Prediction Markets Are Structurally Inaccurate When Forecasting Democratic Decline

Prediction markets are structurally incapable of pricing long-horizon democratic decline, and EAs should stop treating them as a credible counterweight to expert opinion on this question.
This is a follow-up post to an October 2025 EAF article that documented prediction markets and expert sentiment towards the chance of authoritarian capture of the US government. After revisiting the polls, we see that the median market moved only 3pp across the nine questions we tracked, while expert assessments of American democracy fell by 14 points on a 0–100 scale across the same timeframe. Long-horizon questions like these are systematically mispriced and suffer from chronically low liquidity, as users realize they can leave their money in a high-yield savings account and see higher returns than buying the underpriced option and waiting years to cash out. This outcome, combined with an analysis of persistent flaws in the construction of prediction markets as a whole, lead us to conclude that prediction markets should not be prioritized over expert opinion with regard to long-horizon predictions.

Prediction Markets Barely Moved Over 5 Months

Will Trump seriously attempt a third term? – Manifold
- Oct 8 2025: 15% → March 5 2026: 18% → +3%
If elected in 2024, will Trump attempt a third term? – Manifold
- Oct 8 2025: 23% → March 5 2026: 20% → –3%
Trump chances of winning 2028 Presidential – Polymarket
- Oct 8 2025: 3% → March 5 2026: 2% → –1%
Will Trump retain executive power past 2028? – Metaculus
- Oct 8 2025: 3% → March 5 2026: 3% → +0%
Will The Economist stop classifying the US as a democracy before 2030? – Metaculus
- Oct 8 2025: 25% → March 5 2026: 17% → –8%
Will the US still be a liberal democracy at the end of Trump’s 2024 term? – Manifold
- Oct 8 2025: 69% → March 5 2026: 40% → –29%
Will Trump 2.0 be the end of democracy as we know it? – Manifold
- Oct 8 2025: 26% → March 5 2026: 23% → –3%
Will the USA enter a second civil war before 2031? – Metaculus
- Oct 8 2025: 3% → March 5 2026: 3% → +0%
Will there be a civil war in the US before 2030? – Manifold
- Oct 8 2025: 13% → March 5 2026: 8% → −5%

Analysis:
The median prediction market moved just 3 percentage points over five months. Markets with baseline scores at or below 5% (both civil war questions, Trump retaining power, and Trump winning in 2028) were structurally locked near the floor and moved only 0.5pp on average, while the five markets with room to move averaged 9.4pp, driven almost entirely by the liberal democracy question’s 29-point drop. If we remove that outlier, the remaining four actionable markets averaged only 3.75pp. Most markets that could have repriced in response to democratic erosion didn’t. A few forecasts even moved against expert opinion; the poll that measured the chance that the Economist would stop classifying the US as a democracy fell 8pp over the same time period in which experts registered a 14-point drop in democracy performance scores.

We can still see the imprint that some scandals and political upheavals left in a few of the forecasts: there was a 30% jump in pro-third-term sentiment across April 2025 in the pre-election Manifold poll, consistent with the administration’s Liberation Day tariffs kicking in on April 2 and the subsequent stock market crash. A different Manifold poll experienced fluctuations after resolution criteria changes and a conclusive 20% drop after the ICE shooting of Renee Good. But these brief spikes were exceptions that the markets quickly absorbed and smoothed over. The market is moving in a similarly pessimistic direction to expert opinion, but the dominant forecast pattern is underpriced stasis.

Prediction markets work well for short-horizon electoral questions, which makes their failure on long-horizon democratic-health questions all the more notable as a structural rather than general problem. Looking at forecasts for the midterms, prediction markets have a generally favorable outlook towards the Democratic party winning back the house, with closer to even odds of them sweeping both chambers of the Senate, and it’s easy to trace a jump in pro-Democratic sentiment to April 2025’s Liberation Day tariffs. Election Betting Odds maintains a tracker of opinions across the biggest prediction markets, and it shows the clearest picture. This spike in sentiment is slightly visible in Metaculus’ poll as well, which has otherwise held steady across its time on the platform. A similar spike in sentiment can be seen in a Manifold poll, with day-long spikes in October (coinciding with the government shutdown) and early January 2026 (Venezuela military operation and killing of Renee Good).

Why Prediction Markets Fail On Long-Horizon Questions
The standard defense of prediction markets, best articulated in Scott Alexander’s Prediction Market FAQ, is as follows: if prediction markets were systematically worse than experts, someone would notice and bet against expert consensus, correcting the mispricing in the process. This is a strong argument in favor of short-horizon, high-liquidity markets. It breaks down for long-horizon democratic decline questions for three reasons:

Opportunity Costs: Even if a long-horizon market is clearly mispriced, the profit from buying the underpriced option and waiting years to cash out may be less than putting your money in a savings account. A Works in Progress article noted that sophisticated traders are structurally less interested in long horizon bets because locking capital into a multi-year prediction market contract means foregoing years of interest earnings. Corroborating their analysis is a 20212 paper by Lionel Page and Robert Clemen, which analyzed over 500,000 real transactions across 1,787 markets on Intrade and found that time until expiration negatively affects price accuracy, with high-likelihood events systematically underpriced and low-likelihood events overpriced—markets were reasonably well calibrated when time to expiration was short but significantly biased for events farther in the future. The measured bias was 4.7 to 10.9 percentage points for markets with more than 100 days to expiration. This is the nail in the coffin of treating prediction market stasis as evidence of epistemic stability, rather than evidence that nobody has a financial incentive to correct the pricing error.

Fake Money And Misaligned Incentives: Manifold requires users to bet with fake currency, whereas Polymarket requires USD and Metaculus uses accuracy ratings for each bettor. The opportunity cost of making a wrong bet on Manifold is thus much lower than on Polymarket or Metaculus, where you would lose real currency or social currency on the back of a wrong prediction. This explains Manifold’s higher volatility, since users bet more freely and create larger price swings. Manifold’s long-horizon forecasts are thus more accurate than those of Metaculus or Polymarket, with the added caveat that they lose a vaunted feature of prediction markets that ensures their accuracy; namely, that predictions become more accurate when bettors are forced to suffer real financial loss if they overindex on an incorrect forecast.

Confusing Resolution Criteria: the criteria for a lot of predictions we’ve been looking at are often more convoluted or counterintuitive than they might seem at first glance: for example, this Metaculus poll about civil war in the US requires at least 500 combat deaths to occur within a 6 month period, but ONLY after the Insurrection Act has been invoked. This probably wouldn’t align with the average bettor’s understanding of a civil war. In another example, a suggested criteria change on the Manifold outlier poll caused an 18pp drop, demonstrating that user sentiment can be driven by question wording rather than genuine belief updating. This compromises the trustworthiness of prediction markets against, say, public polling, in which standard practice is to have experts word their polling questions extremely carefully to prevent confusion and skewed results.

Expert Opinion Vs. The Wisdom Of Crowds
While prediction markets drifted, expert assessments of American democracy underwent a dramatic decline. Bright Line Watch, which surveys political scientists at American colleges and universities, recorded a drop from ⁶⁷⁄₁₀₀ in November 2024 to ⁵³⁄₁₀₀ by April 2025; a 14-point collapse in five months. When asked to forecast their rating for 2027, experts projected a continued decline to 48.

The US Democracy Threat Index, a Bright Line Watch creation hosted on Metaculus that tracks 39 specific questions about democratic erosion, currently sits at 39.7/100, but this metric tracks threats and trajectory rather than democracy health. The expert panel provides a comparative snapshot (the US currently functions more like a mixed regime than a full democracy). The Threat Index instead captures the rate and direction of erosion: specific guardrails being dismantled, active violations, and institutional stress tests. Notably, the Threat Index questions were designed by experts, eliminating some of the confounding resolution criteria problems that plague the prediction markets tracked above.

The expert opinions on democracy erosion captured by Bright Line Watch should hold more weight for EAs than prediction markets for three reasons:
Domain expertise. Political scientists studying democratic backsliding track comparative cases (eg, Hungary, Poland, Turkey, India), and have frameworks for recognizing early-stage erosion patterns that prediction market users lack.
No financial incentive distortion. Experts’ assessments are not subject to the opportunity cost dynamics that cripple long-horizon markets. Their opinions reflect genuine well-informed belief rather than risk-adjusted self-interested financial decisions.
Lagging vs. leading indicators. Expert ratings are updated frequently in response to events. The Bright Line Watch 14-point drop is a leading indicator of institutional erosion that prediction markets, for the structural reasons described above, are too sluggish to register.

BLW Limitations: Bright Line Watch solicits opinions from a population of political scientists that has an 8.5:1 ratio of Democrats to Republicans. Still, BLW’s democracy ratings showed real variation under Biden, suggesting that partisanship plays a minor role in the expert forecasts, and the questions BLW tracks (executive constraints, press freedom, agency independence, etc) are procedural, not policy-based, so experts can rate them badly without it being straightforwardly partisan.

Conclusion:
The key takeaway from the polls we examined is this: EAs who are evaluating democracy as a cause area should avoid looking to prediction markets for evidence of democratic decline. Prediction markets are useful for short-horizon electoral questions, but are structurally unable to price the slow, incremental erosion that characterizes democratic backsliding. EAs should start instead with expert opinion and structured indices (Bright Line Watch, V-Dem, etc) to set their priors on the topic of democratic erosion in the US.
Of the three platforms hosting the markets tracked in this post, only Polymarket has taken direct action on the long-horizon problem, introducing a 4% annualized yield on long-term political and geopolitical positions in September 2025. Manifold’s play-money structure and Metaculus’s reputation-based system have no equivalent mechanism to address opportunity cost mispricing, and given their design, likely never will.
The fact that a sizeable minority of prediction market users are willing to stake their money or reputations on an authoritarian takeover of the US should concern any EA, but the market’s lack of responsiveness to contemporary events is an artifact of financial incentives overruling the tendency of prediction markets to forecast true outcomes. Long-horizon forecasts in particular lose much of their value after the first month or so, and only retain any explanatory power inasmuch as they reflect expert opinion, due to their pervasive mispricing.