Forecast-Gated Passive Liquidity Provision in Five-Minute Bitcoin Prediction Markets

This paper studies liquidity provision in five-minute Bitcoin Up/Down prediction markets using resolved markets, sampled top-of-book states, L1 transitions, and Chainlink BTC/USD reference prices. A parsimonious Chainlink-implied probability has a Brier score of 0.1684, versus 0.2529 for the displayed midpoint. Edge filters consequently select quotes with positive ex-post terminal payoff, although these conditional quote-quality statistics are not execution returns. Displayed liquidity is short-lived: the median best-quote lifetime is 59 ms, and visible complete-set violations are usually single-update events. We propose TAHMM, a transparent time-to-expiry and adverse-selection haircut rule with market-specific inventory control. Its empirical contribution is a reproducible state-selection and wallet-accounting framework; exact fill profitability remains unidentified because the archive lacks trades, order identifiers, and queue position. A funded replay on the final 20 percent of markets starts from 1,000 USDC and ends between 2,466.11 and 5,363.87 USDC across three L1-based fill conventions. This wide range is itself a central result: top-of-book changes cannot identify passive fills, so the wallet experiment is a sensitivity analysis rather than evidence of executable returns. The analyzed sample contains 2,034 resolved five-minute markets from 2026-05-21 12:50:00 UTC to 2026-05-28 15:10:00 UTC, 53,975 sampled Up top-of-book quote states, and Chainlink reference observations covering 2026-05-21 12:45:34 UTC to 2026-05-28 15:11:05 UTC. The raw L1 archive spans 2026-05-14 09:10:00 UTC to 2026-05-28 16:55:00 UTC; the narrower analyzed interval is imposed by Chainlink coverage and the requirement that every market has start and end reference prices.