Post-Rejection Follow-up Sampling: A Methodology for Counterfactual Outcome Measurement in Algorithmic DEX Trading

Battle Count: 69

Directly addresses a fundamental gap in algorithmic trading evaluation: the inability to measure counterfactual outcomes of rejected trading candidates. PRFS provides an empirical methodology for filter precision evaluation on DEXs, which is critical for calibrating entry/exit rules in automated trading systems. The methodology is dataset-independent and generalises to any algorithmic decision system where rejections outnumber executions. Highly relevant to DEX trading, MEV strategies, and automated token screening. The companion papers extend this to precision auditing and benchmark datasets.

The PRFS architecture is structurally straightforward: a fan-in/fan-out design with a rejection event store, a subscriber, and a sampling scheduler querying a price oracle. The scheduler is single-threaded but asynchronous. Key complexity lies in: (1) integrating with the existing scanner's rejection stream without perturbing production trading, (2) managing adaptive sampling cadence and termination conditions (horizon expiry, token disappearance, retry budget exhaustion), (3) respecting price-oracle API throughput limits, and (4) handling the bimodal distribution of outcomes (early-disappearance vs. long-lived tokens). No ML model training is involved; it is primarily a data-engineering and systems-integration task.

The companion dataset (2,997 rejection events, 67,000 forward-observation rows, 457 unique mints) is publicly deposited on Zenodo with a DOI. The deposit format, data architecture, and sampling methodology are described in detail. However, the specific filter rule descriptions, threshold values, mapping between anonymised filter labels and internal rule identifiers, and per-category functional taxonomy are proprietary and not disclosed. The per-category horizon mapping is also undisclosed. The trading system is subject to a pending U.S. provisional patent application. Replication of the exact deployment is limited by these proprietary elements, though the methodology itself is generalisable.