Shortly after the emergence of the initial US election forecasting models, political economists began constructing predictive models for British and French elections. Voting behavior and popularity studies from these countries, reviewed in Lewis-Beck and Stegmaier 2000, provided the theoretical foundations for these studies. UK parliamentary elections present a challenge to forecasters because it is not the party vote share that ultimately matters for determining the prime minister and the government, but rather the parliamentary seat share. Additionally, the rise of the Scottish National Party and the UK Independence Party (UKIP) required forecasters to consider the impact of “third parties” in their forecasts. As Fisher and Lewis-Beck 2016 discusses in the introductory article to the special issue of Electoral Studies, guest edited by the authors, because public opinion polls were so inaccurate in 2015, models that relied on public opinion data greatly underestimated the Conservative seat share. None of the models in this special issue predicted an outright Conservative majority in parliament. This is in contrast to the success of forecasts in 2010, which highlighted the predictive power of approval ratings. For example, Lebo and Norpoth 2011 uses approval of the prime minister along with the cyclical nature of elections to predict the vote, which is used to forecast seats. Lewis-Beck, et al. 2011 uses the average of prime minister and government approval, a notion that is supported by a separate explanatory model. In the 2015 election, the Party Leadership Model, developed in Murr 2015, eschews public opinion data and accurately predicted that Conservative prime minister David Cameron would be reelected. In France, election forecasting is complicated by its multiparty system and the two rounds of balloting necessary for the victor to garner 50 percent of the vote in presidential and National Assembly elections. Nadeau, et al. 2010 simplifies both of these issues by forecasting the first-round vote share for the left-wing presidential candidates, using this prediction as a signal of left-wing candidate support on the second ballot. Another challenge for modelers is the small number of direct presidential elections that have been held during the French Fifth Republic. Foucault and Nadeau 2012 and Foucault 2012 have worked around this small sample size by using local-level data to predict the 2012 presidential and National Assembly election results.

  • Fisher, Stephen D., and Michael S. Lewis-Beck. “Forecasting the 2015 British General Election: The 1992 Debacle All Over Again?” Electoral Studies 41.1 (2016): 225–229.

    DOI: 10.1016/j.electstud.2015.11.016Save Citation »Export Citation »E-mail Citation »

    Introduction to the special symposium on forecasting the 2015 UK parliamentary election. They summarize the twelve models presented in the issue, assess the accuracy and lead time of these forecasts, and compare them to public opinion polls. Problems of polling accuracy and the implications for forecasts are also discussed.

    Find this resource:

  • Foucault, Martial. “Forecasting the 2012 French Legislative Election.” French Politics 10 (2012): 68–83.

    DOI: 10.1057/fp.2012.2Save Citation »Export Citation »E-mail Citation »

    Foucault develops a political-economy model to predict the 2012 French legislative elections. Using local and national-level data covering 1986–2007, the model predicts the vote share of the incumbent parties on the Left or on the Right, depending on which side is in power.

    Find this resource:

  • Foucault, Martial, and Richard Nadeau. “Forecasting the 2012 French Presidential Election.” PS: Political Science & Politics 45.2 (2012): 218–222.

    DOI: 10.1017/S1049096512000066Save Citation »Export Citation »E-mail Citation »

    The authors use unemployment and previous election results measured at the level of the département along with national-level presidential approval to forecast the second-round vote share of the right-wing presidential candidate.

    Find this resource:

  • Lebo, Matthew, and Helmut Norpoth. “Yes, Prime Minister: The Key to Forecasting British Elections.” Electoral Studies 30.2 (2011): 258–263.

    DOI: 10.1016/j.electstud.2010.09.004Save Citation »Export Citation »E-mail Citation »

    First, the vote is predicted by the cyclical pattern of elections and prime minister approval, which is shown to perform better than government approval. Then the authors use the vote, with the help of an autoregressive model, to predict the number of seats for the two major parties.

    Find this resource:

  • Lewis-Beck, Michael S., Richard Nadeau, and Éric Bélanger. “Nowcasting v. Polling: The 2010 UK Election Trials.” Electoral Studies 30.2 (2011): 284–287.

    DOI: 10.1016/j.electstud.2010.09.013Save Citation »Export Citation »E-mail Citation »

    Forecast of incumbent party vote share as a function of incumbent support, measured as the average of prime minister and government approval, three months before the election. The authors produce a series of “nowcasts,” which forecast the election three months from the current time.

    Find this resource:

  • Lewis-Beck, Michael S., and Mary Stegmaier. “Economic Determinants of Electoral Outcomes.” Annual Review of Political Science 3 (2000): 183–219.

    DOI: 10.1146/annurev.polisci.3.1.183Save Citation »Export Citation »E-mail Citation »

    This article reviews the economic voting literature on national elections in the United States, France, Britain, and Denmark. The authors provide examples of vote and popularity functions in these countries, highlighting the different types of political and economic measures used, and confirming the explanatory power of these factors.

    Find this resource:

  • Murr, Andreas E. “The Party Leadership Model: An Early Forecast of the 2015 British General Election.” Research & Politics 2.2 (2015): 1–9.

    DOI: 10.1177/2053168015583346Save Citation »Export Citation »E-mail Citation »

    The Party Leadership Model uses the margin by which the party leaders won the leadership elections to predict the next prime minister. Murr’s model, predicting the reelection of Conservative prime minister David Cameron in 2015, is especially noteworthy for its correct forecast in an election when nearly all other predictions were wrong.

    Find this resource:

  • Nadeau, Richard, Michael S. Lewis-Beck, and Éric Bélanger. “Electoral Forecasting in France: A Multi-equation Solution.” International Journal of Forecasting 26.1 (2010): 11–18.

    DOI: 10.1016/j.ijforecast.2009.04.002Save Citation »Export Citation »E-mail Citation »

    The authors apply their two-step modeling approach, used in the UK context, to predict and explain the first-round vote share for the left-wing candidates in French presidential elections. The prediction model is a function of the executive’s popularity six months prior to the election, covering presidential elections 1965–2007.

    Find this resource:

Elections in Other Democracies

The success of forecasting models in the United States, France, and the United Kingdom has spurred interest in applying and adapting these models to other democracies. This enthusiasm can be seen in the publication of forecasting symposia in academic journals, including one on forecasting in Europe (Lewis-Beck and Bélanger 2012) and another on forecasting in neglected democracies (Lewis-Beck and Jérôme 2010). The fundamentals undergirding the reward-punishment political-economy forecasting models in the United States, Britain, and France can be observed in these models, with government or prime minister approval featuring prominently. Economic conditions appear directly in some models, while in others they are captured within the popularity or approval measure. The models developed in Bélanger and Godbout 2010, predicting Canadian elections, and Norpoth and Gschwend 2003, forecasting German elections, are exemplars of this latter approach. In other cases, forecasters have had to adapt their models to country contexts to improve predictive accuracy. In the Japanese case, the authors of Lewis-Beck and Tien 2012 tested a variety of measures to account for the short tenures of prime ministers and election timing. In the end, they settled on the days since the last election. Fewer days means that the prime minister is acting strategically by calling a new election at a time when the party has the greatest chance of winning. In the Turkish example, the author of Toros 2011 modifies the forecasting model to include a unique variable that captures whether the center (establishment) sanctions or attempts to exclude a party of the periphery from politics. One important statistical problem faced by forecasters in some countries is the small number of elections on which to base a forecast. When the sample size is too small, the predictive accuracy of the forecast is weakened. Different approaches have been use to get around the small N problem. In the case of Hungary, the newness of democracy and few elections led the authors of Stegmaier and Lewis-Beck 2009 to predict vote intention for the Hungarian Socialist Party using quarterly time-series data. Another approach to increase the sample size is to expand the model internationally. The authors of Kennedy, et al. 2017 gather data on 621 national executive elections globally spanning 1945–2012 to predict the fortunes of the incumbent party.

  • Bélanger, Éric, and Jean-François Godbout. “Forecasting Canadian Federal Elections.” PS: Political Science and Politics 43.4 (2010): 691–699.

    DOI: 10.1017/S1049096510001113Save Citation »Export Citation »E-mail Citation »

    Canadian parliamentary election vote share model, based on incumbent party approval, government longevity, and the unemployment rate, measured in election years 1953–2008. Incumbent seat share is projected from the predicted vote share, using a “swing ratio” model.

    Find this resource:

  • Kennedy, Ryan, Stefan Wojcik, and David Lazer. “Improving Election Prediction Internationally.” Science 355.6324 (2017): 515–552.

    DOI: 10.1126/science.aal2887Save Citation »Export Citation »E-mail Citation »

    In models predicting directly elected executive elections worldwide, the authors find that level of democracy, incumbent running for reelection, international aid, and public opinion polls enhance the accuracy of the forecast. Economic conditions have little impact, possibly because their effects are conditioned by political institutions.

    Find this resource:

  • Lewis-Beck, Michael S., and Éric Bélanger, eds. Special Issue: Election Forecasting in Neglected Democracies. International Journal of Forecasting 28.4 (2012).

    Save Citation »Export Citation »E-mail Citation »

    The editors of this special issue present a compilation of election forecasting models by leading scholars for countries where these efforts are still very new: Spain, Belgium, Norway, Japan, Brazil, Turkey, and Lithuania.

    Find this resource:

  • Lewis-Beck, Michael S., and Bruno Jêrôme, eds. Special Issue: European Election Forecasting. International Journal of Forecasting 26.1 (2010).

    Save Citation »Export Citation »E-mail Citation »

    This special issue includes articles on the forecasting of French, German, and Italian national elections, forecasting support for the radical right, predicting the influence of ideological groups in Europe, and improving the accuracy of polls.

    Find this resource:

  • Lewis-Beck, Michael S., and Charles Tien. “Japanese Election Forecasting: Classic Tests of a Hard Case.” International Journal of Forecasting 28.4 (2012): 797–803.

    DOI: 10.1016/j.ijforecast.2012.04.005Save Citation »Export Citation »E-mail Citation »

    The authors augment the classic political-economy forecast model with the number of days since the last election to account for the strategic nature of election timing in Japan. Their preferred seat-share model includes this new variable as well as change in gross domestic product and prime minister approval.

    Find this resource:

  • Norpoth, Helmut, and Thomas Gschwend. “Against All Odds? The Red-Green Victory.” German Politics and Society 21 (2003): 15–34.

    DOI: 10.3167/104503003782353619Save Citation »Export Citation »E-mail Citation »

    The model relies on chancellor preference, past party support, and terms in office to forecast the vote of the governing coalition. The model predicted the victory of the SPD-Greens coalition in the 2002 election with precisely the vote share it received, beating both pre-election and exit polls.

    Find this resource:

  • Stegmaier, Mary, and Michael S. Lewis-Beck. “Learning the Economic Vote: Hungarian Forecasts, 1998–2010.” Politics and Policy 37.4 (2009): 769–780.

    DOI: 10.1111/j.1747-1346.2009.00197.xSave Citation »Export Citation »E-mail Citation »

    The authors forecast Socialist Party support based on quarterly vote intention data, covering 1991–1998 and then 2002–2009, as a function of previous quarter vote intention and unemployment lagged two quarters. The authors observe a pattern of learning as the electorate shifted from policy-oriented support to the classic reward-punishment approach.

    Find this resource:

  • Toros, Emre. “Forecasting Elections in Turkey.” International Journal of Forecasting 27.4 (2011): 1248–1258.

    DOI: 10.1016/j.ijforecast.2011.01.002Save Citation »Export Citation »E-mail Citation »

    Toros applies the political-economy model with context-appropriate modifications. In addition to growth in gross national product, the model accounts for the difference in lead-party vote share in the general and local elections, and a dummy variable to capture tensions between the center and the periphery.

    Find this resource:

Election Markets

Often considered a “new approach” to forecasting elections, election markets have existed in the United States for more than one hundred years, though changes in government regulations moved the markets underground in the 1930s (Rhode and Strumpf 2004). Academic interest in the power of predictive markets (Wolfers and Zitzewitz 2004) led a group of economists at the University of Iowa to establish the Iowa Electronic Markets (IEM) in 1988, launching a new approach to election forecasting. The 2016 election witnessed the arrival of another online electoral market—PredicIt, which attracted a large clientele. Anyone can trade in these prediction markets, though with limited investments (a maximum of $500 at IEM, $850 at PredictIt per contract). The IEM hosts markets for presidential elections, party nominations, and control of the US House and Senate. In the winner-takes-all markets, candidate (or party) contracts are traded based on market values ranging from $.00 to $1.00. Contracts on the winning candidate (or party) have a value of $1, while the loser’s value is $.0. When the market is active, the contract values reflect investors’ views on how likely the candidate or party is to win the contest. Given this, how well do these markets predict election outcomes? This has been the source of much controversy, and the answer depends on the data being compared. When the comparison is made between raw market and poll data across countries or over the course of election campaigns, the markets on average outperform the polls. This is particularly true early in the campaign (Berg, et al. 2008a and Berg, et al. 2008b). Erikson and Wlezien 2008 argues that directly comparing markets to polls is flawed, because markets measure who investors think will win, while polls capture who would win if the election were held that day. After methodologically accounting for this, they find that US polls outperform markets. And, prior to World War II, in a time before scientific public opinion polls, the historical New York markets were actually more accurate predictors of elections than the IEM is today (Erikson and Wlezien 2012). Instead of using IEM data, Rothschild 2009 compares Intrade election shares to FiveThirtyEight polls. This analysis shows that the debiased market values outperform the debiased polls, especially early in the campaign and when the race is close. While betting on politics is illegal today in the United States (with the exception of the IEM), these markets exist in other democracies and have been used by scholars for forecasting. Using Betfair data, the authors of Wall, et al. 2012 introduce a novel approach to the challenge of predicting party seat shares in UK parliamentary elections. The authors predict which candidate will win in each constituency based on the betting data, and then forecast the overall parliamentary seat share for the main parties.

  • Berg, Joyce E., Robert Forsythe, Forrest Nelson, and Thomas Rietz. “Results from a Dozen Years of Election Futures Market Research.” In The Handbook of Experimental Economics Results. Vol. 1. Edited by Charles R. Plott and Vernon L. Smith, 742–751. Amsterdam: North Holland, 2008a.

    DOI: 10.1016/S1574-0722(07)00080-7Save Citation »Export Citation »E-mail Citation »

    A detailed overview of how the Iowa Electronic Markets (IEM) operate. American and other national election markets run by the IEM are compared to polls during the last week of the campaigns. Across these elections, on average, the market has less error than the polls.

    Find this resource:

  • Berg, Joyce E., Forrest D. Nelson, and Thomas A. Rietz. “Prediction Market Accuracy in the Long Run.” International Journal of Forecasting 24.2 (2008b): 285–300.

    DOI: 10.1016/j.ijforecast.2008.03.007Save Citation »Export Citation »E-mail Citation »

    An assessment of election prediction market versus poll accuracy in US presidential elections, 1988–2004. The markets typically offer a more accurate prediction than the polls. When the lead time is more than one hundred days prior to the election, the markets prove to be much better predictors than the polls.

    Find this resource:

  • Erikson, Robert S., and Christopher Wlezien. “Are Political Markets Really Superior to Polls as Election Predictors?” Public Opinion Quarterly 72.2 (2008): 190–215.

    DOI: 10.1093/poq/nfn010Save Citation »Export Citation »E-mail Citation »

    Analysis of the predictive accuracy of polls compared to both the vote-share and winner-take-all markets. The analysis shows greater forecasting accuracy of polls. The vote-share market lags behind the polls, and the winner-take-all market appears to overestimate the impact of campaign events and shocks.

    Find this resource:

  • Erikson, Robert S., and Christopher Wlezien. “Markets vs. Polls as Election Predictors: An Historical Assessment.” Electoral Studies 31.3 (2012): 532–539.

    DOI: 10.1016/j.electstud.2012.04.008Save Citation »Export Citation »E-mail Citation »

    Evaluation of the accuracy of election prediction markets before and after the development of scientific public opinion polling. They find that the pre–World War II New York markets had better predictive accuracy than modern markets, and that political market share prices today respond to the polls.

    Find this resource:

  • Rhode, Paul W., and Koleman S. Strumpf. “Historical Presidential Betting Markets.” Journal of Economic Perspectives 18.2 (2004): 127–142.

    DOI: 10.1257/0895330041371277Save Citation »Export Citation »E-mail Citation »

    A fascinating look at pre–World War II New York election betting markets. Through newspaper reports, the authors document the investment volume and predictive accuracy of these markets. The betting odds quite accurately projected election outcomes well in advance of Election Day.

    Find this resource:

  • Rothschild, David. “Forecasting Elections: Comparing Prediction Markets, Polls and Their Biases.” Public Opinion Quarterly 73.5 (2009): 895–916.

    DOI: 10.1093/poq/nfp082Save Citation »Export Citation »E-mail Citation »

    Comparison of the accuracy of polls and markets in the 2008 US election campaign. The author compares FiveThirtyEight’s debiased poll forecasts to debiased Intrade prices and finds that the debiased market values are more accurate predictors than the polls, particularly early in the election cycle.

    Find this resource:

  • Wall, Matthew, Maria Laura Sudulich, and Kevin Cunningham. “What Are the Odds? Using Constituency-Level Betting Markets to Forecast Seat Shares in the 2010 UK General Elections.” Journal of Elections, Public Opinion and Parties 22.1 (2012): 3–26.

    DOI: 10.1080/17457289.2011.629727Save Citation »Export Citation »E-mail Citation »

    Using the public betting market data for the 2010 UK parliamentary election in each constituency, the authors forecast the seat shares for each of the main parties. This approach performs worse than other forecasting methods, primarily because the markets overestimate the chances of long-shot candidates.

    Find this resource:

  • Wolfers, Justin, and Eric Zitzewitz. “Prediction Markets.” Journal of Economic Perspectives 18.2 (2004): 107–126.

    DOI: 10.1257/0895330041371321Save Citation »Export Citation »E-mail Citation »

    This article discusses the development and uses of prediction or event futures markets, placing political markets, such as the Iowa Electronic Markets, in the context of the variety of prediction markets that exist today. Available online as NBER Working Paper 10504.

    Find this resource:

Other Approaches to Election Forecasting

The previous sections have covered the dominant approaches to election prediction: polls, models, and markets. Alternative approaches with theoretical grounding in a variety of disciplines have also demonstrated accuracy in predicting national election results. One method combines forecasting approaches. The synthetic forecast developed in Lewis-Beck and Dassonneville 2015 melds the static political-economy model with polling data updates in the United Kingdom, France, and Germany. Graefe 2015 offers an elaborate method for combining forecasts. The author generates the average forecast within four forecasting methods and then takes the average of the four to predict daily the German election result. Another method, “citizen forecasting,” bases the prediction model on a survey question that asks who the respondent thinks will win. The authors of Lewis-Beck and Tien 1999 aggregates responses to this question from American National Election Study data from 1956 to 1996, and uses the percentage who think the incumbent will win as the predictor in their vote share forecast model. Murr 2011 enhances this technique by using citizen forecasts to predict UK constituency winners in 2010 and then aggregates these to calculate the partisan composition of parliament. Accounting for the traits or looks of successful presidential candidates is another line of election forecasting. This has been done through competence ratings based on photos of the 2008 Democratic and Republican primary contenders (Armstrong, et al. 2010), and through the development of indexes, which score candidates as either possessing or not possessing attributes on a variety of indicators. The candidate with the highest total score is projected to win the presidency. The “Keys to the White House” index (Lichtman 2016) comprises thirteen measures tapping into incumbent party policy performance, candidate characteristics, and aspects of the campaign. The authors of Armstrong and Graefe 2011 develop an index based on candidate biographies. Armstrong and Graefe’s fifty-nine factors cover everything ranging from candidate education, family, and political experience to physical attributes. The rise of social networks and online information sharing has provided new opportunities for citizens to express their political opinions. To what extent do these aggregated online exchanges reflect public opinion? This is the core question addressed in Tumasjan, et al. 2011, a study of Twitter tweets prior to the 2009 German federal election. The authors found remarkable accuracy between the share of party mentions in the tweets and the actual party vote shares, suggesting that, at least in this particular case, online information sharing reflects public opinion, and therefore could be used for election prediction.

  • Armstrong, J. Scott, and Andreas Graefe. “Predicting Elections from Biographical Information about Candidates: A Test of the Index Method.” Journal of Business Research 64.7 (2011): 699–706.

    DOI: 10.1016/j.jbusres.2010.08.005Save Citation »Export Citation »E-mail Citation »

    Armstrong and Graefe create a predictive index based on candidate biographies. Candidate education, family, political experience, and adversity traits compose this fifty-nine-item “bio-index.” From 1896 to 2008, the index predicts twenty-seven of the twenty-nine presidential winners correctly, outperforming polls, markets, and models.

    Find this resource:

  • Armstrong, J. Scott, Kesten C. Green, Randall J. Jones Jr., and Malcolm J. Wright. “Predicting Elections from Politicians’ Faces.” International Journal of Public Opinion Research 22.4 (2010): 511–522.

    DOI: 10.1093/ijpor/edq038Save Citation »Export Citation »E-mail Citation »

    Using black-and-white photos of the 2008 Democratic and Republican Party primary contenders, the authors ask people to evaluate candidate competency. They average these ratings and use them to predict performance in the nomination contest and in the general election.

    Find this resource:

  • Graefe, Andreas. “German Election Forecasting: Comparing and Combining Methods for 2013.” German Politics 24.2 (2015): 195–204.

    DOI: 10.1080/09644008.2015.1024240Save Citation »Export Citation »E-mail Citation »

    The article reviews the different forecasting approaches for the 2013 German election: polls, prediction markets, expert judgements, and models. Graefe generates a Combined PollyVote Model, averaging within and across the four components daily, to predict party vote shares. This combined model reduces forecasting error.

    Find this resource:

  • Lewis-Beck, Michael S., and Ruth Dassonneville. “Comparative Election Forecasting: Further Insights from Synthetic Models.” Electoral Studies 39 (2015): 275–283.

    DOI: 10.1016/j.electstud.2015.03.018Save Citation »Export Citation »E-mail Citation »

    The synthetic forecasting model combines a traditional static forecasting model that measures government approval and the economy six months prior to the election, and continuously updates this forecast with vote intention polls. The addition of vote intention polls captures campaign effects on the election result.

    Find this resource:

  • Lewis-Beck, Michael S., and Charles Tien. “Voters as Forecasters: Micromodels of Presidential Election Prediction.” International Journal of Forecasting 15.2 (1999): 175–184.

    DOI: 10.1016/S0169-2070(98)00063-6Save Citation »Export Citation »E-mail Citation »

    The authors use citizen vote expectations (“Who do you think will win?”) to assess how well Americans can predict presidential elections. In elections 1956–1996, 71 percent of respondents made a correct prediction. These aggregated citizen predictions are used to predict incumbent party vote share.

    Find this resource:

  • Lichtman, Allan J. Predicting the Next President: The Keys to the White House 2016. Lanham, MD: Rowman & Littlefield, 2016.

    DOI: 10.1016/j.ijforecast.2008.02.004Save Citation »Export Citation »E-mail Citation »

    The Keys to the White House index, made up of thirteen indicators, heavily weights the president’s economic and foreign policy performance. In addition, it accounts for features of the election campaign, scandal, and candidate charisma.

    Find this resource:

  • Murr, Andreas Erwin. “Wisdom of Crowds”? A Decentralised Election Forecasting Model That Uses Citizens’ Local Expectations.” Electoral Studies 30.4 (2011): 771–783.

    DOI: 10.1016/j.electstud.2011.07.005Save Citation »Export Citation »E-mail Citation »

    Murr uses 2010 British Election Study Internet survey data, with more than 16,000 respondents, to predict the winner in each UK constituency. He makes this prediction based on who the plurality of respondents think will win. From these constituency predictions, he projects the party composition of parliament.

    Find this resource:

  • Tumasjan, Andranik, Timm O. Sprenger, Philipp G. Sandner, and Isabell M. Welpe. “Election Forecasts with Twitter: How 140 Characters Reflect the Political Landscape.” Social Science Computer Review 29.4 (2011): 402–418.

    DOI: 10.1177/0894439310386557Save Citation »Export Citation »E-mail Citation »

    This analysis of more than 100,000 political tweets prior to the 2009 German federal election shows that the share of mentions each party received closely reflected the vote share received by that party. The authors compare this method of information aggregation for prediction purposes to the Iowa Electronic Markets.

    Find this resource:

back to top