BL-SD-026: Adverse Selection Differs by Aggressor Identity

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Hyperliquid perpetuals Programme: I Verdict: CONFIRMED Blocking axis: not applicable, this is a fact rather than a hypothesis Transferability: not applicable, confirmed rows carry no transferability classification

1. Identification

This is the entry gate for the identity-aware maker family, and it was pre-registered as a fact test deliberately separated from any profit-and-loss question. The design's own kill criterion states the logic plainly: if adverse selection does not differ by the identity of the aggressor filling a resting order, then knowing that identity is useless for a maker, the strategy degrades to a generic maker, and the design is rejected before a single virtual fill is simulated. If it does differ, the fact becomes a necessary but not sufficient input to a separate PnL question, tested afterward in BL-SD-027.

2. Economic thesis

On a venue where every trade print exposes both counterparties, a resting maker can in principle know, at the moment of a fill, which class of participant just took its quote: Hyperliquid's own liquidation vault, a forced liquidation, a large account, or ordinary retail. The thesis under test is narrower than a trading strategy: does the price impact that follows a fill, the maker's adverse selection, actually differ across those classes, or is identity a distinction without a difference once trade size is accounted for.

If a maker quotes indiscriminately, the counterparty structurally forcing it to lose is any aggressor whose trade carries information the maker does not have, because the maker cannot decline a fill once it is queued. The question this gate answers is whether that structural loss is concentrated in one identifiable class of aggressor rather than spread evenly, because only a concentrated loss can be avoided by conditioning the quote on venue state.

3. Falsifiable claim

Pre-registered before any adverse-selection number was computed:

Adverse selection, measured as markout at 5 seconds from the maker's side, differs significantly between fills from large accounts and fills from retail accounts, with the ordering large at or above retail, and the effect does not vanish under a permutation null.

4. Mathematical specification

Notation per BL-D-011.
 
Aggressor and class, resolved at fill time from venue state (classifier G1, precedence
HLP > FORCED > LARGE(account value >= $250,000 asof) > RETAIL):
 
  aggr_i  =  user_a_i  if side_i == 'B'  else  user_b_i
  class_i =  classify_aggressor(aggr_i, t_i)
 
Aggressor direction (maker's resulting position, from the fill's own side flag):
 
  adir_i  =  +1  if side_i == 'B'   (buy, takes the ask; maker goes SHORT)
             -1  if side_i == 'A'   (sell, takes the bid; maker goes LONG)
 
Adverse-selection markout at horizon Delta, from the maker's side:
 
  AS_i(Delta) = adir_i * ( mid(t_i + Delta) - px_i ) / px_i * 1e4
 
             px_i = the print price (the maker's fill price)
             Positive AS = price moves toward the aggressor's direction = maker loses.
 
Primary horizon Delta* = 5 s. Secondary {1, 10, 30} s reported as robustness only.
 
Contrast statistic:
 
  delta_AS = mean{ AS_i(Delta*) : class_i = LARGE } - mean{ AS_i(Delta*) : class_i = RETAIL }
 
Inference:
 
  CI95(delta_AS)  via day-block bootstrap, 10,000 resamples
  p_perm          via label-permutation null, 10,000 resamples, shuffling class_i
                  across fills while holding timing and direction fixed

Causality holds by construction: class_i is resolved from venue state observable at t_i, and AS_i(Delta) looks forward only to t_i + Delta, which is a post-hoc markout used to characterise the fill's outcome rather than an input to any decision at t_i. There is no entry, exit, or sizing rule specified in this dossier, because this measurement characterises a microstructure fact; it does not place a trade. The virtual-fill PnL question that this gate feeds is specified separately in BL-SD-027.

5. Operating envelope

6. Instrument

BL-IS-006, test-driven, implements the markout calculation, the class-mean contrast, the bootstrap confidence interval, and the permutation p-value as pure functions. It reuses classify_aggressor from the orthogonality instrument built for the multi-confirmation programme.

At this sample size the contrast is well powered without waiting for further accumulation, which the pre-registration notes explicitly as a departure from the event-limited hypotheses elsewhere in this venue's record.

7. Results

Adverse selection by class, at the primary 5-second horizon:

Persistence across horizons, reported as robustness rather than as a separate decision:

The contrast does not decay after the primary horizon; it grows modestly and then holds flat through 30 seconds.

8. Verdict, blocking axis, and transferability

Verdict: CONFIRMED.

This is a venue fact, not a strategy, and carries neither an axis of death nor a transferability classification. What was confirmed: adverse selection experienced by a resting maker on Hyperliquid differs sharply and reliably by the identity of the aggressor that fills it. Large accounts impose more than three times the markout of forced-liquidation flow, and the gap between large accounts and retail, 0.906 basis points with a confidence interval entirely above zero and a permutation p-value of zero on 4.51 million fills, is one of the best-powered findings in this venue's record. The ordering matches the informed-participant thesis exactly: forced flow, which is price-insensitive by construction, carries the lowest adverse selection; large accounts, plausibly the most informed participants on the venue, carry the highest.

The persistence result strengthens the finding rather than merely confirming it. The contrast grows from 0.15 basis points at 1 second to roughly 1.0 basis point by 10 seconds and holds there through 30 seconds without reverting. A markout that reverts after an initial spike would be more consistent with a transient liquidity effect, a large trade temporarily walking the book and then recovering. A markout that persists without reverting is more consistent with informed trading whose impact on price is permanent, which is the reading this dossier adopts, with the caveat stated below.

Why it could not be monetised as measured. A microstructure fact that identity separates adverse selection is a necessary condition for a state-conditional maker strategy, not a sufficient one. The gate confirms that a maker who could quote selectively, avoiding fills from large accounts and welcoming fills from forced or retail flow, would face materially less adverse selection than a maker who quotes indiscriminately. It says nothing about whether the spread available on this venue is wide enough to cover even the lowest-adverse-selection class after fees. That question was tested directly in BL-SD-027, using a virtual maker-fill replay built specifically to answer it, and the answer there was no: gross returns were negative for every class of filler tested, including the most favourable one, because half-spread on this venue's majors is thinner than even the retail and forced adverse-selection figures reported here. The identity fact is real and would matter enormously on a venue or at a spread regime where the arithmetic worked in the maker's favour; on this venue, at this measurement, it did not.

One qualification on interpretation, carried forward honestly rather than smoothed over. The contrast between classes is confounded, in part, with trade size: a large account's trades are larger by definition, and larger trades produce larger price impact through mechanical market impact alone, independent of any informational content. The persistence of the contrast through 30 seconds is the argument offered against a pure size-impact reading, since mechanical impact from a single large trade would be expected to mean-revert as the book replenishes, and this contrast does not. But the confound is not separately decomposed in this measurement, and a reader should treat the informed-participation interpretation as the more likely reading rather than as fully isolated from a size effect.

9. Cost and capacity analysis

Not applicable. No trade PnL, fee, or capacity question was addressed by this gate; those belong to BL-SD-027, which used this confirmed fact as its precondition.

10. Envelope sensitivity and reopening conditions

Not applicable in the form used for FALSIFIED or CLOSED rows. This dossier records a fact about the venue's microstructure rather than a verdict that reverses under a cost or latency substitution. The relevant question for an outside reader is not what parameter would reverse this finding, but what execution structure would let an operator act on it: a spread regime, a fee schedule, or a venue where the half-spread available to a maker exceeds the adverse selection reported here for at least the lowest-adverse-selection class, forced flow at 0.10 basis points or retail at 0.32 basis points.

11. Reference implementation

BL-IS-006 specifies the markout instrument, published as 05_INSTRUMENTS/src/markout.py. The identity classifier this hypothesis rests on, from the dated re-grounding note of 2026-07-09, is published as 05_INSTRUMENTS/strategies/python/src/aggressor_identity.py, with 13 tests. It is shared with BL-SD-029 and published once rather than twice.

12. Evidence register

This measurement is recorded in the pre-registration ledger itself rather than as a separately registered evidence artifact, and is registered as D-020 in BL-A-006.

13. Lessons extracted

This is the strongest single microstructure finding in the Hyperliquid record by statistical power, 4.51 million fills with a permutation p-value of exactly zero, and it is also the clearest illustration in this venue's record of the difference between a confirmed fact and a monetisable edge. The design's own architecture, splitting the identity-aware maker hypothesis into an entry gate tested purely as a microstructure fact and a separate PnL screen tested only if the gate passed, is the reason this dossier can report a strong positive result honestly while the strategy built on top of it still failed. Conflating the two, reporting the gate's confirmation as if it settled the strategy question, would have overstated what was known at this stage by a wide margin.

14. Cross-references

BL-SD-027 for the virtual maker-fill PnL screen that used this confirmed fact as its precondition and failed on gross spread economics. BL-SD-029 for the aggregated-burst hypothesis that also drew on this identity classifier, on the taker side rather than the maker side. Ledger row S4-G1.

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