The New York Times recently wrote about how “The Fed’s Decisions Now Could Alter the 2024 Elections” (11/17/2023) and cited Yale Professor Ray Fair’s economy-driven election forecasting model:
Professor Fair’s pioneering U.S. elections model does something that was fairly radical when he created it in the 1970s.
It analyzes politics without really considering politics.
Instead, Professor Fair focuses on economic growth, inflation and unemployment. With a few tweaks through the years, he has used economics to analyze elections since 1978, based on data for elections going back to 1916.
What he’s found is that the economy sets the climate for national elections. The candidates and the political parties must live within it.
Professor Fair is not just a pioneer in election modeling but also the only election forecaster to run a sub-3 hour marathon (let’s see Nate Silver try that). What’s interesting about this particular election model (Fair has other models that we’re not getting into here) is twofold:
- It leans on economic indicators independent of presidential candidates, campaigns, or issues.
- Rather than predicting winners state-by-state, Fair tries to estimate the vote share between the two major parties, ignoring any third-party vote.
His 1978 paper “The Effect of Economic Events on Votes for President” was an early version of the model that Fair has updated over the years with new information from the election results and economic activity. Here is the model for the 2024 presidential race with a Democratic incumbent running for re-election:
Democratic Vote Percentage = 49.48 + 0.708*G – 0.606*P + 0.865*Z
If a statistical model seems intimidating, please just think of a model as a math formula for predicting something, in this case, the Democratic presidential candidate’s vote percentage. Each variable (like GDP or incumbent party) affects the predicted vote share differently. In this formula, coefficients (0.708, 0.606, and 0.865) are numbers telling us how much their respective variables (G, P, and Z) change the vote percentage. If the coefficient for G is 0.708, it means a one percentage point increase in G leads to a 0.708 percentage point increase in the Democratic vote percentage. A positive number means the Democratic vote percentage goes up as the coefficient increases, and a negative number means the vote percentage goes down. Essentially, coefficients help us understand how much and in what way each part of the formula influences the final prediction for the Democratic vote percentage. We’ll summarize the individual model components next.
Looking at the economic variables used:
- G = growth rate of real per capita GDP in the first 3 quarters of 2024 (annual rate)
- P = growth rate of the GDP deflator in the first 15 quarters of the Biden administration, 2021:1-2024:3 (annual rate)
- Z = number of quarters in the first 15 quarters of the Biden administration in which the growth rate of real per capita GDP is greater than 3.2 percent at an annual rate
The constant of 49.48 is derived from the following “non-economic variables” that Fair uses in the model:
- I = 1 (Democratic presidential incumbent at the time of the election)
- DPER = 1 (a Democratic presidential incumbent is running again))
- DUR = 0 (neither party has been in the White House for more than one term)
- WAR = 0 (No present world war)
- A model intercept of 48.22 (the predicted value if all variables equal zero)
Here is the formula for arriving at the constant of 49.48; it will not change as long as President Joe Biden runs for re-election. The non-economic variables are multiplied by their coefficients and added together, along with the model intercept of 48.22.
Constant = (I * -0.85) + (DPER * 2.11) + (DUR * -3.45) + (WAR * 3.90) + Intercept
Constant = (1 * -0.85) + (1 * 2.11) + (0 * -3.45) + (0 * 3.90) + 48.22
Constant = -0.85 + 2.11 + 0 + 0 + 48.22 = 49.48
Right now, G = 1.40, P = 4.78, and Z = 4, which gives us the current forecast for the 2024 election:
Democratic Vote Percentage = 49.48 + 0.708*G – 0.606*P + 0.865*Z
Democratic Vote Percentage = 49.48 + (0.708 * 1.40) – (0.606*4.78) + (0.865*4)
Democratic Vote Percentage = 49.48 + 0.9912 – 2.89668 + 3.46 = 51.03
The model currently predicts (with data as of 10/26/2023) that the Democratic share of the two-party vote will be 51.03% in the 2024 Presidential election.
Fair’s economy-driven model has proven robust over time. Here is a table of forecasts and model errors, courtesy of PollyVote:
In this article we’ll be looking at the forecasts above from Professor Fair’s two-party vote share model and comparing them to the Iowa Electronic Markets (IEM) share prices. There have been many comparisons between FiveThirtyEight and PredictIt, but today we’re interested in a head-to-head comparison between the grandaddies of election models and prediction markets.
The longest-running (legal in these United States) prediction market, the University of Iowa was first granted a no-action letter by the Commodity Futures Trading Commission for the “Iowa Presidential Stock Market” in 1992, and then another no-action letter in 1993 for the “Iowa Electronic Markets” that IEM has been operating under ever since. Clunky but reliable, IEM allows traders to buy and sell different election futures contracts through its website (although traders must first snail mail their money and applications to IEM.) As a research project, there are restrictions IEM must abide by, such as a market limit of $500 per trader.
IEM has multiple markets for Presidential elections, but we’ll be looking at its two-party vote share market:
Payoffs in the 2024 Presidential Vote-Share Market will be determined by the percentages of the two-party popular vote received by the official Democratic and Republican nominees in the 2024 U.S. Presidential Election. Contracts will be associated with the candidates officially nominated by each party. For instance, contracts for a candidate who receives 32.4% of the popular votes cast for the Democratic and Republican candidates, will be worth 32.4 cents each. Payoffs are NOT affected by votes received by nominees from other parties, the outcome of the electoral college or any vote taken by the House of Representatives should such a vote be necessary.
Below are the Democratic vote share forecasts from the Fair model along with the corresponding Democratic vote share contract prices from IEM for presidential elections from 2008 to 2020:
| Election | Result | Forecast date | Days out | Fair VS | Fair Error | IEM VS | IEM Error | AVG | AVG Error |
| 2008 | 53.7 | 11/1/2006 | 734 | 53.5 | -0.2 | 55.0 | 1.3 | 54.3 | 0.5 |
| 2008 | 53.7 | 1/31/2007 | 643 | 53.4 | -0.3 | 53.1 | -0.6 | 53.3 | -0.5 |
| 2008 | 53.7 | 4/27/2007 | 557 | 53.2 | -0.5 | 52.0 | -1.7 | 52.6 | -1.1 |
| 2008 | 53.7 | 7/27/2007 | 466 | 52.0 | -1.7 | 51.9 | -1.8 | 52.0 | -1.8 |
| 2008 | 53.7 | 10/31/2007 | 370 | 51.9 | -1.8 | 52.7 | -1.0 | 52.3 | -1.4 |
| 2008 | 53.7 | 1/31/2008 | 278 | 52.0 | -1.7 | 52.6 | -1.1 | 52.3 | -1.4 |
| 2008 | 53.7 | 4/30/2008 | 188 | 52.2 | -1.5 | 53.3 | -0.4 | 52.8 | -1.0 |
| 2008 | 53.7 | 7/31/2008 | 96 | 51.5 | -2.2 | 54.4 | 0.7 | 53.0 | -0.8 |
| 2008 | 53.7 | 10/30/2008 | 5 | 51.9 | -1.8 | 54.3 | 0.6 | 53.1 | -0.6 |
| 2012 | 52.0 | 7/31/2011 | 464 | 53.4 | 1.4 | 54.0 | 2.0 | 53.7 | 1.7 |
| 2012 | 52.0 | 10/30/2011 | 373 | 50.0 | -2 | 49.9 | -2.1 | 50.0 | -2.1 |
| 2012 | 52.0 | 1/28/2012 | 283 | 50.3 | -1.7 | 49.0 | -3.0 | 49.7 | -2.4 |
| 2012 | 52.0 | 4/27/2012 | 193 | 50.2 | -1.8 | 52.7 | 0.7 | 51.5 | -0.5 |
| 2012 | 52.0 | 7/27/2012 | 102 | 49.5 | -2.5 | 51.3 | -0.7 | 50.4 | -1.6 |
| 2012 | 52.0 | 10/26/2012 | 11 | 49 | -3 | 52.0 | 0.0 | 50.5 | -1.5 |
| 2016 | 51.1 | 1/31/2015 | 647 | 46 | -5.1 | 52.3 | 1.2 | 49.2 | -2.0 |
| 2016 | 51.1 | 4/29/2015 | 559 | 48.6 | -2.5 | 53.4 | 2.3 | 51.0 | -0.1 |
| 2016 | 51.1 | 7/31/2015 | 466 | 46.4 | -4.7 | 49.3 | -1.8 | 47.9 | -3.3 |
| 2016 | 51.1 | 10/31/2015 | 374 | 45.8 | -5.3 | 51.4 | 0.3 | 48.6 | -2.5 |
| 2016 | 51.1 | 1/30/2016 | 283 | 45.7 | -5.4 | 54.6 | 3.5 | 50.2 | -0.9 |
| 2016 | 51.1 | 4/28/2016 | 194 | 45 | -6.1 | 59.0 | 7.9 | 52.0 | 0.9 |
| 2016 | 51.1 | 7/29/2016 | 102 | 44 | -7.1 | 51.1 | 0.0 | 47.6 | -3.6 |
| 2016 | 51.1 | 10/28/2016 | 11 | 44 | -7.1 | 58.0 | 6.9 | 51.0 | -0.1 |
| 2020 | 52.3 | 4/26/2019 | 557 | 45.4 | -6.9 | 55.6 | 3.3 | 50.5 | -1.8 |
| 2020 | 52.3 | 7/26/2019 | 466 | 46.2 | -6.1 | 54.1 | 1.8 | 50.2 | -2.1 |
| 2020 | 52.3 | 10/31/2019 | 369 | 45.9 | -6.4 | 50.3 | -2.0 | 48.1 | -4.2 |
| 2020 | 52.3 | 1/30/2020 | 278 | 45.6 | -6.7 | 56.2 | 3.9 | 50.9 | -1.4 |
| 2020 | 52.3 | 5/15/2020 | 172 | 54.2 | 1.9 | 51.0 | -1.3 | 52.6 | 0.3 |
| 2020 | 52.3 | 8/21/2020 | 74 | 50.5 | -1.7 | 51.1 | -1.2 | 50.8 | -1.5 |
| 2020 | 52.3 | 10/29/2020 | 5 | 47.9 | -4.4 | 53.4 | 1.1 | 50.7 | -1.7 |
In these 30 head-to-head comparisons, both model forecasts were shockingly close, even over a year out from the election.
The Fair model underestimated the Democratic candidate’s share of the two-party vote in 28 out of 30 predictions. On the other hand, IEM traders underestimated the Democratic vote share exactly half the time in 15 out of 30 predictions.
Given that Fair and IEM are using different methodologies, we decided to include a column for their average vote share prediction (AVG) and the absolute error of the averaged predictions (AVG Error). This ensemble performed better than either Fair or IEM using root-mean-square error (RMSE) on those 30 predictions:
| Forecast | RMSE |
| Fair | 4.06 |
| IEM | 2.58 |
| AVG | 1.79 |
Please note the following:
- This comparison only uses Fair predictions that overlap with IEM market data. There were six forecasts not used because they were before an IEM vote share market had opened.
- IEM is using “Last Price” which is the last trade before midnight of that day. The trade generating this price may have occurred on a previous day if there was no trading of the contract on that specific day.
- The Fair model forecasts both go farther back in time (1980) than the IEM has Presidential vote share market data publicly accessible (2008).
- Dates compared are very far out from Election Day, up to 734 days out, and so there are events the model couldn’t have taken into account, but which very well helped determine the outcome.
This is not meant to be representative of the entire work of either Professor Fair’s vote share model or the IEM traders, but simply to compare them head-to-head when possible. While Fair’s model is static, or at least dependent on economic statistics, IEM traders can react to events in real time. And it’s also possible for traders to be making decisions based off of Fair’s model.
Finally, for whatever reason, neither Professor Fair nor the IEM have convenient user interfaces to their websites or intuitive archives for their historical data. Below are links to election forecasting information that may be of interest.
Ray Fair:
Official website
2024 vote share predictions
2024 vote share model inputs
2020 vote share predictions
1996-2018 vote share predictions
“Papers on Voting”
Ranking Assumption model
Predicting Presidential Elections and Other Things
An Election Economics Post-Mortem (PBS)
Forecasting U.S. Presidential Elections: A Brief Review (Foresight)
Iowa Electronic Markets:
IEM Registration
Trader Manual
2024 Presidential vote share market
2024 markets
1998-2022 markets
1988-1996 markets