Per-Market Information Leakage and Order-Flow Skill: Two Methodological Lenses on Informed Trading in Decentralized Prediction Markets

Battle Count: 81

Highly relevant to prediction market trading, market microstructure research, and regulatory compliance in decentralized finance. The paper directly addresses order-flow analysis, information asymmetry detection, and price discovery mechanisms in binary prediction markets. The ILSdl framework provides a per-market measure of information front-loading relevant to trading strategy evaluation. The sign-randomization methodology distinguishes skill from luck in directional trading. However, the paper is primarily methodological/comparative rather than providing directly implementable trading signals. Most relevant for prediction market participants, compliance officers, and market surveillance teams rather than traditional quantitative trading.

The sign-randomization classifier requires processing 1.72 million accounts across 210,322 markets with 10,000 permutations per account, demanding substantial computational infrastructure. The ILSdl framework requires LLM-assisted T_event recovery (~$0.09/market), complete CLOB price coverage, scope-condition verification, and anchor-sensitivity analysis. The combined surveillance pipeline requires category-conditioned decomposition, per-account context features (wallet age, funding provenance, sybil-cluster membership), and multi-stage threshold calibration. The lifecycle heuristic is simpler but requires per-event account creation timestamps and activity windows. Overall, operational deployment requires significant data infrastructure and ongoing calibration.

The paper explicitly notes that Gomez-Cram et al. [2026] account-level classifications, the labelled list of 1,950 heuristic-flagged accounts, category-conditioned decompositions, and source code for sign-randomization have NOT been released. Polymarket transaction history is publicly available on-chain, making the methodology reproducible in principle but requiring significant infrastructure. The author releases three accompanying datasets (polymarket-tnews-tevent-recovery-v1, polymarket-hazard-rates-v1, polymarket-ils-corpus-v1) on GitHub. The ILSdl computation on the Maduro cluster is proposed but not executed. The author did not independently reproduce Gomez-Cram et al.'s classifications.