Outcome-Classified Precision Auditing of Filter Rules in Algorithmic DEX Trading: Evidence from 2,400 Rejection Events

Battle Count: 70

Directly addresses a core operational question in algorithmic trading: whether rejection filters are net-positive. Provides a rigorous, reproducible methodology (PRFS + five-tier classification) for auditing filter precision using live counterfactual data rather than reconstructed backtests. The conservative 3.7:1 save-to-miss ratio and per-filter breakdowns are immediately actionable for filter tuning. The external validation of the early-death tier against lifecycle data adds methodological rigor. However, the paper is specific to DEX/Solana microstructure and does not address portfolio-level or multi-asset trading strategies.

The audit methodology itself is straightforward: a deterministic single-pass classification script over a JSONL file with well-defined thresholds (50% drawdown, 200% run-up, 60-minute age cutoff). The primary complexity lies in the data collection infrastructure (PRFS tracking subsystem sampling price/liquidity at configurable intervals over 24 hours) and the external validation join against a separate lifecycle dataset. The classification logic, tie-breaking rules, and ratio computations are simple arithmetic. No ML model training or complex optimization is involved.

The paper provides a fully deposited companion dataset (Zenodo DOI 10.5281/zenodo.19987695) including the raw JSONL observation file (99,510 rows), a deterministic single-pass audit script (Kamat_FilterPrecision_2026_audit.js), figure-generation scripts, provenance documentation, and a whitelist-based verification snippet. The audit script is described as producing identical classification distributions on any run against the deposited file. Version history documents corrections from v1 to v2. All filter labels are anonymised deterministically by descending count.