The Anatomy of a Decentralized Prediction Market: Microstructure Evidence from the Polymarket Order Book

Battle Count: 65

Highly relevant for quantitative traders operating in prediction markets and decentralized CLOB venues. The paper provides critical microstructure evidence (spreads, depth profiles, maker concentration, liquidity dynamics) essential for market-making strategies, execution algorithms, and information-trading cost estimation in prediction markets. The trade-direction inference finding is a methodological warning that directly impacts any quantitative strategy relying on order-book-derived aggressor signs. The Glosten-Harris decomposition and Kyle's lambda analysis are standard tools in quantitative trading. However, the paper is primarily descriptive/analytical rather than prescriptive for trading strategies, and the prediction market domain is narrower than traditional equity/crypto trading.

High complexity: requires processing 30 billion tick-level events (624GB), scraping 255 million on-chain OrderFilled events from Polygon with adaptive batch sizing and reorg-buffer handling, performing schema reconciliation across off-chain and on-chain data sources via CLOB REST API, implementing multiple microstructure estimators (effective spread, realized spread, Kyle's lambda, Amihud, Roll, Abdi-Ranaldo), Glosten-Harris decomposition, bootstrap confidence intervals with market clustering, pre-registered stratified panel construction, and linear regression with HC3 robust standard errors. The on-chain join pipeline and trade-direction inference validation are particularly non-trivial.

Exceptional reproducibility: pre-registered panel with committed random seed (20260424), deterministic build script with SHA-256 hash recorded in pre-registration document, full code on GitHub (https://github.com/philippdubach/polymarket-microstructure), Zenodo archive (DOI 10.5281/zenodo.19811426), panel parquet artifacts directly provided, on-chain scrape pipeline reproducible against any Polygon RPC provider with archive-node access. The only gap is the 624GB raw WebSocket archive which is not redistributed but access can be arranged via the author.