# Post-Rejection Follow-up Sampling: Measuring Outcomes of Rejected Decisions in Algorithmic DEX Trading

Crypto.world

> Filter-gated trading systems on decentralised exchanges reject most of the tokens they evaluate, and the operating record usually preserves only the reason for rejection — never what the token's…

**arXiv**: 2606.08228

**Published**: 2026-06-06

**Authors**: Arati Uday Kamat

## AI summary

## Overview

**Research area:** Quantitative finance / market microstructure — specifically measurement methodology for algorithmic trading on decentralised exchanges (DEXs), sitting at the intersection of the "selective labels" literature in machine learning and empirical DeFi research.

**Technical level:** Intermediate. The core idea is intuitive and explained in plain terms, but the paper uses formal specification language, Wilson confidence intervals, chi-square tests and clustering caveats.

**Scope (one sentence):** The paper defines and demonstrates Post-Rejection Follow-up Sampling (PRFS), a method for directly observing what the market did to tokens an algorithmic DEX trading system *rejected*, using a companion dataset of 67,000 observation rows covering 2,997 rejection events across 457 tokens.

## What This Paper Is About

Filter-gated trading systems on decentralised exchanges reject most of the tokens they evaluate, and the operating record usually preserves only the *reason* for rejection — never what the token's price and liquidity did afterwards. Because every attempt to audit, compare or tune those filters needs evidence about the rejected side, the paper asks whether an operator can cheaply record the post-rejection trajectory of each rejected token using the *same live price oracle* the rejecting scanner already uses, without touching the scanner itself. The goal is a formal, reproducible measurement primitive — not a better trading rule — and the paper ships a reference implementation, a frozen dataset, and an explicit accounting of what the data can and cannot support.

## Key Contributions

1. **A formal specification of PRFS as a measurement primitive.** PRFS is defined as a triple of decoupled subsystems (a durable rejection log, a scheduler that enqueues follow-up samples at fixed offsets, and an oracle adapter writing to an append-only outcome log) with explicit rules for event identity, the observation window (DELTA_MAX_MS = 743,040,000 ms, i.e. 8.60 days), the default schedule ({5, 15, 60, 240, 1440} minutes), a reason-attribution rule, censoring labels, and the definition and scope of the estimand.
2. **A reference implementation with executable checks.** `prfs v2.0.0` regenerates every headline number from the deposited CSVs under continuous integration, with tests for restart determinism (`tests/test_restart.py`), reason-set semantics (`tests/test_reason_ledger.py`), and cross-implementation agreement (`test_oracle_agreement.py`).
3. **A secondary, independent cross-check.** A separate aggregator (`oracle/oracle_aggregate.py`) using different code and a different accumulation strategy, plus a third stdlib-only script (`scratchpad/p2_repro/independent_repro.py`), both reproduce 1,455 / 2,997 primary coverage, 1,641 / 2,997 secondary coverage, 253 conflicts, 263 orphan keys and the seven per-filter rows.
4. **A frozen, provenance-locked dataset with disclosed gaps.** 2,997 rejection events and 67,000 forward-observation rows on a single Solana pump.fun sub-venue, with SHA-256 checksums on deposit, plus full disclosure of the residual attribution gap (253 conflict events, 263 orphan reason emissions, 8,593 unaligned outcome rows) and the non-random missingness patterns.

## Main Findings

- **Primary coverage is 48.55 percent.** Under the reason-aware three-field key (mint, timestamp, reason), 1,455 of 2,997 events receive a matched forward observation; the naive event-independence Wilson 95 percent interval is [46.76 percent, 50.34 percent].
- **Diagnostic coverage is 54.75 percent.** Under the reason-agnostic two-field key (mint, timestamp), 1,641 of 2,997 events match. The 186-event surplus (1,641 − 1,455) is exactly the set of matches where the recorded reason changed between the registry and the outcome side.
- **Mint-level coverage is complete.** All 457 of 457 unique mints have at least one forward observation, even though event-level coverage is below half.
- **Coverage is not missing-at-random, but the corrected reading is about the tracker, not the sampler.** A Pearson chi-square test on the filter-versus-coverage table gives chi-square(6) = 401.29 (p approximately zero) and a two-proportion z-test on the source difference gives z = 14.78 (p approximately zero). The October 2026 data-collection note states these associations reflect the production tracker's start time and record keying, not sampler reach.
- **The deposited data were not collected under the PRFS schedule (a published correction).** Version 2 described the deposit as collected under the fixed 8.60-day PRFS schedule. The v3 note corrects this: the production tracker sampled every tracked token every 10 minutes (99.41 percent of same-token sample gaps fall between 9.5 and 10.5 minutes), followed each record for at most 1,440 minutes, and kept one record per token, so a new rejection of a token ended the previous follow-up.
- **Coverage gaps are fully explained by start-up and one-record-per-token tracking.** 731 events (24.4 percent) precede the tracker's first sample; after it, every event not followed within 10.5 minutes by another rejection of the same token has an outcome record (894 of 894).
- **All 186 reason mismatches are artefacts of record keying.** Each carries the label of an earlier rejection of the same token.
- **Events are heavily clustered within mints.** 2,997 events come from 457 mints: mean 6.56 events per mint, median 3, ninetieth percentile 14, maximum 89, and 96 percent of events belong to multi-event mints. The reported Wilson intervals are therefore described as lower bounds on true uncertainty, with a mint-level aggregation given as a sensitivity check.
- **One filter dominates the sample.** `filter_2` alone contains 48 percent of the deposit, and the two source labels (`source_a`, a lower-latency direct-launch feed; `source_b`, a higher-latency aggregated feed) partition events roughly evenly, so per-filter and per-source comparisons need their denominators reported alongside every estimate.
- **No profit-and-loss claim is made.** The reported estimand is the observed post-rejection forward return r(e, tau) = (obs_price(e, tau) / p0(e)) − 1; a hypothetical executed-PnL estimand is defined but requires an explicit execution model, and the reference implementation raises `prfs.estimand.MissingExecutionModel` without one.

## Methodology in Plain English

The setup is simple to state. A trading bot looks at candidate tokens and mostly says "no," recording only a reason code. PRFS adds a bystander process that listens to that stream of "no" decisions and, at fixed delays after each rejection (5, 15, 60, 240 and 1,440 minutes), asks the same live price oracle the bot used: what is this token worth now, and how much liquidity does it have? Every answer is appended to a log. Because the three parts — rejection log, scheduler, oracle adapter — share no memory and talk only through durable append-only files, the scheduler can restart without losing pending lookups, and the oracle adapter can be swapped out.

Two practical obstacles get explicit treatment. First, tokens can vanish from the oracle entirely; rather than silently dropping them, the design records a terminal "disappeared" marker or an "oracle absence" placeholder after a bounded number of retries with jittered exponential backoff, so a failed lookups is visible, not invisible. Second, the reason label attached to a rejection can differ between the moment the scanner emits it and the moment the sampler writes its row. Rather than overwriting one side, the design keeps both, treats the registry side as authoritative for per-filter splits, and files every disagreement in a conflict ledger with a disposition of `KEEP_REGISTRY_AS_PRIMARY`.

Matching is done two ways to separate two questions. A strict three-field key (token, timestamp, reason) asks "did we observe this exact rejection, with its reason intact?" — the basis for any per-filter statement. A looser two-field key (token, timestamp) asks only "did the sampler ever reach this token at this moment?" — an integrity check on the plumbing, reported only as a diagnostic.

On the statistics, the paper deliberately stays descriptive. Coverage proportions get Wilson score intervals rather than p-values; the only two hypothesis tests are framed as tests of the missing-completely-at-random null, not as claims about filter or source effects. The authors flag openly that treating 2,997 events as independent is wrong given only 457 mints, and point readers to a mint-level aggregation and to cluster-robust errors for any downstream work.

The reproducibility scaffolding is unusually heavy for a measurement paper: byte-level SHA-256 checksums on the two CSVs, a primary implementation, a cross-check aggregator that shares no code with it, and a third stdlib-only recompute script — all gated by the test suite so that any change breaking agreement fails CI.

## Why This Matters

**Impact on research.** The paper reframes a familiar inference problem as an observation problem. In credit scoring or bail decisions, a rejected applicant leaves no outcome trace, so the literature (Lakkaraju et al. 2017; Coston et al. 2020; Crook and Banasik 2004) must identify the counterfactual through selection correction. On a DEX, the rejected token's price keeps moving whether or not anyone watches, so a sampler can simply look. The paper is careful about what this buys: the observed trajectory is explicitly *not* a Rubin-Neyman potential outcome, because a hypothetical order would have moved the price, changed fill probability, and engaged execution logic. Naming that gap precisely is itself a contribution, and the paper argues no prior deposit in the surveyed literature (Fang et al. 2022; Daian et al. 2020; Bailey and Lopez de Prado 2014) provides a cadence-matched, disappearance-terminated follow-up stream tied to a live rejecting scanner.

**Real-world applications:**

- **Filter auditing and threshold tuning.** An operator can compute the forward-return distribution of everything a given rule rejected, rather than only the population the rule admitted.
- **Comparative rule benchmarking.** Two candidate filters can be compared on the same rejected population instead of on what each happened to accept.
- **Backtest base-rate construction.** A cadence-matched observation stream replaces offline reconstruction from price-feed archives, whose snapshot policy and oracle-absence pattern differ from the live decision environment.
- **Model risk and incident review.** The reason ledger and censoring summary make it possible to distinguish "the rule was wrong" from "the data pipeline lost the token."

**Industry relevance.** The method is deliberately low-cost and non-invasive: it runs beside the scanner, shares the scanner's own oracle, and requires no change to the decision path. That makes it adoptable as operational hygiene in trading stacks that already keep scan logs, and the paper's framing of ad-hoc logging versus a formalised primitive with a defined window, cadence, attribution rule and reproducibility contract is aimed squarely at practitioners who currently have the former and not the latter.

## Future Directions

- **Replication on other venues and windows.** The paper repeatedly states that the deposit is a single-venue, single-window instance and that PRFS is meant to be re-run on other AMM venues, Ethereum layer-2 surfaces, and non-launch-focused systems before any coverage figure is generalised.
- **Running the schedule as specified.** The correction note reveals a gap between the specified PRFS design (fixed 5/15/60/240/1440-minute ticks out to 8.60 days) and what the production tracker actually did (10-minute polling, one record per token, follow-up capped at 1,440 minutes). A deposit collected under the actual fixed schedule, and with the adaptive extras enabled, is the obvious next empirical step.
- **Clustering-aware inference downstream.** The paper hands off to companion work (Kamat 2026a, 2026b) and explicitly warns that any downstream inference must use cluster-robust standard errors or mint-level bootstrapping; how much the naive intervals understate uncertainty per filter and per day is left open.
- **An execution model for a PnL estimand.** The machinery for a hypothetical executed-PnL estimand exists but raises `MissingExecutionModel`; specifying entry and exit slippage, fees, exit rule and portfolio constraints would let the observation stream answer profitability questions rather than only return-distribution questions.

## Target Audience

Quantitative researchers and trading-system engineers working on DEX or AMM strategies will get the most direct value, particularly those who own a filter stack and have wondered what their rejections actually did. It also serves methodologists interested in selective labels and reject inference who want a concrete case where direct observation is available, and reproducibility-focused readers, for whom the checksum-locked deposit, dual implementations and CI-gated agreement tests are a useful template. Statisticians will find the honest clustering caveats and the explicit refusal to claim MCAR or MAR a refreshing change from the usual overreach — and the published v2-to-v3 correction, which leaves the design, deposit and headline counts unchanged while rewriting what the missingness patterns mean, is itself a small case study in how measurement claims should be revised.

## Authors’ abstract

Note (Oct 2026): data were collected by a tracker with a 10-minute sweep, 24-hour cap and one record per token, not the PRFS schedule of v1-v2; headline counts unchanged. Filter-gated algorithmic trading systems on decentralised exchanges reject most candidate tokens they evaluate, yet the observed forward market trajectory of those rejected candidates is rarely measured on the same live venue that produced the rejection. This paper introduces Post-Rejection Follow-up…

[The full abstract on arXiv](https://arxiv.org/abs/2606.08228)

[Read the original paper](https://arxiv.org/abs/2606.08228)

---

Canonical: https://crypto.world/paper/post-rejection-follow-up-sampling-measuring-outcomes-of-rejected-2606.08228
