Excess Movement in High-Frequency Prediction-Market Quotes: The Entry-Delay Frontier — replication package

Replication package for the working paper Excess Movement in High-Frequency Prediction-Market Quotes: The Entry-Delay Frontier The paper takes a known identity — for a martingale settling at 1 or 0, expected quadratic variation equals q0(1−q0) — and asks what happens to it when the measured object is a quoted midpoint sampled at one minute rather than a belief observed sparsely. In 7,118 Polymarket sports contracts the aggregate ratio of realized variance to the binary variance budget is 1.53 at one-minute sampling, 1.05 at sixty minutes, and 1.009 [0.997, 1.020] at the coarsest grid. An exact finite-grid regrouping splits that gap into initial calibration and return cross-products and resolves the second by lag. Miscalibration is small; the excess sits at short lags. A non-zero lag-two raw cross-moment rejects the joint null of a martingale fair value, zero-mean temporally independent quote error, and orthogonality between quote error and fair-value innovations. Serially correlated error remains viable. The resulting ten-point one-minute midpoint fade is then priced against the venue's own tape of 1,153,546 transactions. Across 36 combinations of entry delay and holding horizon applied to 3,000 signals in 396 markets, every policy point estimate is negative; the best is −0.327 points [−0.70, +0.04], while the market-clustered simultaneous one-sided upper bound on the grid maximum is +0.44, so the grid shows no profitable policy rather than proving none exists. Among tape-covered signals, 39% carry no transaction of any kind, and their midpoint reversion is 5.234 points against 1.069 where an entry print exists. This deposit contains the paper and its figures, the full analysis pipeline (25 scripts), all 27 result artifacts backing every reported figure, and the retrieval code that reconstructs the raw inputs from the venue's documented public API. All computation is CPU-only and deterministic: seed 20260721, with every bootstrap resampling markets rather than the signals nested inside them, at B = 150,000 replications. Interval endpoints are quoted only to the precision their seed spread identifies, under a rule fixed before the spreads were measured. Appendix A of the paper maps each headline number to the script and artifact that produce it; data/decomp/RESULTS.md is the longer provenance record, with sixteen addenda documenting results that were withdrawn, reversed or re-measured during the work and the evidence that forced each change. Scope: one venue, two U.S. sports leagues (MLB and NBA), 2024-10 to 2026-07. This is a warning case study about computing high-frequency statistics on raw quote midpoints at decentralized or thinly quoted central limit order books. It is not a general law about prediction markets and not a finding about market efficiency.