Prediction markets let participants trade contracts that pay one dollar when
a given outcome occurs. Equivalent and logically related markets on different platforms frequently diverge in price, violating the law of one price and
opening arbitrage opportunities. This thesis presents a system that detects
combinatorial arbitrage across four prediction-market venues (Kalshi, Polymarket, Predict and Limitless). Its core is a deterministic-first pipeline that
builds a cross-platform identity layer, together with a linear program over the
outcome space Ω that proves the existence of arbitrage rather than relying
on price correlation. The system is evaluated on two levels: the correctness
of the identity layer in defining the outcome space Ω (candidate-generation
recall measured on our own verified set, with no labelled precision evaluation carried out) and the economic soundness of the detected opportunities,
measured in a nine-day continuous passive live observation run in which no
orders were placed. That run recorded 796 arbitrage structures carrying
a modelled payoff floor of $6,148 at the order-book depth actually quoted,
which amounts to a capital-weighted aggregate return of 0.52 % on the capital those structures would tie up. Of the 308 structures that resolved by
the end of the observation window, 306 paid at or above their own computed floor; those 306 episodes paid $533 against a computed floor of $522.
Two structures deviated from their floors. The first deviation, a shortfall of
roughly $0.38, was caused by a divergence between the settlement rules of
the two venues involved, while the second structure had not yet fully settled
by the end of the window.
|
