Market-Making Research on Polymarket
Incentives, Risks, and Mitigation
Erin Li · McGill Desautels
Introduction
Polymarket operates as a hybrid prediction market that combines an off-chain central limit order book (CLOB) for fast order matching with on-chain settlement on Polygon for final execution. Liquidity in this system is primarily provided by market makers, who continuously post bid and ask orders on the order book. In practice, these market makers are highly algorithmic, relying on bots to actively update quotes and manage inventory in real time. Their role is to ensure that traders can enter and exit positions efficiently, even in markets driven by uncertain and rapidly evolving information. However, the binary outcome structure of prediction markets and the platform's reward design create a set of incentives and risk exposures that distinguish market making on Polymarket from traditional venues.
Market Maker Reward Structure
Market makers on Polymarket earn revenue through three channels: the bid-ask spread, maker rebates, and liquidity rewards.
1. Bid-Ask Spread
The spread is the most fundamental source of market-making revenue. By posting a bid below and an ask above fair value, a market maker earns the difference on each completed round-trip as order flow arrives from both sides.
2. Maker Rebates
Beyond spread capture, market makers also earn rebates when their resting limit orders are filled by incoming takers. Beginning March 30, 2026, Polymarket implemented a revised fee structure across all markets using the formula:
where C is the number of shares traded, p is the transaction price, and feeRate and exponent vary by market. The fee curve is symmetric and peaks when the market is near 50% probability, declining toward zero as prices approach either extreme — meaning trades near the most uncertain outcomes carry the highest fees. Approximately 20–25% of collected taker fees are redistributed daily to market makers using the same formula, so rebate flow is also largest in active, competitive near-50% markets and diminishes as markets approach resolution.
3. Liquidity Rewards
In addition to filled-order earnings, Polymarket distributes approximately $12 million annually in liquidity rewards to market makers with resting orders on the order book — rewards that can be earned without any orders being executed. Distribution is determined by a scoring system summarized by the following parameters:
- Maximum incentivized spread — quotes must sit within a defined distance from the midpoint (e.g., ±4 cents) to qualify.
- Minimum order size — a size threshold filters out token quotes.
- Two-sided score — gives roughly 3× the rewards of single-sided quoting; in extreme markets where p < 0.1 or p > 0.9, liquidity must be double-sided to be eligible.
- Reward pool share — each market has a daily reward pool, and a maker's share is proportional to their contribution relative to all other qualifying quotes.
Some markets also include sponsored reward pools funded by users, further increasing participation incentives.
This taker-fee and maker-reward structure shares properties with liquidity incentive mechanisms in decentralized exchanges. Hasbrouck, Rivera, and Saleh (2022) show that higher trading fees on a DEX can increase trading volume and improve liquidity by incentivizing inventory provision; the additional inventory reduces price impact, which lowers the marginal cost of trading and can offset the higher fees. While Polymarket uses an order book rather than an AMM, a similar economic logic applies: because fees are highest near the midpoint where uncertainty is greatest, they attract more aggressive quoting precisely where liquidity is most needed, improving depth and tightening spreads in the most active regions of the book. However, the same conditions that reward liquidity provision also create unique risks.
Risk Structure
The binary outcome structure of prediction markets, combined with a reward design that heavily incentivizes liquidity close to the midpoint, creates risk exposures unique to market makers on Polymarket.
1. Inventory Risk Under Terminal Binary Settlement
In equity markets, inventory risk is manageable because positions have continuous terminal values and can be hedged against the underlying asset. In prediction markets, neither applies. Every position on Polymarket resolves to exactly $0 or $1.00, and there is no underlying asset for hedging. This transforms inventory accumulation from a manageable inconvenience into a hard financial exposure with a defined worst-case loss at resolution.
2. Adverse Selection and Information Asymmetry
Adverse selection is the risk that a maker's quote is executed against a counterparty with superior information about the asset's true value; the bid-ask spread exists to compensate for this informational disadvantage. In the Glosten-Milgrom (1985) framework, the zero-profit ask equals the expected value conditional on a buy order arriving, with μ representing the fraction of informed traders (between 0 and 1). The bid and ask are:
In other words, the spread is essentially the probability that any incoming order comes from someone with superior information. Makers therefore want μ to stay small so they can quote tight spreads profitably. On Polymarket, μ starts relatively low early in a market but rises sharply as resolution approaches; near resolution μ effectively approaches 1, almost every order becomes informed, and tight spreads become extremely dangerous.
Kyle (1985) complements this by showing how informed trading creates price impact. The key parameter is Kyle's λ, which measures how much each unit of net order flow moves the price:
Higher λ means prices move more in response to a given order, reflecting lower liquidity and stronger adverse selection. On Polymarket, λ spikes near resolution as casual uninformed traders exit and the remaining flow becomes increasingly one-sided and informed. When both μ and λ are elevated in the final hours before settlement, adverse selection risk reaches its peak.
3. Event-Driven Jump Risk
Standard market-making models assume prices follow continuous diffusion. The Avellaneda-Stoikov (2008) framework, used to determine reservation prices and optimal spreads, assumes the mid-price evolves smoothly and can move freely without boundaries, so risk is driven by gradual, small price changes. Prediction markets behave fundamentally differently: contract prices are strictly bounded between 0 and 1 and must resolve to a discrete 0 or 1. Volatility does not decline steadily toward expiry — it undergoes a sharp phase transition as resolution approaches — and price movements are often driven by sudden, discontinuous jumps triggered by new information rather than smooth diffusion. This mismatch creates a distinct event-driven jump risk that is structural rather than statistical: a maker holding inventory at the moment a high-impact event resolves faces significant potential losses, and traditional reservation-price models can significantly understate true risk exposure near resolution.
Risk Mitigation
Polymarket's reward design creates attractive economics for liquidity providers, but the unique microstructure — terminal binary settlement, no delta-hedging, and extreme adverse selection near resolution — generates risks that require active, sophisticated mitigation. Makers must simultaneously manage downside exposure while positioning to capture spread, rebate, and liquidity-reward revenue.
1. Inventory Management
Because prediction markets allow no delta-hedging, inventory risk must be managed entirely through position constraints and quote adjustment. The Guéant, Lehalle & Fernandez-Tapia (2013) (GLFT) framework extends Avellaneda–Stoikov by introducing explicit inventory bounds |q| ≤ Q, where Q is the maker's maximum tolerable loss on a single contract, and derives closed-form optimal quotes. Optimal bid and ask quotes consist of two components: a symmetric baseline spread plus an inventory-dependent adjustment. As current inventory q grows, the (Q − q) term shrinks on the ask side, tightening the ask to encourage selling, while (Q + q) grows on the bid side, widening the bid to discourage further buying. When q reaches Q, quotes effectively disappear from the book, preventing further inventory accumulation. In an environment where contracts can resolve to a total loss with no hedge available, position limits become a hard constraint; combining dynamic inventory skewing with strict position limits provides effective control.
Directional exposure can also be partially reduced through cross-market conditional hedging — taking offsetting positions in correlated markets (e.g., hedging a state-level outcome against the national election using implied conditional probabilities) to lower overall portfolio delta without fully flattening any single market.
2. Adverse Selection Detection and Response
Adverse selection in prediction markets is discontinuous and intensifies with new information or as resolution approaches, so makers must actively detect and respond to elevated information asymmetry. Easley, López de Prado & O'Hara (2012) introduced the Volume-Synchronized Probability of Informed Trading (VPIN), a real-time estimate of the proportion of toxic (informed) flow:
where \(V_\tau^{S}\) and \(V_\tau^{B}\) are sell and buy volume in each bucket and \(V\) is the bucket size. VPIN measures buy/sell imbalance over a rolling window; a higher value signals increasingly one-sided, toxic flow. When VPIN rises beyond a threshold, makers can widen spreads or temporarily withdraw liquidity to avoid being adversely selected. Spikes in Kyle's λ serve as an additional early-warning signal. Near resolution, staged withdrawal — gradually reducing quote size and widening spreads, then fully removing liquidity in the final stages — reflects the reality that as μ approaches 1, the Glosten-Milgrom adverse-selection cost dominates all other considerations.
3. Event-Driven Risk Prevention
To capture event-driven dynamics, Dalen (2025) proposes a logit jump-diffusion framework:
The drift term \(\mu(P_t)\,dt\) captures gradual belief updating; the Brownian term \(\sigma\,dW_t\) captures small day-to-day fluctuations; and the jump term \(J\,dN_t\) explicitly models sudden, discontinuous moves triggered by news, polls, rulings, or other high-impact events. The model recognizes that prediction-market prices move in two regimes — gradual belief updating and sudden event-driven jumps. In practice, makers complement this with pre-event risk management: Good-Till-Date (GTD) orders automatically expire quotes before known events, preventing stale orders from being adversely filled and reducing exposure to jumps that standard reservation-price models fail to anticipate.
Conclusion
Polymarket's hybrid microstructure and multi-layered reward structure — bid-ask spreads, maker rebates, and liquidity rewards — create meaningful incentives for makers to supply continuous, tight, two-sided quotes. But the same features that make liquidity provision attractive also introduce risks standard frameworks are not designed to handle: no underlying asset eliminates delta-hedging, binary terminal settlement creates hard inventory constraints with defined worst-case losses, adverse selection intensifies discontinuously near resolution, and event-driven jumps violate continuous-diffusion assumptions. Addressing these requires a layered approach: GLFT bounded inventory control with dynamic skew, real-time toxicity detection via VPIN, staged withdrawal near resolution, and logit jump-diffusion pricing paired with pre-event order management. Ultimately, the core challenge is not managing risk or capturing reward in isolation, but navigating the tension between the two — the reward structure pushes makers toward tight, deep, two-sided quotes near the midpoint, precisely where adverse-selection and jump risk are greatest. Sustained profitability depends on dynamically balancing this trade-off: remaining competitive under normal conditions while withdrawing quickly and decisively when conditions turn informationally adverse.
References
- Avellaneda, M., & Stoikov, S. (2008). High-frequency trading in a limit order book. Quantitative Finance, 8(3), 217–224.
- Dalen, S. (2025). Toward Black-Scholes for prediction markets: A unified kernel and market maker's handbook. arXiv:2510.15205.
- Easley, D., López de Prado, M., & O'Hara, M. (2012). Flow toxicity and liquidity in a high-frequency world. The Review of Financial Studies, 25(5), 1457–1493.
- Glosten, L. R., & Milgrom, P. R. (1985). Bid, ask and transaction prices in a specialist market with heterogeneously informed traders. Journal of Financial Economics, 14(1), 71–100.
- Guéant, O., Lehalle, C. A., & Fernandez-Tapia, J. (2013). Dealing with the inventory risk: a solution to the market making problem. Mathematics and Financial Economics, 7(4), 477–507.
- Hasbrouck, J., Rivera, T. J., & Saleh, F. (2022). The need for fees at a DEX: How increases in fees can increase DEX trading volume. SSRN 4192925.
- Kyle, A. S. (1985). Continuous auctions and insider trading. Econometrica, 53(6), 1315–1335.
- Polymarket. (2026). Market Makers. Polymarket Documentation. https://docs.polymarket.com
