Prediction Market Arbitrage Research and Failure Modes

Two contracts can look equivalent on a dashboard and settle differently in the only state that matters.

A constructed pair of prediction-market contracts can differ by six cents while describing almost the same event. If one uses an official release as its resolution source and the other allows a named media report, the apparent arbitrage may disappear once the two rules are placed side by side.

That is the recurring pattern. A real arbitrage locks a non-negative payoff across every allowed outcome after fees, fills, funding, and settlement risk. Most prediction-market spreads are relative-value trades wearing an arbitrage label.

Prove the payoff before pricing it

Binary YES and NO contracts often redeem to $1 or $0. Within one fully collateralized market, buying complementary outcomes below $1 can create a mechanical spread if both orders fill and fees preserve the gap. Cross-venue trades add a much harder condition: the contracts must define the same event and exhaust the same outcome space.

• Logical identity: wording, thresholds, geography, units, and event subject match.
• Time identity: start, end, timezone, observation window, and publication cutoff match.
• Source identity: both rules rely on the same authoritative result or compatible hierarchy.
• Edge-case identity: postponement, cancellation, recount, revision, and ambiguous outcomes match.
• Payoff identity: voids, ties, partial outcomes, and payout units preserve the hedge.

If one line differs, classify the trade as basis risk and size it accordingly.

Read the market rules as data

Polymarket's official resolution documentation says markets specify a resolution source, end date, and edge-case rules, with outcomes proposed and disputable through its stated oracle process. Those fields belong in the signal table beside price.

Kalshi's official order-book documentation explains that the API returns YES and NO bids. A YES bid at one price implies the complementary side of the binary book, though researchers still need the market's own rules and lifecycle fields.

Archive rule text and metadata at signal time. A current API response cannot prove which wording a trader saw before a clarification or venue update.

The spread needs executable prices

Displayed midpoints create imaginary arbitrage in thin books. Compute the complete quantity available at the prices you can actually take, then apply per-market fees, order minimums, and rounding.

Polymarket's current fee documentation describes market-dependent taker fees and fee-free categories. Kalshi's official fee page says fees depend on expected earnings and that some markets may charge maker fees. Query or calculate the exact fee that applies when the order is submitted.

Execution creates a legging problem

The first order can fill while the second moves away. A six-cent screen spread may become a directional position in a few hundred milliseconds. Browser automation adds render and session latency, while direct APIs still face network delay, rate limits, queue position, and stale snapshots.

Our cross-venue prediction-market study measured the distinction between browser tooling and direct book access. The general lesson is to timestamp market data at receipt and save the order acknowledgment from each venue.

1. Subscribe to both books and reject stale or crossed snapshots.
2. Calculate maximum matched quantity across the two executable ladders.
3. Submit under a predefined sequence, timeout, and price limit.
4. Update the hedge only from confirmed fills, never requested size.
5. Cancel residual orders when the minimum net spread disappears.
6. Reconcile positions and buying power after every partial or reject.

Simultaneous submission reduces delay and still cannot make two venues atomic. Your failure policy is part of the strategy.

Transfer time creates another non-atomic layer. Cash or collateral parked at one venue may not be available where the hedge is needed, and moving it after a spread appears is usually too slow. A realistic capacity estimate uses prefunded balances, venue-specific limits, and the smaller immediately executable leg. It does not assume capital can teleport to the richer book.

Resolution risk can dominate price risk

A disputed or delayed result locks capital and can break the hedge. Venue rules may rely on different sources, treat corrections differently, or resolve an event on different clocks. Legal access and account restrictions can also change which side remains tradeable.

Our options-versus-prediction-market investigation shows how two apparently similar payoffs separate once contract details and execution enter. The bonding-bot study shows the tail hidden in repeatedly buying contracts near $1.

• Map each venue's resolution source and dispute process.
• Set a maximum settlement delay in the funding model.
• Treat unclear edge cases as separate payoff states.
• Stress one side voiding while the other resolves normally.
• Do not reuse capital in simulation before it is actually released.

A cross-venue arbitrage validator

normalize(contract_a, contract_b):
  compare subject, predicate, threshold, unit
  compare start, end, timezone, source hierarchy
  enumerate void, delay, correction, dispute states
  return equivalent only if every payoff pair is hedged

net_edge(size):
  proceeds = minimum guaranteed settlement payoff
  cost = executable book cost on both venues
  friction = fees + funding + transfer + expected leg risk
  return proceeds - cost - friction

Store the validator output with a human-readable difference report. A model can help match candidate contracts, though deterministic fields and manual review should decide whether capital treats them as identical.

The point-in-time data guide covers rule and market snapshots. The event-driven tooling guide explains how to retain source provenance when an agent or service builds the match.

Backtest the failures, not just the spread

Historical research needs full depth or a defensible top-of-book execution model, market-rule versions, fee schedules, lifecycle events, and order outcomes. A last-trade series cannot reproduce whether both legs were available together.

• False match rate: candidate pairs rejected after rule comparison.
• Fill ratio: quantity filled on both sides relative to requested hedge.
• Leg loss: cost of flattening unmatched exposure.
• Resolution divergence: frequency and loss when rule paths differ.
• Capital days: time from trade to withdrawable settlement proceeds.
• Net edge: guaranteed payoff less executable cost and every measured friction.

Start in a demo or read-only environment where available, then shadow live books without sending orders. The official Alphanume documentation is useful for adjacent event datasets, though venue-native order and resolution records remain the primary evidence.

Replay the shadow system through maintenance windows and thin hours, not only liquid news events. Save rejected orders, disconnected streams, changed tick sizes, and rule updates. These observations estimate the probability that an apparent locked payoff was never jointly tradeable, which is often the dominant denominator in a small sample.

The honest conclusion

Prediction-market arbitrage exists, though the easy dashboard version usually omits contract identity, size, fees, and atomicity. The strongest opportunities are operational research problems before they become trading problems.

Build the rule diff first, walk both books second, and simulate the broken second leg. A spread that survives those three steps has earned a small live test.

Keep the word arbitrage for trades whose payoff remains locked after fees, partial fills, disputes, clarifications, and the ugly resolution states are included. Your final artifact should show the exact rule clauses that make the payoffs mutually exhaustive.