BL-SD-018: Large-Cap Flow Leading Small-Cap Price

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-SD-018: Large-Cap Flow Leading Small-Cap Price

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Venue: Bybit Programme: I Verdict: FALSIFIED Axis of death: Data Transferability: ENVELOPE-BOUND

1. Identification

Field	Value
Internal code	leadlag_probe, later leadlag_recorder
Family	Parent-to-child lead-lag, large-cap flow ahead of ecosystem alt price
Venue	Bybit perpetuals
Programme	I
Pre-registered	2026-06-27, the probe's own design comment, written before it was run
Verdict issued	2026-06-27
Pre-registration document	The lead-lag probe's own design comment in its source header, written before it was run
Predecessor	None

2. Economic thesis

A large-cap ecosystem anchor, SOL or ETH, moves first; same-ecosystem alt children follow with a lag and a higher beta, because information about the parent's move takes time to reach and be priced into thinner child order books. The edge, if any, is the width of that lag relative to this programme's own reaction speed.

Who is forced to lose. Not named in the source. The design frames the question as an empirical one about lag width rather than as a mechanism with a named, structurally compelled counterparty.

3. Falsifiable claim

Reproduced from the probe's own design comment, written before it was run:

A cross-correlation lag profile peaking at lag zero is contemporaneous, already arbitraged below one minute, and is not a slow edge at this speed. A profile peaking at a lag of one or more, confirmed by an event study in which entering the child after observing the parent's move beats a matched random-time null and clears the round trip, is a capturable lagged response.

Metric	Threshold	Decision if crossed
Cross-correlation peak lag	at lag 0	Contemporaneous; no slow edge at this resolution
Event-study NET vs random-time null	does not beat the null or is not positive	No capturable response confirmed

4. Mathematical specification

Notation per BL-D-011.
 
logret(close)      = diff( log(close) )
 
cross-correlation:  ccf(pr, cr, lag) = corr( pr_t , cr_{t+lag} )
                     positive lag means the child follows the parent
 
event study, causal:
  z_t = ( pr_t - mean(pr) ) / std(pr)
  fire when |z_t| >= 2.0
  enter the child at t+1 (AFTER the parent's move at t is observed), direction = sign(pr_t)
  cret_H  = sum( cr_{t+1 : t+1+H} ) * 1e4,   H in {1, 3, 5} minutes
  real_ret = direction * cret_H - c_rt
  null: same count, random entry time t', same-time parent sign, same cost
  c_rt = 11.0 bps, taker to taker

Causality holds by construction: the child entry occurs at t+1, strictly after the parent's move at t is observed, never on the same bar the move is detected.

5. Operating envelope

Parameter	Value for this measurement
Round-trip cost assumed	11.0 bps, taker to taker
Execution channel	Crossing on both legs of the child position
Latency band	Cold path; the probe reads one-minute OHLCV over REST
Measurement cadence	One-minute bars, approximately 1,000-bar fetch limit per symbol
Capital scale	Programme default, `BL-FM-002` section 6
Venue and instruments	Bybit USDT perpetuals; parents SOLUSDT and ETHUSDT against 8 ecosystem children
Data era	Approximately 16 hours of one-minute history, 2026-06-27

A follow-on instrument, built the next day to test the same question below one-minute resolution, is described in section 10 rather than in this section, because it never produced an analyzed result within the programme's runway.

6. Instrument

The lead-lag probe, self-described in its own header as a probe on a small sample, a directional indication rather than a verdict. No archived output file with the per-pair cross-correlation values or event-study net returns was located for this dossier; the only surviving record of the result is the one-line qualitative summary in the Bybit post-mortem: the adjustment is contemporaneous at one-minute resolution. That absence is reported here rather than reconstructed, and it is one of the two figures this dossier could not corroborate with an archived number.

7. Results

Arm	n	Finding	Threshold	Result
Cross-correlation lag profile, all parent-child pairs	approximately 16 hours of 1-minute bars	Peak correlation at lag 0	Peak at lag 1 or more required for a slow edge	FAIL, contemporaneous
Event study, NET vs matched null	not preserved as an archived figure	Narrative record states no residual capturable response at this resolution	NET positive and beats the null	Not corroborated with an archived number; the narrative conclusion is negative

8. Verdict, axis of death, and transferability

Verdict: FALSIFIED. Axis: data. Transferability: ENVELOPE-BOUND.

At one-minute resolution, any adjustment between a large-cap parent and its ecosystem children completes inside the same bar, faster than this instrument can resolve. That is a negative result at the resolution actually tested, and it is reported as one. It is not a claim that no lag exists at a finer resolution, because this substrate cannot see one if it does. The axis is data, specifically measurement resolution, rather than idea, and the transferability classification follows BL-FM-002 section 5's distinction between a recorder that has falsified a phenomenon and one that has failed to resolve it.

9. Cost and capacity analysis

The round-trip cost of 11.0 basis points is immaterial to this result; the finding is a resolution limit, not a cost limit. Capacity was not measured and the question did not reach it.

10. Envelope sensitivity and reopening conditions

Binding parameter: measurement resolution, one-minute bars in the probe that produced this verdict.

Reversal condition:  tick-level or sub-second capture of the same parent
                      and child instruments, resolving a lag this probe
                      cannot see.

A follow-on instrument for exactly this purpose was built the next day, the tick-level lead-lag recorder, bounded to the substrate's own publish cadence of approximately 33 milliseconds for active parents and 140 milliseconds or more for small children, sufficient to resolve a lag from roughly 50 milliseconds to several seconds. It accumulated 2,000 bounded event summaries and an 856 megabyte raw lake before being stopped on 2026-07-14, for runway reasons stated at the time as the consumer not being scheduled within the remaining window, not because the recorder failed. The data is preserved and the decision is described as reversible.

What would reopen it: analysis of the existing tick-resolution lake, or a fresh capture, at sub-minute resolution.

What would not reopen it: re-running the one-minute probe. That resolution already produced its answer.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/python/src/lead_lag.py, with 19 tests, carrying the probe and the tick-level recorder. The runner harness is transport and is not reproduced. An earlier, unbounded version of the recorder's cache sink caused a 38-hour outage of the Bybit engine through an out-of-memory crash loop before being bounded. The bound is in the published module and the comment beside it says why it is there, because a bound whose reason is not recorded is a bound somebody removes.

12. Evidence register

Related artifact	Carries
None assigned	Neither the one-minute probe's output nor the tick-resolution lake has been registered as a machine-readable artifact in `BL-A-001`

13. Lessons extracted

The unbounded-sink outage is a durable operational lesson independent of the hypothesis: a list-type Redis sink for a research instrument must be bounded from its first deployment, not retrofitted after a crash loop. This generalises the data-integrity discipline in BL-D-006.

A second, quieter lesson belongs here: a dataset can be collected honestly, at real infrastructure cost, and never analyzed within a runway. That is reported as what happened, not folded into a false claim of falsification at the resolution the recorder was built to test.

14. Cross-references

BL-FM-002 section 5 for the cadence and coverage distinction this dossier applies directly. BL-D-006 for the data-integrity protocol relevant to the unbounded-sink incident. BL-SD-000 ledger row LEADLAG, and its envelope-bound entry naming measurement resolution as the binding parameter.