Orhan Torul bio photo

Orhan Torul

Associate Professor of Economics at Boğaziçi University

  X LinkedIn   Google Scholar   ORCID E-Mail

When the Odds Do Not Agree

Through most of 2024, you could bet on the American presidential election in at least seven different places. Polymarket, Betfair, Bovada, BetOnline, William Hill, Unibet and Everygame all posted a running probability that Donald Trump would win, updated continuously as news arrived. They did not agree with one another.

Usually the disagreement was small. On an average day the seven estimates sat within about 2.6 percentage points of their own average. Around the big shocks of the campaign the spread more than doubled, passing 6 percentage points after the July assassination attempt and again after President Biden withdrew from the race. In a recent paper, my co-author Emrehan Aktuğ and I ask whether that spread, the bare fact of the platforms disagreeing, tells us anything about the stock market.

Why the Disagreement Rather Than the Odds Themselves

Most work on elections and markets asks a directional question: what are the odds, and which way are they moving? A single platform’s price answers that. If Polymarket says 60%, that is its traders’ collective best guess.

Disagreement asks something different. It is not about where the odds point, but about how firmly anyone holds them. The platforms differ in who trades on them, in which countries those traders sit and in how quickly each updates, so their numbers were never going to coincide exactly. When all seven look at the same campaign and arrive at visibly different answers, though, the natural reading is that the evidence itself has become hard to interpret. Investors then face a wider range of possible policy outcomes, and a wider range of outcomes leaves more room for prices to swing.

There are other ways to measure election uncertainty, and each has a drawback that disagreement across platforms avoids:

  • Disagreement among pollsters comes from surveys, which appear every few days at best. It can tell you nothing about a particular Tuesday afternoon.

  • Options prices yield a well-known gauge of expected stock market turbulence, updated constantly, but that gauge blends election worries together with everything else investors are worried about. The electoral part cannot be separated out.

  • Disagreement across betting platforms is specific to the election, comes from real money changing hands rather than from stated intentions, and can be read every five minutes.

Data and Design

We pair five-minute prices from the seven betting platforms with five-minute stock prices for 71 American companies, running from January 22 to November 5, 2024. The start date is the first trading day after Ron DeSantis dropped out of the Republican primary, so that the betting odds reflect the general election rather than the nomination contest.

We sort the companies by which party they give money to, using campaign contribution records compiled by OpenSecrets over the 2020 to 2024 cycles. Thirty-six lean Republican and thirty-five lean Democratic. We then build two equally weighted portfolios and study the difference between them. The reason for working with the difference is that most of what moves a large company’s share price has nothing to do with politics, whether interest rates, oil prices or consumer demand. Those forces push both portfolios in the same direction and largely cancel when one is subtracted from the other. What should survive the subtraction is sensitivity to who wins.

That sets up the test. If disagreement among the betting platforms genuinely captures election-specific uncertainty, it should predict turbulence in this politically sensitive portfolio more strongly than it predicts turbulence in the stock market as a whole. The comparison is what interests us, rather than the raw finding, because a comparison of this kind is much harder to produce by accident.

1. Disagreement Predicts Turbulence in Politically Exposed Stocks

When the betting platforms disagree more in one hour, the political portfolio swings more in the next. A typical widening in disagreement, slightly under one percentage point, is followed by a 4% to 11% increase in the portfolio’s volatility. By volatility we mean simply the size of its price swings, in either direction, over the following hour. The relationship holds after accounting for how turbulent the broad market was and for what options prices implied about expected turbulence.

Note that the prediction concerns the size of the swings and not their direction. That distinction turns out to matter a great deal, and we return to it below.

2. The Effect Is Concentrated Where Politics Matters

Disagreement predicts swings in the overall market far more weakly than it predicts swings in the political portfolio, and the gap between the two is large enough that it is unlikely to be chance. One subtlety had to be addressed before we could believe it, however. Our broad market measure is the average of the same 71 politically connected firms, which makes it an imperfect comparison. So we built a cleaner one: sixteen large utility and consumer staples companies, firms that sell electricity and groceries, which take no political side and appear nowhere among the original 71. Disagreement predicts swings in that basket about half as strongly as in the political portfolio, and the difference again holds up. Election disagreement is not entirely unrelated to ordinary market turbulence, but its predictive power is clearly concentrated where political exposure is greatest.

3. It Predicts How Much Prices Move, Not Which Way

This is the sharpest test available to us, and the one that separates an uncertainty story from a simple directional one. If disagreement merely signaled that Trump was gaining ground, it ought to predict which way the political portfolio moves. It barely does so. What it predicts strongly is the size of the move regardless of direction, with roughly ten times more explanatory power for size than for direction.

A small directional effect does survive, and we report it rather than rounding it away. In 2024 the platforms happened to drift apart at moments when the news favored Trump, and a thin risk premium that our measure of the consensus odds does not capture may also be at work. The channel is mostly, though not purely, about uncertainty.

4. How Long It Lasts, and Which Platforms Drive It

Two further results are worth reporting:

  • The effect lasts about a trading day: this fits the behavior of disagreement itself. Once the platforms drift apart they converge slowly, taking roughly 33 hours to close half the gap, so the underlying confusion stays elevated for hours rather than minutes.

  • The slower platforms matter more than the fast ones: a standard statistical method can rank the seven by how much each contributes to discovering the correct price. Polymarket and Betfair lead. The traditional bookmakers lag, and update their numbers far less often. It is disagreement among the laggards, not among the leaders, that tracks stock market turbulence most closely. One explanation is that the slow, retail-facing venues reflect the confusion of the general public and drift apart exactly when news is genuinely hard to read, while the leaders converge quickly and add little. We treat this as suggestive rather than settled, since these platforms also differ in liquidity and in who trades on them.

A Caveat About the Numbers

We gave the headline effect above as a range, 4% to 11%, rather than as a single number. The reason deserves a paragraph, because it is the sort of thing academic papers tend to bury.

Our main results use intraday stock prices from one data vendor, Twelve Data, and imply an 11% effect. Running exactly the same analysis on prices for the same companies from a second and more expensive vendor, DataBento, implies 4%. The direction and the statistical significance are identical; the magnitude is not. The explanation is mundane. The two vendors fill in their five-minute price grids at very different rates, roughly 95% of the time for DataBento against roughly 70% for Twelve Data, and they timestamp their data slightly differently. Since our hourly volatility measure is built from twelve five-minute price changes, it matters a great deal how many of those twelve are genuine observations rather than gaps filled in from the previous price. We report both figures rather than choosing the more flattering one.

What the Evidence Cannot Establish

Three limitations bound what we can claim:

  • This is prediction, not proof of cause: everything we can check points the same way. Disagreement comes first and turbulence follows rather than the reverse, the previous hour’s disagreement works as well as the current hour’s, and setting aside the big campaign event days changes nothing. Even so, we cannot rule out that some third factor moves both at once. The clean way to settle it would be an accident of nature, such as a platform going offline for technical reasons, which would scramble the disagreement measure for reasons unconnected to the election. No such outage occurred during our sample, and we flag the idea for whoever gets the data next time.

  • The event evidence is uneven: disagreement and market turbulence both jumped after the July 13 assassination attempt, and both fell after President Biden withdrew and removed a major unknown. Around the two televised debates the relationship is much weaker. The pattern hints that what matters is surprise, meaning unscheduled events that abruptly widen the range of possibilities rather than scheduled ones everyone has had weeks to anticipate.

  • It is one election: earlier American elections cannot be checked, because the data do not exist. Polymarket had not been founded and Betfair was effectively the only liquid market in 2016. We therefore present this as something true of 2024, not as a law of political markets.

What It Is Good For

For anyone holding shares in politically exposed companies, the practical use is ordinary risk management. Betting odds are public and free, they update in real time, and the disagreement among them forecasts turbulence in exactly the stocks whose value depends on the election. An investor trying to hold risk steady would trim a politically tilted position by roughly 4% to 10% when disagreement widens by a typical amount. As a forecasting matter, adding disagreement to the standard predictors raises the share of hourly turbulence we can explain from 42% to 45%.

The wider point concerns what we choose to read off these markets. The headline probability attracts nearly all the attention, and it answers the question everyone asks, which is who is going to win. The disagreement around it answers a quieter question about how much anyone really knows, and our results suggest that this quieter number carries real-time information about election uncertainty that the headline figure does not.

Whether the same holds for other high-stakes political events with several active betting markets remains open. Given how quickly these platforms have grown since 2024, it should not stay open for long.

References