Testing whether apparent pricing edge survives real execution frictions.
Illustrative interface values - not live trading performance or empirical research results.
Early tick replay suggested that short-horizon prediction markets might contain pricing edge. Live-like execution produced a weaker result once network latency, API response time, stale order books, failed fills, and reversals near settlement were included.
The research question therefore shifted from whether a model could find a price discrepancy to whether that discrepancy could survive the execution funnel and become reliable filled exposure.
Instead of presenting the strongest early replay result, the project made execution degradation the main research object. This gives up an easier strategy narrative in favor of a question that better reflects real market conditions.
ML is applied after edge detection to decide whether a candidate deserves exposure, with chronological splits used for later-sample checks. This avoids treating an expected-value score as a profitability guarantee.
The public version keeps anonymized, downsampled, and normalized data with simplified diagnostics while excluding wallets, signers, private ledgers, model artifacts, and strategy-sensitive parameters.
The public execution sample records 1,000 attempts with 5.8% accepted and filled rates. Across 986 joined market-level observations, market-implied probability was slightly better calibrated than the fair-probability model on both Brier score and log loss.
The sample does not support a stable profitability claim. Its main result is diagnostic: most apparent edge did not become filled exposure, extreme probability buckets were sparse and unstable, and reasonable ML or fill-probability gates could fail on later data or suppress nearly all trade flow.
Illustrative interface values - not live trading performance or empirical research results.
Tracing how public-sample attempts move through submission, acceptance, and fill states.
Comparing fair and market-implied probabilities against 986 joined public-sample outcomes.
A public-sample interface for execution, calibration, ML-filter, and risk diagnostics.
This project is presented as a public research and portfolio artifact. It does not represent financial advice, trading advice, or a claim of trading profitability.