The US Elections: Prediction Markets vs Forecasting Models
Preface
This is about predicting election outcomes for the President, Senate, and House using prediction markets and forecasting models.
Background
Prediction markets
A prediction market is a betting site where the traders trade binomial options (Yes/No) and pay X cents on the $dollar. If their bet wins, they get 100 cents on the $dollar. If their bet loses, they get 0 cents. As an example, at PredictIt, you can buy a Trump wins-Yes for 54 cents. If he wins, you will get 100 cents. If he loses, you will get 0 cents.
The beauty of this is that these binomial option prices can readily be translated into probabilities. The probability that Trump will win can be calculated as:
54/(54+50) = 51.9%
If you are a believer in the Efficient Market Hypothesis (EMH), you will give much cred to this type of information.
For the most part, prediction markets are deemed more predictive than polls because:
- They are live 24/7. They respond to events as they occur. They do not lag by a week or more.
- The volume of traders is far larger than the polls samples that are typically limited at 1,000 with an associated error margin of + or — 3%. Thus, the “small sample error” factor may be a lot stronger within polls vs prediction markets.
- Prediction markets may be less biased than polls. Traders have skin in the game. They have money at stake. Therefore, regardless of how they will vote, they are still mostly interested in being right and avoiding losing money.
- Prediction markets leverage the information from polls, not the reverse. Prediction markets also factor external economic, political factors that polls do not factor.
You may hear about prediction markets price manipulation. But, it may be challenging to pull this off. Many of the prediction markets have limitations on the size of the trades which may render manipulation challenging.
Whether prediction markets are more predictive than forecasting models is an open question. not
Forecasting models
Forecasting models are econometrics models using regression methodology (most probably Logistic Regression). And, they include two sorts of inputs:
a) Polls at the State level so the model can figure out the resulting Electoral College aggregation reasonably accurately; and
b) Economic variables including inflation, unemployment rate, GDP growth, etc.
These models are trained on an extensive history of previous elections going back several decades.
Early into the Presidential election cycle, these models weight more on the economic fundamentals than the polls input. As we get closer to Election Day, they progressively weight more on the polls and less on the economic fundamentals.
A big missing piece: Super Forecasters
Philip Tetlock is the premier social scientist of predictions you may not have heard off.
While Daniel Kahneman got the credit for documenting the flaws in human judgment, Tetlock has gone far beyond in testing how bad the predictions of “experts” are. He has come up with the Good Judgment Project. And, he has trained a group of human beings who are unusually good at predictions. There is much literature confirming that the Super Forecasters are even more accurate than prediction markets.
https://goodjudgment.com/resources/the-superforecasters-track-record/
I would have loved to obtain the Super Forecasters election forecasts. I contacted the organization several times with no answer. But, you can obtain their election forecast for $200.
https://goodjudgment.io/FFSubscribe/index.php#US%20Elections
Introduction
I will compare the elections predictions of four prediction markets with four forecasting models.
The prediction markets are:
- Predict It
- Polymarket
- Smarkets
- Kalshi
The forecasting models are:
- The Economist
- 538
- Split Ticket
- Race to White House
The probabilities for the Presidency, Senate, House
Within the table below see their respective probabilities for the Presidency, Senate, and House.
See below the probabilities’ averages.
Regarding the Senate and House there is much convergence. The prediction markets and the forecasting models converge towards a 2/3d to 3/4th probability that Republicans win the Senate and around a mid-fifties probability that the Democrats win the House (538 is a notable exception that assigns a similar probability but in the Republicans’ favor).
When focusing on the Presidency, there is a marked demarcation between the two. The prediction markets favor Trump (avg. 55.4%). Meanwhile, the forecasting models favor Harris (53.1%).
Visualizing the President probabilities divergence
The data visualization for the different prediction markets and forecasting models is very uneven. Given that I will focus only on the few with a reasonably good data visualization interface.
Prediction markets
Here is a 30 day-trend from Predict It. We can see that Harris was leading for a majoriy of the time. But, her lead was on a declining trend throughout the entire 30 day period.
If we focus over the week from October 8 to the 15, there is a clear jump in Trump’s favor from October 8 to the 11th. Additionally, you can see a big jump in trading volume. So something occurred over that short time frame which rendered traders bullish on Trump.
You can see a similar trend at Kalshi. Notice the big jump for Trump around October 9th.
The other prediction markets confirmed a similar trend. There may be some arbitrage opportunities across these prediction markets. So, their respective prices/probabilities do converge quite a bit.
Forecasting models
The trend for the three forecasting models with good data visualization (538, The Economist, Race to the WH) are all very similar. Harris probabilities peaked in mid-September. This was probably within a week of the Harris — Trump debate. During that debate Harris performed well as she successfully “prosecuted” Trump. Meanwhile, Trump’s performance was as embarrassing to himself and his party as Biden’s was during their respective debate.
After the mid-September peak, Harris probabilities progressively mean-reverted to closer to 50%. However, unlike within the prediction markets they never ever came close to flipping over in Trump’s favor.
Even though 538, The Economist, and ‘Race to the WH’ developed their models independently, their respective trends since early September are very similar. This may be due to their weighting polls more than economic fundamentals as we get closer to the election. They apparently aggregate the polls and weight them in a similar fashion.
Explaining the divergence between prediction markets and forecasting models since October 8th?
I asked Perplexity. And, it mainly focused on the difference in methodology between the two that I reviewed earlier.
https://www.perplexity.ai/search/among-presidential-election-pr-oCJCLsE7ToSaf18GwwW6IA
At random, I can think of a few things explaining the divergence including:
- The VP Debate on October 1st, when J.D. Vance made a far better case for Trumpism than Trump did. And, Tim Walz did not “prosecute” Trumpism as aggressively as Harris did.
- Mounting tension in the Middle East causing voters to become increasingly inclined toward Isolationism favoring Trumpism.
But, I am just grabbing at straws. And, I have to wonder if the divergence between the two is spurious.
Maybe the prediction markets are veering off the road. And, the forecasting models are closer to the prospective reality. Or maybe not…
Keep in mind that whether you go with the prediction markets at about 43%/57% in Trump’s favor or 56%/44% in Harris’ favor, you are still very close to a coin flip.
It will be interesting to check if the prediction markets and the forecasting models do converge by November 4th.
THE END
