Settlement Manipulation in Prediction Markets

Abstract: Prediction markets increasingly list contracts settling on an asset price that holders can move by trading the underlying. We build a model showing that such contracts transfer wealth from prediction-market liquidity traders to manipulators and harm price discovery in the underlying, even as it becomes more liquid. After the launch of Polymarket's five-minute Bitcoin contract, settlement-time spot order flow spikes, causing large price reversals after settlement. Manipulators capture a large amount of profit, mostly from retail. Manipulation is largely absent in the fifteen-minute contracts: lengthening the contract horizon removes it, providing the market-design remedy our model and evidence support.

• The paper combines a sequential game-theoretic model with transaction-level evidence from more than 60 million Polymarket Bitcoin fills to show that manipulators can profit by pushing spot prices across settlement strikes.
• Empirical results show a 50% near-settlement order-flow increase, post-close price reversals, stronger effects near-the-money, and manipulation losses of $7.6 million for retail traders versus $8.2 million in manipulator gains.
• The findings indicate that longer contract horizons, averaged settlement references, and closing auctions can reduce manipulation incentives by making last-second price pushes less likely to determine outcomes.

Overview and contribution

This paper studies a structural vulnerability in the fastest-growing class of prediction-market contracts: binary "up/down" contracts that settle on an asset's price at a fixed close. Because the settlement reference is itself a price that traders can move by trading the underlying, a holder of the contract can profit by pushing the spot price across the strike in the final seconds before settlement — the cash-settlement manipulation mechanism of Kumar and Seppi, now operating inside retail prediction markets with horizons as short as five minutes. The paper combines a sequential game-theoretic model with transaction-level evidence from Polymarket's five-minute Bitcoin contract to show two results: such contracts degrade price discovery in the underlying spot market, and they transfer wealth from ordinary traders to a small cohort of manipulators. The authors state this directly against the prevailing view, inherited from Hanson's manipulator result and documented manipulation episodes in political markets, that manipulation need not impair prediction-market accuracy: for asset-price contracts, they argue, it does.

The institutional setting matters. Polymarket's crypto up/down contracts resolve against a Chainlink oracle that aggregates prices across exchanges rather than any single venue. The paper validates Binance as a proxy for this benchmark: the Binance mid sits about 2.5 basis points from the Chainlink print, moves essentially one-for-one within each window, and finishes on the same side of the strike as the resolution about 85% of the time at the five-minute horizon. A push of a few basis points past the strike therefore reliably carries the resolution — and median five-minute spot moves are only about 6.6 basis points, so a modest push is often decisive.

Theoretical framework

The model extends Glosten–Milgrom sequential trading to two linked venues. An informed trader and liquidity traders populate the spot market; a risk-neutral competitive market maker quotes break-even prices on each venue; and a manipulator may buy the binary contract and then trade spot in the settlement window. The contract settles on the spot market's posterior mean of the fundamental value plus small noise, which captures the fact that the oracle aggregates venues the manipulator does not fully control. Under moderate adverse selection ($\varepsilon < \sigma\pi < 1/2$), the equilibrium has three distinctive features.

First, price discovery strictly falls whenever the manipulator participates with positive probability, through two channels: displacement (she occupies the settlement-window slot, thinning informed flow) and dilution (her uninformed buys pool with informed ones, so a buy reveals less). This is the mirror image of Zhu's dark-pool result: there, a second venue concentrates informed flow and sharpens the exchange price; here, the linked contract injects uninformed, settlement-motivated flow into the price-setting market.

Second, the prediction-market maker earns a rent at a zero spread. The break-even ask falls below $1/2$ because the manipulator's contract is worth more than a liquidity trader's, but the non-crossing constraint floors the posted price at $1/2$, leaving the maker a strictly positive rent per contract sold despite quoting no spread. The paper derives an exact accounting identity: the uninformed liquidity trader's expected loss funds the maker's rent, the manipulator's profit, and the manipulator's spot-push cost. Notably, the harm to the liquidity trader operates cross-market: holding the "up" contract makes a settlement-window spot buy look possibly manipulative, so the spot maker lowers its ask while the bid is unchanged, biasing the settlement against the up side even though the holder never trades spot.

Third, the horizon is a design lever. With $n$ ordinary trading rounds before settlement, the pre-settlement order imbalance follows a $\pm 1$ random walk whose dispersion grows as $\sqrt{n}$ while the band of imbalances where a push is pivotal stays fixed width. The pivotal probability, the manipulator's conditional profit, and her unconditional expected profit all vanish as $n$ grows. Entry probability oscillates with parity of $n$, but the paper shows this oscillation is irrelevant to safety; what governs safety is how often a push can be pivotal at all. Aggregating settlement references (window averages, closing auctions) supplies a parallel lever, and the paper notes that the binary Nasdaq-100 index options approved in April 2026 sit on the safe side of both levers — long expirations settling on the Nasdaq Closing Cross.

Empirical design and footprint of manipulation

The empirical strategy exploits the staggered rollout of Polymarket's Bitcoin contracts: a pre-launch baseline (July–October 2025), an intermediate period with fifteen-minute and four-hour contracts only, and a post-launch period beginning February 12, 2026 when the five-minute contract went live. The sample covers over 60 million on-chain fills on the five-minute BTC contract alone, linked to Binance tick data.

Four parts of the predicted footprint appear, and only after launch:

• Near-settlement activity spike: final-10s absolute order flow rises about 50% relative to the cycle body versus the pre-launch baseline (log coefficient 0.406), with absolute return up 15% and depth up 5%, all significant at 1%.
• Post-settlement reversal: the first post-close ten-second return reverts against the final-10s return with slope $-0.10$ once the contract is live, against a baseline indistinguishable from zero — inconsistent with information-driven price impact.
• Moneyness concentration: in the roughly 6% of cycles still near-the-money (Up-token price within 0.10 of 0.5) ten seconds before close, the order-flow spike is about 3.9 times that in decided cycles, and reversal deepens to roughly 25% versus 10%. Liquidity diverges exactly as the model predicts: Binance depth rises while Polymarket depth falls sharply (log coefficient $-1.02$), as binary makers withdraw from exposure to a flipping settlement.
• Horizon attenuation: the same difference-in-differences applied to the fifteen-minute contract yields an order-flow increment of only 0.068 (versus 0.406) and no detectable reversal, consistent with the model's prediction that longer horizons shrink rather than eliminate the incentive.

Who wins, who loses, and why it is not hedging

Using a within-cycle push-intensity statistic built solely from Binance order flow, the paper flags the top decile of cycles (~1,600) as manipulated. These cycles cluster in thin-liquidity windows (56% overnight, 44% weekends) where a dollar of flow moves the resolution most, and are 4.2 times more likely than normal cycles to be near-the-money — a concentration that cannot be mechanical because the classifier uses no outcome or price information.

The pushes change outcomes. In cycles where Polymarket priced one side at 90–100% just before the close, a push against the favored side reversed the resolution 34.2% of the time, against 1.0% in unpushed cycles. Wallet-level PnL attribution identifies 821 manipulators (0.34% of roughly 243,000 traders) who earn +$1/2$07.6M there, bear 70.5% of the manipulated-cycle loss pool, and sit on the losing side in 65% of manipulated cycles versus 48% of normal ones. Market makers largely avoid the transfer (−$0.6M), consistent with their book withdrawal. These attributions are robust across a wide grid of cohort-definition thresholds.

The paper rules out the leading innocent explanation — gamma hedging by prediction-market makers — on three grounds. In the near-certain cycles a push most often flips, the binary's delta has collapsed and there is essentially nothing to hedge. Timing contradicts dynamic hedging: Polymarket positions accumulate linearly through the body while cumulative Binance flow is approximately zero until the final minute, when ~$1/2$11,400 of risk per cycle, three orders of magnitude below the spot trade the hedging story would attribute to it.

Limitations and open questions

Several caveats bear directly on the strength of the conclusions. Manipulator identification rests on realized profitability in flagged cycles plus participation thresholds, not direct observation of intent; the paper acknowledges a trader could in principle qualify by luck or unrelated information, though the break-even performance of the cohort in normal cycles mitigates this. PnL is gross of exchange fees, which changed mid-sample and peak precisely at even odds where these contracts spend much of their lives; net-of-fee accounting is left open. The Binance-as-proxy validation shows strike-side agreement of only 84.8% at the five-minute horizon, so some measured pushes fail to carry the resolution, and the both-NTM overlap interaction rests on very few cycles (11 for activity, 9 for reversal). The theoretical model assumes a single settlement-window slot, binary fundamentals, and moderate adverse selection; whether its quantitative welfare decomposition extends to richer settings is not established. Finally, the paper documents one contract on one platform over two months; whether the footprint generalizes to other coins, to Kalshi's centrally cleared fifteen-minute contract, or to the equity-index binaries now pending before the SEC remains an open empirical question.

Conclusion

The paper establishes that ultra-short, oracle-settled asset-price prediction contracts create a settlement-manipulation channel that degrades spot price discovery and transfers wealth from retail traders to a small manipulator cohort, while paradoxically improving spot liquidity. Both theory and data point to contract horizon as the operative design lever: lengthening it renders a feasible push almost never pivotal, and aggregating or randomizing the settlement reference offers complementary protection. The regulatory implication is specific — the "not readily susceptible to manipulation" standard is best met through settlement design, since manipulation here is a property of the settlement rule rather than of trader intent.