BL-SD-029: Aggregated Informed-Flow Bursts, Directional

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Venue: Hyperliquid perpetuals Programme: I Verdict: CLOSED Axis of death: Cost Transferability: ENVELOPE-BOUND

1. Identification

Pre-registered and settled the same day. The instrument already existed, having been built for a prior hypothesis in the same family, so the measurement was a matter of hours rather than weeks.

2. Economic thesis

Informed participants do not transact continuously. They arrive in bursts, because information arrives in bursts and because working a large position generates a cluster of executions rather than a single one.

If a burst of aggressive flow from participants classified as large can be detected as it happens, and if that flow carries information rather than merely size, then price should continue in the direction of the burst for as long as it takes the market to absorb what the participant knows.

Who is forced to lose. The passive side of those trades. A resting order filled by an informed aggressor is filled at a price that is about to be wrong, and the maker cannot decline because the fill has already happened. That is a structural loss rather than a skill difference, and it is the same effect the predecessor hypothesis measured directly.

3. Falsifiable claim

Pre-registered before measurement:

Aggregated directional bursts from large aggressors are followed by price drift in the burst direction, of sufficient magnitude to clear the taker round trip at tier zero.

The third and fourth criteria are the ones that make this a test of the thesis rather than of a signal. If the effect is real and driven by informed participation, restricting to larger and more clearly informed bursts should strengthen it.

4. Mathematical specification

Notation per BL-D-011.
 
Burst aggregation over a trailing window w:
 
  V_t     =  sum over trades i in (t-w, t] of  q_i * eps_i * 1{class_i = LARGE}
 
            q_i     signed notional of trade i
            eps_i   +1 if the aggressor bought, -1 if sold
            class_i aggressor classification, resolved from venue state at t
 
  sigma_t =  standard deviation of V over the trailing L observations ending at t
 
  s_t     =  V_t / sigma_t
 
Entry:
 
  fire at t when  |s_t| >= theta
  direction  d_t = sign(s_t)
 
Exit:
 
  at t + h, h = 30 s, fixed. No discretionary exit, no stop, no target.
 
Sizing:
 
  equal notional per event
 
Cost:
 
  c_rt = 2 * fee_taker = 9.0 bps at venue tier 0
 
Net return per event:
 
  r_net(t) = d_t * ( p_{t+h} - p_t ) / p_t * 1e4  -  c_rt

Causality holds by construction. V_t aggregates only trades at or before t; sigma_t is estimated on the trailing window ending at t; class_i is resolved from venue state observable at t. Every term entering s_t is measurable with respect to F_t.

The exit is fixed rather than mechanism-bound, which is a departure from BL-D-007 rule 9 and is acknowledged. The thesis does not name a condition under which absorption is complete, so no mechanism-bound exit was available. A fixed horizon was chosen in advance and not varied, precisely so that horizon shopping could not occur afterwards.

5. Operating envelope

Crossing on both legs is the conservative assumption and it was chosen deliberately. A burst-triggered entry has no time to rest an order, so the entry leg is a taker fill by construction. The exit leg could in principle be passive, and section 10 treats that case.

6. Instrument

Measurement ran on the identity-classified trade record from the venue's shadow recorder, using the aggressor classifier built for the predecessor hypothesis and its accompanying test suite.

The block bootstrap is required rather than optional here. Bursts cluster within days, so events are not independent draws and per-event intervals would overstate significance.

7. Results

Drift at the 30-second horizon, before costs:

The signal is real on all three instruments and well powered on two.

Net of the round trip:

The interval does not merely span zero. It lies entirely below it.

Selection response, the test that distinguishes an informed-flow effect from a size effect:

Flat. Restricting to larger and more clearly informed bursts does not strengthen the drift, which is not what an informed-participation thesis predicts.

Attribution:

Roughly three quarters of the drift is explained by the size of the trade rather than by who placed it.

8. Verdict, axis of death, and transferability

Verdict: CLOSED. Axis: cost. Transferability: ENVELOPE-BOUND.

The signal exists. Drift of +3.04 basis points at a sign-permutation p-value of zero on 486 events is not a marginal detection, and it survives on two further instruments. What defeats it is a 9.0 basis point round trip against a 3.04 basis point effect.

The classification requires the test in BL-D-002 section 4: if every envelope parameter were arbitrarily favourable, would this produce a positive expectation?

At zero cost, yes. The drift is positive, significant, and consistent in sign across three instruments. The hypothesis is therefore envelope-bound rather than falsified.

Two qualifications belong on that classification and both weaken it.

The identity component is not significant on its own. What clears at a lower cost is a volume-impact effect with an identity flavour, not the informed-participation thesis as stated. A reader reproducing this should expect to be trading price impact.

The selection response is flat, which is evidence against the mechanism even though it is not a cost finding. A stronger version of the thesis would have produced a stronger signal under tighter selection, and it did not.

9. Cost and capacity analysis

Gross drift, ETH, 30 s          +3.04 bps
Round trip, tier 0 taker         9.00 bps
Net                             -5.96 bps

Capacity was not measured. At the sample rate observed, roughly one event per instrument per hour on the majors, and at the declared capital scale, capacity was not the binding constraint and no measurement was taken. A reader operating at size should treat capacity as unmeasured rather than as adequate.

10. Envelope sensitivity and reopening conditions

Binding parameter: round-trip cost.

Reversal condition   c_rt  <  3.04 bps

The available channels on this venue at tier zero:

The cheapest reachable channel leaves +0.04 basis points, which is inside the noise of the measurement and is not an edge.

This is the honest reading and it matters more than a clean reversal point would. The hypothesis does not clear at any tier-zero channel. It clears only below tier zero, which on this venue requires the maker rebate at 500,000,000 USD of fourteen-day volume.

There is a second obstacle that a cost substitution alone does not remove. The maker-to-maker column assumes both legs fill passively. The entry leg cannot: the signal is a burst, and by the time it is detected there is no time to rest an order and no reason to expect it filled at a favourable price. Realistically the entry leg is a taker fill, which puts the floor at 6.00 basis points and the net at -2.96.

Reopening conditions, stated precisely:

A round trip below 3.04 basis points achieved with a passive entry leg, which requires either a rebate tier or a venue with a materially different schedule.

Or a demonstration that the drift is larger at a horizon other than 30 seconds. This was not tested and cannot be tested on this data without horizon shopping, which R-12 prohibits. A reader would need to pre-register the horizon and use data this programme has not seen.

What would not reopen it: tighter selection. That was tested and the response was flat.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/python/src/identity_burst.py, with 20 tests. The aggressor classifier it depends on is published beside it as 05_INSTRUMENTS/strategies/python/src/aggressor_identity.py, with 13 tests. The classifier resolves participant class from venue state and is shared with the predecessor hypothesis in BL-SD-026, which is why it is a module of its own rather than a private function here.

12. Evidence register

The measurement for this hypothesis is recorded in the venue verdict report rather than as an emitted artifact. It is registered as D-022 in BL-A-006.

13. Lessons extracted

This hypothesis is the clearest argument in the archive for the layer-zero cost check, and it appears as the worked example in BL-D-008 section 2.

Everything about the measurement was correct. The instrument was tested, the inference was correlation-aware, the null was matched, the sample was adequate, and the finding was real. The answer was still available on the first day for the price of comparing a plausible drift magnitude against a published fee schedule.

The attribution result produced a second lesson that did not become a numbered rule but should have. A hypothesis naming a specific cause, here participant identity, must attribute the effect to that cause rather than merely observing the effect in its presence. Three quarters of this drift was volume impact, and without the attribution arm the record would have carried an identity finding it had not earned.

14. Cross-references

BL-SD-026 for the identity fact this hypothesis attempted to monetise. BL-D-008 section 2, the cost-first doctrine worked example. BL-D-002 section 5 on computing a reversal point. BL-P1-004 section 5 for the narrative. Ledger row IB1.

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