Self-resolving Prediction Markets

Have thoughts on this topic? Join the conversation on X.

Introduction

Current prediction market design is most effective for events with objective, verifiable outcomes. They resolve and reward participants based on known future results. Whether the resolution system uses an external oracle, AI-based analysis, or human judges, the underlying model remains the same: the market resolves based on clearly observable evidence.

While controversies and disputes occasionally arise due to poorly written rules, these markets largely function as intended.

However, beyond verifiable events, people also enjoy making predictions about topics that may never be officially resolved. These can include the long-term impact of a policy, controversial issues where a global consensus is impossible, or subjective topics like public sentiment and cultural trends.

Can a prediction market exist for these inherently subjective and unverifiable outcomes? This is the question explored by Siddarth Srinivasan, Ezra Karger, and Yiling Chen (SKC) in their paper, "Self-Resolving Prediction Markets for Unverifiable Outcomes."

The paper’s key idea is that instead of waiting for an actual event to occur, the market resolves itself by using the crowd's final consensus as the outcome. An oversimplified explanation of the process is as follows:

• Users participate in the market sequentially, each updating a shared forecast.
• After each prediction, the process randomly stops with a probability, α.
• When it stops, the last participant's prediction is treated as the final outcome.
• Earlier participants are paid based on how much they moved the market toward this final outcome, considering both the direction and degree of change.
• The last few participants, including the final predictor, receive a fixed reward instead of a score-based payment to avoid edge-case gaming.

The reasoning behind this mechanism is that it requires no external resolution. Instead, it self-resolves by using its most informed prediction (i.e. the final one) as a proxy for truth.

• Overview of the mechanism
• Core elements
• More on the payoff structure
• Conclusion
• References

Overview of the mechanism

The SKC paper is dense with mathematical proofs and admittedly stretches beyond my full comprehension. But its core fascination is a single, exciting question:

• How can we design a system where the act of truthfully sharing information becomes the most rational and rewarding strategy, even when no objective "right answer" will ever exist?

Solving this would be a big deal. It would extend the power of prediction markets (i.e. aggregating diverse, private knowledge) to a much more expansive domain of unverifiable but crucial questions, from policy impacts to cultural sentiments.

Let’s explore the mechanism in more detail:

• A market is created for a non-verifiable question, such as “How likely is it that Policy X will ultimately reduce crime?” or “What is the probability that Theory Y is correct?”
• Traders participate one after another. Each enters their probability that the statement is true, which updates the shared forecast.
• The market would close at a random time and use the final predicted probability from the last user as the resolution.
• The market closes at a random time. The final probability stated by the last participant is treated as the "reference forecast". Because this last participant has seen all previous forecasts and possesses their own private information, their view is theoretically the most informed.
• Participants (except the last few) are paid based on how closely their forecast aligns with the final reference forecast, using a proper scoring rule like negative cross-entropy. This rewards them for moving the market toward what becomes the collective best estimate.
• The last k participants receive a fixed reward. This is an anti-gaming measure that removes the incentive to withhold information strategically in order to be the final, decisive "reference voice."

The core incentive is designed so that the most profitable strategy is to honestly report one's best estimate given all available information. Participants are ultimately scored against the market's culminating wisdom.

The paper goes into details to prove that the mechanism can work under economic assumptions similar to normal markets (e.g. rational, self-interested participants with access to private information) but now applied to settings with no verifiable ground truth.

Core elements

If this still sounds very confusing, let’s focus on building an intuition of the core elements.

Sequential Reporting with Random Stop

Agents (as they’re called in the paper) enter one by one to report probabilities. After each agent’s report, you flip a weighted coin with probability α of stopping. If the coin comes up “stop”, the market ends. Otherwise, you proceed to the next agent.

This random termination scheme makes the length of the game unpredictable. No one can be sure they’re the final trader or how many more will come after them, which, as discussed in the paper, helps keep everyone honest.

Final Agent as Reference Outcome

When the process stops at some agent T, that agent’s reported number - call it q(T) - is declared the final market prediction and the outcome of the market. This final agent would have seen all prior forecasts. They presumably used all that information plus their own signal to make the best possible prediction. Thus, q(T) is the most informed consensus and serves as the stand-in for the true answer. For example: “Final market prediction: 82%.”

Payoff Calculation (Cross-Entropy Scoring)

All participants except the last k agents are rewarded based on how close their reported probability was to q(T).

The specific rule is the negative cross-entropy market scoring rule. In plain English, they compute what your score would be if the final prediction q(T) was the actual truth, and you had predicted q(T). Importantly, the payoff is not based on your final absolute closeness to the reference number, it’s based on your marginal improvement toward it.

The system measures a score against the final reference (the yardstick). What you get paid is the increase in that score caused by your move (after you vs. before you). This is the “delta” in the cross-entropy market-scoring rule. Your payoff is tied to how much you helped move the needle toward what the final outcome ended up being.

Flat Reward for Last k Agents

The last k agents (including the final one) each receive a fixed reward R, independent of their predictions. This flat reward is basically a participation reward. It solves a few issues:

• It removes any incentive for those late agents to try to manipulate the final outcome. For instance, if you’re the second-to-last agent (T-1), you know the next agent (T) will be final. If you were going to be scored on how close you are to T, you might think “Maybe I should skew my report to influence T’s report.” But because you’re instead just getting a fixed payout, you have no financial stake in what T does.
• It also gives the final agent a reward even though no one scores them. This ensures people are willing to be the final predictor. If the final agent got nothing (or had to pay others without reward), nobody would want to go last. By giving them R, you at least compensate them for their time and information. And since it’s flat, the final agent can’t improve it by lying or telling the truth – but we assume they’ll tell the truth because there’s no incentive to do otherwise and they might as well maintain consistency/honesty.

More on the payoff structure

At first, the payoff structure may be surprising. Normally, you'd assume that being close to the final outcome means you were "right" and should be rewarded for your accuracy.

But that's not how this works. If the market probability is already at 80% and you simply report the same number, you haven't actually contributed any new information to move the probability toward its eventual conclusion.

The mechanism pays you for information you add, not for how close you end up to the crowd once it’s finished. Formally, each trader’s payoff is:

• score(final forecast vs. old price) – score(final forecast vs. your new price).

That’s the “difference-in-score” logic of a market-scoring rule. More simply:

• If you shift the market toward where it finally settles, you’re rewarded for that push.
• If you barely move the needle, your contribution—and thus your reward—is small.
• If you move the price in the wrong direction, you effectively pay a penalty.

So this means an early trader who nudges the number from 40 → 50 % (a 10-point improvement) gets more than a late trader who nudges 60 → 61 %, even if the late trader’s number is “more accurate” in an absolute sense.

This design choice prevents free-riding. Without it, someone could wait until the market has converged, copy that near-final number, and collect a prize for “accuracy” without adding insight. The delta-based rule rewards marginal contribution instead.

This also addresses the Keynesian beauty contest problem. Keynesian beauty contest refers to the behaviour where participants anchor their beliefs based on what others will think, instead of their own true belief.

You should by now realize that in this mechanism you don’t get paid for matching what you think others will say. Instead you get paid for how much your move pushes the market toward where it ultimately ends. It rewards marginal information you add on your turn, which makes sincere reporting the winning play.

All the design elements here (the α and k settings especially) are to deter dishonesty. The goal is that for each participant, if they think it through, to realize: “Whatever I report, at the end I’ll be judged against where the market ends up. I can’t really sway the end much alone because others will come after me. So the best I can do is make the forecast as accurate as possible given what I know right now.”

Conclusion

A core problem with existing prediction markets is their reliance on an ultimate oracle, which often is not available. Many questions that are subjective would still be extremely valuable to resolve, such as:

• “How safe does new drug X seem right now given today’s evidence?”
• “How likely is policy Y to reduce unemployment in the next year?”
• “Is this research claim credible yet?”

Or culture-vibe debates like:

• “Is Vision Pro a flop so far?”
• “Who wins: 1 gorilla or 100 men?”

Or:

• “By 2030, will historians describe today’s AI safety concerns as overblown?”
• “Will quantum computing be viewed as transformative by 2040 textbooks?”
• “Will ETH be considered the ‘world reserve settlement asset’ in a 2030 IMF paper?”

The mechanism proposed by SKC is an exciting new way to resolve these questions. We’d love to see teams prototype it on crypto-native rails! If it clicks, the world gets a completely new way to settle uncertain questions, allowing for precision and efficiency in decision-making.

If you are interested in the topic, please reach out on X.

References

• Self-Resolving Prediction Markets for Unverifiable Outcomes Siddarth Srinivasan, Ezra Karger, and Yiling Chen