本文由 AI 分析生成
建立時間： 2026-02-11 來源： https://x.com/RohOnChain/status/2023781142663754049
Summary
Roan walks through how institutions extract alpha from prediction-market data using Jon Becker’s newly public dataset of 400M+ Polymarket/Kalshi trades back to 2020 (tick-level, MIT-licensed, 36GB). After step-by-step setup with uv/DuckDB/Parquet, he details three institutional methods: empirical Kelly sizing with Monte Carlo drawdown analysis, calibration-surface analysis across price and time, and order-flow decomposition showing makers systematically beat takers.
Roan 講解機構如何利用 Jon Becker 新近公開的數據集（4 億+ Polymarket/Kalshi 交易，回溯至 2020 年，tick 級、MIT 授權、36GB）從預測市場中提取 alpha。在用 uv/DuckDB/Parquet 完成逐步設置後，他詳述三種機構方法：結合蒙地卡羅回撤分析的經驗 Kelly 倉位法、跨價格與時間維度的校準曲面分析，以及顯示做市方系統性勝過吃單方的訂單流分解。
Key Points
- The dataset (400M+ trades, tick-level with taker direction, resolutions included) gives retail the granularity institutional vendors charge $100K+/yr for.
- Method 1 — Empirical Kelly: textbook Kelly assumes edge is known; real edge is a distribution, so they pattern-match historical analogs, build empirical (fat-tailed) return distributions, run Monte Carlo resampling for path-dependent drawdowns, and haircut sizing by f = f_kelly × (1 - CV_edge). Sizing for the 95th-percentile drawdown, not the median, is the institutional vs retail divide.
- Method 2 — Calibration surface C(p,t): longshot bias is real and measured — at 1-cent contracts takers win only 0.43% vs 1% implied (-57% mispricing); takers have negative excess returns at 80 of 99 price levels. Institutions extend this temporally (early bias → mid efficiency → late reversal hypothesis).
- Method 3 — Maker vs taker: makers earn +0.77% to +1.25% excess return structurally (symmetric YES/NO, Cohen’s d ≈ 0.02), proving the edge is structural spread capture + biased taker flow, not superior forecasting; risks are inventory and adverse selection.
Insights
The unifying thesis subverts the retail assumption that hedge funds win on better information: the edge is process, not prediction — risk management (sizing for the distribution), time-varying calibration exploitation, and structural positioning (being the maker). The prediction market is treated as a “laboratory” for patterns that inform billions in traditional positions, not as the venue for capital deployment. The honest separation of “verified research” (price-dimension longshot bias) from “institutional hypothesis” (time dimension, requires validation) is methodologically careful and worth noting.
Connections
Raw Excerpt
The edge is in: Risk management … Time-varying strategies … Structural positioning … None of these require better prediction. They require better process. The Becker dataset gives you the laboratory to build that process.