Polymarket arbitrage bots look for complete sets that cost less than their settlement value: a $0.46 YES ask plus a $0.51 NO ask returns $100 for $97 per 100 shares. The gap counts only if both legs fill and it covers fees; a model-versus-price gap alone is a directional signal. To run a bot near the book, QuantVPS offers a Polymarket VPS in Dublin, under 1ms to Polymarket's order book.
Keep your strategy running while you're away.
How Polymarket market structure creates arbitrage candidates
Outcome tokens trade in a central limit order book, with off-chain matching and Polygon settlement. Quotes for different outcomes can become temporarily inconsistent, but displayed prices alone do not show whether enough size is available to complete a position (see prediction market liquidity and spreads). To compare distance to the book by location, see trading VPS latency by city and venue.
Complete-set and related-market arbitrage
For a binary market, compare the executable cost of a compatible YES and NO set with its combined settlement value. A displayed total below one dollar is only a candidate: fees, available depth and incomplete execution can remove the difference.
Across related contracts, build an explicit payoff table. Outcomes must have the same event definitions and cover the intended states without gaps or double counting. Similar titles are insufficient when one market resolves on a different date, source or condition.
Buying a high-probability outcome near resolution remains a directional trade if the adverse outcome is still possible. It should not be presented as a risk-free tail-end arbitrage strategy.
For a position involving a second venue, use the cross-market arbitrage guide. That workflow has additional settlement, currency and leg-execution differences.
Worked complete-set calculation
Consider a hypothetical compatible YES and NO pair in the same binary market. An executable YES ask is $0.46 and a NO ask is $0.51, each with 100 shares available. Buying 100 of each costs $97 before fees. Assuming the normal binary payout relationship holds, the complete set returns $100 across the two outcomes, leaving a $3 gross difference.
The difference must cover both legs' fees and any additional execution or settlement costs. If their combined cost is $3 or more, there is no positive net spread at those prices. If only 40 NO shares fill after all 100 YES shares execute, the position consists of 40 complete sets and 60 unmatched YES shares. A balanced payoff calculation for 100 sets no longer describes the actual inventory.
| State after submission | Matched sets | Residual exposure | Required response |
|---|---|---|---|
| Both legs fill 100 shares | 100 | None from this pair | Verify fees and positions |
| YES fills 100, NO fills 40 | 40 | 60 YES shares | Apply the preplanned residual limit |
| Only NO fills 100 | 0 | 100 NO shares | Reconcile before hedging or unwinding |
| Neither leg fills | 0 | None | Re-evaluate against a fresh book |
This is illustrative arithmetic, not an observed opportunity or a return claim. Confirm the actual market's settlement and token relationship before using a complete-set calculation. The Polymarket CLOB explanation describes why order-book depth and executed quantities belong in that assessment.
Verify related-market payoff coverage
Related contracts need a more detailed state table than a simple YES/NO pair. List every relevant event outcome and the payment from each proposed leg in that state. Include ties, cancellations and disputed resolution where applicable. Two contracts that usually move together can still leave a gap when their written definitions differ.
For example, a contract about a candidate winning an election and a contract about that candidate being inaugurated have related but distinct conditions. A model may rank them as close matches, but the intervening event matters. Treat such a pair as directional or basis exposure unless its full rules establish the required payoff relationship.
Keep the contract descriptions and resolution sources attached to each candidate. Recheck them when a market update changes the relevant conditions. A cached title match should not silently remain approved for execution after the underlying definition has changed.
AI-assisted mispricing detection
Models can rank candidate relationships, parse market descriptions and estimate liquidity. Check their proposed matches against the actual resolution rules before using them. A fluent interpretation can still pair contracts that settle differently.
Evaluate a model on held-out events and retain the executable book observed when each decision was made. Avoid using final outcomes, later prices or future liquidity in a replay. Predictions and actual arbitrage payoff relationships should remain separate in the logs.
HFT execution and market data
Use streaming books to maintain state and measure feed age, decision time, submission time and acknowledgement separately. The market channel is wss://ws-subscriptions-clob.polymarket.com/ws/market. Faster processing cannot restore a missed quote or make all legs fill.
Current order choices include GTC, GTD, FOK and FAK. FOK applies to its own order, not an entire multi-market strategy. Admission of a batch under a rate limit does not establish atomic execution across its legs.
Fees, rate limits and sizing
The documented taker fee is C × feeRate × p × (1 − p), with C as shares and p as price. The category’s feeRate is a formula parameter, not a flat percentage of trade value; maker fees are zero. Read the market’s configured parameters before assessing net spread.
API/IP throttles and signer trading budgets are separate controls. The published Standard signer tier allows 40 order tokens per second with a 60-token burst, and 80 cancellation tokens per second with a 120-token burst. Monitor tier and remaining-budget headers; the documentation’s rollout notice does not establish a new enforcement date.
Set a maximum residual position and a timeout for unfinished legs. Include cancellation and unwind costs in the decision, and stop when reconciliation cannot establish the current exposure. Do not increase size merely because an opportunity persists in a stale local book.
Low latency
Stay online and closer to execution.
Host your platform beside its market: Chicago for CME futures, New York for US markets, London for FX EAs.
See latency by datacenterFrom $29.99/mo · Windows or Linux
Evaluate AI matching and execution separately
Measure two different error types. A matching error pairs contracts with incompatible outcomes. An execution error uses a valid relationship but assumes prices or size that could not be obtained. An attractive historical result can conceal either problem if the test records only eventual resolution prices rather than the books available at decision time.
Use a held-out collection of events to review proposed relationships before enabling an execution policy. Include rejected candidates and explain the rejection. Accuracy on a set of obvious identical outcomes does not establish accuracy on ambiguous market wording. Human review can be particularly useful for a new contract family, while deterministic constraints should govern admission to the trading process.
The Python backtesting libraries comparison is a starting point for replay tooling. Whatever framework is chosen, preserve event-time inputs and avoid using later outcomes as earlier signals. A general price-series backtest needs additional work to represent order depth, partial fills and token settlement.
Budget requests around uncertainty
Allocate request capacity for reconciliation and cancellation rather than spending the entire budget on new opportunities. If a connection fails during a burst of orders, the system still needs to discover its actual inventory. Increasing the request rate can worsen that problem when throttled responses arrive after their data has become stale.
The official trading rate limits distinguish signer budgets from endpoint limits. Read the returned tier and remaining budget instead of assuming the same allowance for every account. Cancellation throughput and order throughput are separate quantities; they do not establish atomic completion of a multi-leg strategy.
Tools in the algorithmic trading software comparison should be judged by whether they support timestamped replay, bounded exposure and account reconciliation. A faster signal loop has limited value if an uncertain fill can trigger duplicate exposure.
Measure capacity before increasing size
Measure feed age, processing time and acknowledgement delay during busy periods, then inspect how often the observed candidate still exists at the submission time. Keep the maximum unhedged amount independent of the nominal size of a complete set. If completing the second leg would require a worse price than the policy permits, invoke the predefined unwind or pause condition rather than chasing the old quote.
For continuous workloads, measure the process requirements first, then size the algo VPS option or compare other algorithmic VPS resources against them. Hosting capacity supports a workload; it does not validate a mispricing model or guarantee fills.
Cross-venue opportunities have a different payoff and execution problem. The Binance-to-Polymarket arbitrage article covers that specific comparison, while Kalshi VPS hosting concerns a separate venue workload. Neither should be treated as evidence that two similar contracts form a matched hedge.
Arbitrage risk controls
Resolution disputes, missing depth and model matching errors remain relevant even when a payoff table looks balanced. Compare the written resolution sources and cancellation treatment before trading. Keep capital available for the longest plausible settlement period.
For the broader strategy context, see automated trading on Polymarket. This article covers internal arbitrage and HFT rather than general bot definitions or installation.
Uptime
99.999% uptime. Built for 24/7 trading reliability.
Your VPS runs in Tier IV datacenters with redundant power and network, monitored 24/7 on site.
See the uptime SLAFrom $29.99/mo · Windows or Linux
Hosting a Polymarket arbitrage bot
Check that every arbitrage leg settles on the same outcome and is still executable after fees, and measure how much size the book really takes before scaling up. Stay within Polymarket's API rate limits so the bot is not throttled when it needs to trade. A Dublin server for Polymarket bots sits in the closest non-georestricted region to Polymarket's London servers, under 1ms to Polymarket's order book.
FAQs
Is a price difference enough to establish arbitrage?
No. Compare executable size, fees and the full payoff relationship. Partial fills or different settlement conditions can leave a directional position.
Can an AI model guarantee accurate market matching?
No. Its proposed matches need review against the actual market rules. Retain those rules alongside the price data used in the decision.
Is buying a likely winner near resolution arbitrage?
It remains a directional position if the losing outcome is possible and there is no offsetting payoff that covers it. A high quoted probability does not eliminate resolution or execution risk. Keep those trades separate from verified complete-set relationships.
Does FOK make a whole arbitrage strategy atomic?
FOK governs the individual order to which it applies. It does not guarantee simultaneous completion of orders in other markets. Record each leg and limit the residual exposure if the full intended position is not obtained.
