Predicting an Election

I spent an hour or so with Glenn Loury last Sunday on his podcast, and the episode is now available on all the usual streaming platforms.

We scheduled the conversation to talk about election forecasts, but ended up straying far beyond this topic to discuss a recent FIRE conference, Lara Bazelon on the ACLU, a controversy at Hamline University, an old article by Ta-Nehisi Coates, Ralph Wiley’s response to Saul Bellow, Benjamin Franklin’s edits to the Declaration of Independence, and my mother’s connection to the mother of the Vice-President.

Along the way, we did get to the question of why statistical models and prediction markets are generating quite different forecasts for the coming election. To be honest, the following thirteen minute clip encapsulates just about everything useful I can say about the subject, thanks to some precise and thoughtful questions from Glenn:

If you prefer to read rather than listen, here’s the gist of it.

First, there has been a clear and persistent gap between the forecasts of models and markets over the past month or so. For example, on the morning of October 30, the Economist model had the race dead even while Polymarket gave Trump a 67 percent chance of victory. There’s been some convergence since then but a sizeable gap remains.

Second, differences of this magnitude should surprise nobody. Models and markets are distinct forecasting mechanisms based on completely different sources of information. Models are built and calibrated based on historical data and operate under the assumption that the past is a reasonably good guide to the future. Markets are not constrained in this way at all—anything a trader considers relevant comes to be incorporated into the market price. This flexibility is a strength of markets, especially when we are in uncharted waters and models start to generate implausible forecasts.

Third, while markets have strengths, they also have weaknesses. They can come to be dominated by a small set of deep-pocketed traders, whose beliefs then have a disproportionate effect on price and dilute the wisdom of crowds effect. They are prone to irrational exuberance. And there exists potential for manipulation, since beliefs about a candidate’s viability and momentum can affect morale, donations, volunteer effort, and turnout. Pessimism about a campaign can become self-fulfilling, which creates incentives for engineering pessimism about an opponent.

Fourth, markets cannot deviate too far from each other because of arbitrage—if a Trump contract is trading for 60¢ on one market and 50¢ on another, I can bet against him on the first and against his opponent on the second, effectively buying a dollar for 90¢. With many traders wanting to do this at scale, prices across markets tend to converge. This effect is mitigated by market segmentation—Polymarket is crypto-based and formally closed to those in the US, while Kalshi is cash-based and excludes foreign nationals. It is also mitigated by position size limits, such as those on PredictIt. Models, however, can deviate from each other substantially and persistently.

Fifth, the question of whether models or markets generate more accurate forecasts on average is an empirical question—one can’t reason one’s way to an answer based on logic alone. Traditional approaches to evaluating accuracy compare predictions to realizations, but one might alternatively ask whether trading on a market as if one believed a model makes or loses money. This profitability test can be implemented even prior to event realization, by tracking the value of a model-based portfolio.

Finally—and perhaps least obviously—the price on a prediction market is not a reflection of any trader’s belief. A contract that pays a dollar if Trump wins the election is currently trading on Kalshi at 56¢. Does this mean that traders on that market assign a probability of 56 percent to the event? Not at all. There are some who think it will happen with seventy or eighty percent, and others who think that even fifty is too high. Why, then are the former not snapping up contracts at the available price, and the latter not betting against the event? It’s because they have already done so. Having hit their risk limits or budget constraints, they don’t want to accumulate more contracts in the absence of new information.

This last point raises some interesting possibilities. The price doesn’t correspond to any commonly held belief, it is simply a point of balance between competing narratives about the election. So if evidence starts to emerge that punctures a hole in one of these narratives, a very sharp price movement can result.

I have no idea who will win this election, and wouldn’t be surprised if either one of the candidates sweeps all seven competitive states. But I suspect that we will see one or more really wild swings on prediction markets before polls close on election day.