UCLA

This thesis leverages the transparent, on-chain architecture of Polymarket to empirically study algorithmic arbitrage and insider trading within NBA prediction markets. Analysis of high-frequency limit order book snapshots reveals extreme microstructural efficiency; single-market mispricings resolve in a median of 3.614 seconds, while combinatorial arbitrage is economically bounded by shallow liquidity. To detect insider trading without regulatory labels, an unsupervised Isolation Forest model is trained on pre-resolution behavioral features and calibrated against a quantitative Pseudo-Ground Truth (PGT) Oracle. The model’s top 1.0% anomaly cohort captured a statistically significant 11.7% of aggregate market profits. These findings demonstrate that while structural inefficiencies on Polymarket are fleeting, distinct on-chain behavioral signatures effectively isolate insider trading.