Between the outcome and the payout
We measure resolution risk on Polymarket and forecast it in public
10bips team, imagined in Poland
Abstract. Before a Polymarket market pays, a whitelisted proposer reports what happened, and anyone can dispute the report. Disputes are rare per proposal, yet their number more than doubled in the past year, and more of them now land on markets that settle within minutes. When one lands, the payout often changes, and a dispute that goes to a vote holds payment for days. We measure each of these steps from public on-chain data, publish what we find, and test our forecasts in the open, misses included.
i. The problem
An event contract pays only once someone decides what happened. That decision can be disputed, reversed or paid late. 10bips measures that risk, proposal by proposal, from public data.
Near its end, a winning position is the safest asset in the market: the outcome is public and the price sits close to $1. What is left is settlement, and it is a growing problem. Polymarket lists far more markets than a year ago, most proposals now come from a handful of automated proposers, and many markets give only minutes to dispute. Disputes are rarer per proposal, but there are more than twice as many, and one can knock the price down and hold the payment for days. The price near $1 mixes this risk with time value, fees and spread, and cannot separate it.
This now has a cost. Venues and lenders are adding margin and credit to prediction markets, and a lender has to value a position exactly when it nears settlement. Without a price for resolution risk, the safe answer is to stop lending as settlement approaches, so capital that could be working waits until the market pays.
ii. Resolution, live
A Polymarket market resolves in three steps. A whitelisted proposer reports the outcome and posts a bond. A dispute window opens, from ten minutes on fast sports and crypto markets to two hours on the rest, in which anyone can dispute by posting a bond of their own. If no one does, the market settles. Table 1 shows these steps as they happen, read from public on-chain data.
| time (utc) | event | market | proposed | proposer | dispute window |
|---|
Table 1. Proposals, disputes and settlements in the last 6 h. From October, each row carries its resolution risk.
No proposals are in the dispute window right now.
Figure 1. A steady sample of the proposals now inside their dispute window, newest at the top. The bead marks how much of each window has passed. A line turns oxblood when the proposal is disputed.
iii. What we measure
The probability that a proposal is disputed before its window closes. Estimated when the proposal is made, from information public at that moment.
The expected loss to a holder of the proposed outcome, in basis points of notional: the probability of a dispute times its cost, which is a possible change in payout plus the wait until it is paid.
The discount a lender should apply to a position before it settles. Resolution risk is an average; the haircut is set by the bad cases, how far a position can fall when a dispute goes wrong.
iv. Method
Resolution risk is an expected loss. Follow one proposal from start to payment; each step adds a term to equation (1).
RR = 10⁴ · · [ + ](1)
Figure 2. One illustrative market, from proposal to payment. The line is the price of the proposed side. Illustrative, not market data.
D is a dispute and Δ a change in payout. L is the loss per $1 if the payout changes. τ is the time from dispute to settlement, and r the yearly cost of the capital locked during it. The factor 10⁴ turns the result into basis points.
Polymarket resolves markets through UMA's optimistic oracle, in stages that are easy to mistake for a single event [1]. Since September 2025 only whitelisted addresses can propose [6]. A proposer posts a bond and an outcome. If the window closes without a dispute, the outcome is final. A first dispute sends the question back for a new proposal and a new window. A second dispute goes to a vote of UMA token holders, which follows 1 to 2 days of public debate [5] and returns one of four answers: the proposer was right, the disputer was right, the question was asked too early, or the payout is split 50/50. Every path after a dispute changes what the market pays, when it pays, or both.
Estimating each term
A statistical model trained on past proposals and whether they were disputed. It uses only what is public when the proposal is made: how often the proposer's earlier proposals were disputed and how many they have made, how soon after the question reached the oracle the proposal came, and what was proposed. An LLM reading the rules barely beats a coin at predicting disputes [4], which is why the model leans on who proposes and when.
The share of past disputes, by type of market, in which the final payout differed from the first proposal.
For a holder to payment: $1 per share if the other side wins, $0.50 if the market splits 50/50. For a lender, the gap between the collateral value and what the position finally pays.
The expected time from dispute to settlement, measured on past disputes along each path in Figure 2, steps 3 to 5.
The cost of locked capital, set by whoever uses the number; 5% a year by default. Markets already discount near-certain contracts for this lock-up [2].
Not an average. Taken from the distribution of how far the proposed side fell after past disputes, by price level, market size and risk score, at a high percentile.
How it is checked
- 1Out of time. The model is trained on older proposals and tested only on newer ones it has never seen, the way it is used live.
- 2On the rare cases. Disputes are rare, so a model that always answers "unlikely" looks accurate and is useless. We judge it by how many real disputes fall among the proposals it ranks as riskiest.
- 3Frozen daily. Each day's forecasts are written to a file, and its SHA-256 hash is published when the day closes, so no forecast can be changed afterwards.
- 4Misses included. Every dispute is compared with the number published before it, and misses are published in the same format as hits.
Limits
- 1Once a dispute is filed, traders price the likely verdict of the vote [3], and the market price becomes the best forecast of the verdict within minutes. Our number covers the period before a dispute; the verdict itself is left to the market.
- 2The rarest cases are the hardest to estimate. We publish how well the number holds at each level of risk, including where it holds least.
References
- [1] M. Nechepurenko (2026). Resolution Is Not Settlement, Part I: Oracle Adjudication and Semantic Governance on Polymarket. arXiv:2609.15368.
- [2] J. Gebele, F. Matthes (2026). When Certainty Is Not Worth It: Capital Lock-Up and Settlement Discounting in Prediction Markets. arXiv:2605.31431.
- [3] G. Wang (2026). Do prediction markets price events or adjudication? Evidence from the disputed Polymarket markets. Economics Letters 268.
- [4] J. Wen, J. Zhou, J. Huang (2026). Can LLMs Help Decentralized Dispute Arbitration? A Case Study of UMA-Resolved Markets on Polymarket. arXiv:2604.15674.
- [5] Polymarket documentation. Resolution. docs.polymarket.com/concepts/resolution.
- [6] UMA (2025). Managed Proposers Update. blog.uma.xyz/articles/managed-proposers-update.
Full model notes on request: contact@10bips.com