BL-SD-023: Provable-Winner Cohort Persistence

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Hyperliquid perpetuals Programme: I Verdict: FALSIFIED Axis of death: Idea Transferability: UNIVERSAL

1. Identification

The design named this hypothesis primary rather than confirmation: the on-chain fact of a wallet's complete fill history with realised profit and loss was to be the edge itself, not a filter on some other signal. The premise gate was written into the design before a single wallet was scored, on the grounds that a cohort-following strategy has no reason to exist if the wallets being followed do not display skill that persists from one period to the next.

2. Economic thesis

Every wallet's fill history on Hyperliquid, including realised profit and loss per closed position, is public. A cohort formed from wallets that are provably and persistently profitable, weighted by realised PnL rather than by notional size, should see its aggregate position changes lead price, because size is not skill and a copy-trading ecosystem built on dashboards that poll positions at minute resolution is necessarily slower than a system that reads counterparties off every trade print.

Who is forced to lose. Two parties, named at design time: the late copy-trader, who follows the same wallets through a third-party dashboard with a lag measured in minutes and therefore pushes price after this strategy would already be positioned; and the retail flow selling into a cohort that is quietly accumulating. Both losses are structural only if the cohort being followed is actually skilled rather than merely lucky or merely large, which is exactly the premise this hypothesis was designed to test before anything else.

3. Falsifiable claim

Pre-registered in the design document, section II.S2.A.3, before any wallet was scored:

Skill, measured as realised profit and loss per wallet, is persistent across non-overlapping calendar windows. Spearman rank correlation of per-wallet scores between consecutive window pairs is positive with p below 0.01 in at least two window pairs.

The gate is written to kill the hypothesis for free, without a single backtest of aggregate position changes against forward price. If wallets that were skilled in one window are not identifiably the same wallets that are skilled in the next, there is no cohort to follow and the event study the design also sketches is moot before it is built.

4. Mathematical specification

Notation per BL-D-011.
 
Candidate population, enumerated from trade counterparties observed in the lake:
 
  W = { w : sum over trades i with w in {user_a_i, user_b_i} of |notional_i| >= 100,000 USD }
 
Per candidate w, non-overlapping calendar windows n = 0, 1, 2:
 
  score(w, n) = sum over fills f in userFills(w) with f.time in window n of  closedPnl(f)
 
Wash-trading exclusion, applied before scoring:
 
  exclude w if  median inter-fill time gap for w  <  1 second
              ( burst-bot signature, not skill )
 
Span exclusion, applied before scoring:
 
  exclude w if  fill history for w  is truncated by the venue's 10,000-fill retrieval cap
              to a span shorter than 3 days
              ( too little history to assign to 3 non-overlapping windows honestly )
 
Persistence statistic, per consecutive window pair:
 
  rho(n, n+1) = Spearman rank correlation of { score(w, n) } against { score(w, n+1) }
                over the population with qualifying fills in both windows
 
Gate:
 
  ALIVE  if  rho(n, n+1) > 0  and  p < 0.01   for at least 2 of the pairs tested
  DEAD   otherwise

Causality holds by construction: score(w, n) uses only fills timestamped within window n, and rho(n, n+1) is computed only from windows that have both fully closed by the time the statistic is evaluated. No wallet's future window contributes to its own score.

The design document sketches an entry, exit, sizing, and cost specification for a cohort-following event study, at sections II.S2.A.4 through A.5: a breadth filter of at least N_w cohort addresses trading the same direction within a window W, evaluated at horizons of 5 minutes, 30 minutes, 2 hours, and 8 hours. That specification was never exercised. The gate above is the only measurement this hypothesis produced, because the population it would need to follow did not display the property the follow-on design assumed.

5. Operating envelope

There is no execution channel to report because the hypothesis died at the premise gate, before an entry or exit rule was ever run against forward price.

6. Instrument

The cohort candidate-enumeration and persistence-testing instrument enumerates candidates from trade counterparties recorded in the lake, retrieves userFills on demand per candidate, applies the wash and span exclusions, and computes the persistence statistic. A first run of this instrument against the top 400 candidates by notional, at a 60-per-minute pull rate, failed on 335 of 400 fetch attempts and was discarded as an instrument failure rather than reported as a result. The second run, over the top 800 candidates at a slower five-second pace with retry, completed with zero fetch failures and is the run this dossier reports.

Half the enumerated population was excluded as wash-suspect. That is not a tuning choice made after seeing the result; the exclusion criterion is a mechanical property of inter-fill timing, fixed before any wallet was scored, and it happens to remove half the candidate pool. The consequence, reported honestly, is that the population available to test persistence on is small.

7. Results

Zero of two pairs meet the significance threshold. More telling than the individual p-values is the sign: the correlation is positive in the first pair and negative in the second. A population with genuinely persistent skill does not reverse the direction of that persistence between adjacent measurement windows. This is not a story about insufficient power alone; it is a population whose ranking does not hold together from one window to the next.

8. Verdict, axis of death, and transferability

Verdict: FALSIFIED. Axis: idea. Transferability: UNIVERSAL.

Kill criterion (c), pre-registered in the original design, fires exactly as designed: rank correlation of realised skill across non-overlapping windows reverses sign rather than persisting, and zero of the two tested pairs clear the significance bar. The hypothesis dies at the gate, before an event study, a null test, or a single PnL number was produced.

This is a mechanism failure rather than a cost, power, or execution failure. No envelope substitution changes a correlation that reverses sign: a wallet ranked in the top decile of realised profit in one window is not reliably ranked in the top decile in the next, on the history retrievable from the venue's API. At zero cost, at any execution channel, at any capital scale, there is still no cohort to follow, because the property the cohort-following design depends on was not found in the population tested.

One qualification belongs on this classification and it does not rescue the verdict, but it sharpens what was actually falsified. The population tested is 23 wallets, drawn from a candidate pool of 800 after wash and span exclusions removed roughly 96 percent of it. The precise finding is that persistence was not found on the fill history retrievable from the venue's API within the programme's data era, not that no persistent skill exists anywhere on this venue at any timescale. Section 10 states what would be required to test the stronger claim.

9. Cost and capacity analysis

No cost or capacity analysis was performed. The premise gate is designed to kill a hypothesis before a trading cost is ever computed, and it did so here. Nothing in sections 7 or 8 depends on a fee schedule, a maker or taker channel, or a capital scale, which is the entire point of running the gate first: this falsification cost nothing beyond the API pulls used to score the candidates.

10. Envelope sensitivity and reopening conditions

No binding parameter exists to name, because the finding is a sign reversal in a rank correlation, not a magnitude shortfall against a cost.

Reversal condition:  none computable.

There is no cost, latency, or capital substitution that turns a correlation which changes sign between adjacent windows into a persistent one. The BL-D-002 section 4 test asks whether an arbitrarily favourable envelope would produce a positive result; here the envelope was never even reached, because the premise the envelope would have to operate on does not hold.

What would reopen it: a longer, uncapped fill history per wallet. The venue's userFills retrieval caps at roughly 10,000 recent fills per address, which truncated 369 of 800 candidates to a span under three days and forced the tested windows to be short. A different history source, specifically the venue's official monthly settlement archive referenced in the S3 family of measurements, carries fill history without that cap and would allow a genuine weekly persistence test rather than a days-long one. That is a different measurement, on data this programme has not pulled, and it has not been run.

What would not reopen it: loosening the wash-trade exclusion or widening the window definition on the existing sample. Both were fixed before the population was scored, and relaxing either after seeing a null result is the p-hacking path the design explicitly forbids.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/python/src/cohort_persistence.py, with 25 tests, as a worked example of a premise gate rather than as a live instrument. The candidate enumeration and the fill fetching are network calls against one venue's API shape and are not reproduced. It is not scheduled for reuse: the asset register marks this tool as an avenue that should not be re-run expecting a different answer without a change in the underlying data source.

Publishing it is worth more than publishing most positive results, because the shape generalises and the arithmetic is thirty lines. Before copying anybody, measure whether their performance persists. This gate spends no hypothesis budget under R-03 and it closed a branch that would otherwise have consumed a full one.

Two design choices in the published instrument carry the test.

Windows are cut on each wallet's own span rather than on a shared calendar. The question is whether a wallet's early performance predicts its own later performance, and a common calendar would put a newer wallet's entire history into one window and score the other two as zero. A test confirms a wallet active for a week and one active for three months both get a full set of scored windows.

The correlation is a rank correlation, not a correlation of values. Profit is scale-free across a population where one account trades a thousand times the size of another, and a value correlation would measure who is large rather than who persists. A test rescales one wallet by four orders of magnitude and confirms the verdict does not move.

The wash filter reads timing and nothing else: the median gap between profit-bearing fills. A wallet trading against itself submits both sides from one process, so the fills cluster while the sizes and profits look ordinary. Filtering on size would remove exactly the real large traders this premise is about. Fewer than five profit-bearing fills returns not-suspect rather than suspect, because too little evidence is not evidence of wash trading and defaulting to suspicion would strip every small wallet out of a population test.

12. Evidence register

This measurement is recorded in a narrative verdict report rather than as an emitted machine-readable artifact, and is registered as D-012 in BL-A-006.

13. Lessons extracted

The instrument itself produced a lesson before the strategy did. The first attempted run, at a 60-per-minute pull rate against the top 400 candidates, failed 335 of 400 fetches and was correctly discarded as an instrument failure rather than folded into a result. A slower run with retry logic produced zero fetch failures on 800 candidates. Treating a high failure rate as disqualifying for the instrument, rather than as noise to average over, is the same discipline BL-D-006 names for data integrity, applied to an API retrieval pipeline rather than to a market data stream.

The second lesson is about premise gates generally: this hypothesis is the cleanest example in the Hyperliquid programme of a gate designed to kill cheaply firing exactly as intended. No event study, no null test, and no cost model were needed to reach a verdict, because the gate asked the one question that had to be true before any of that machinery would matter, and the answer was no.

14. Cross-references

BL-SD-022 for the multi-confirmation hypothesis whose identity gate depended on this premise and was closed as a consequence of this falsification. BL-D-002 section 4 for the reversal-point test applied in section 10. BL-D-006 for the data-integrity discipline exercised in section 13. Ledger row S2.

Copyright (c) 2026 Bijective Labs. All rights reserved.