BL-SD-019: Tick-Exact Dislocation Fade

Status: WITHDRAWN

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Bybit Programme: I Verdict: WITHDRAWN Axis of death: Data Transferability: -

1. Identification

This entry exists to keep the retraction itself in the record rather than letting it disappear along with the verdict it replaced. The row is instructive precisely because it shows a rejection failing for the same reason a validation would have: an uninspected sample.

2. Economic thesis

A price prints away from where the rest of the book and the recent tape agree it should be, on a burst of volume well above the local baseline. That gap is a dislocation: a local, transient repricing driven by a temporary supply and demand imbalance rather than by new information. If the imbalance is temporary, the price should fade back toward the pre-event level once the imbalance clears.

Who is structurally forced to lose. The side that created the dislocation by consuming the book faster than it could be replenished. A market order large enough to walk several levels pays the average price across those levels, not the best price, and the gap between the average paid and the level the book snaps back to is a real, structural cost paid by whoever caused the move. The fade strategy is a claim on the other side of that cost.

3. Falsifiable claim

Pre-registered before the first measurement:

A displacement of at least min_disp_bps basis points, accompanied by a volume spike of at least min_vol_ratio times the trailing baseline, is followed by a partial reversion of sufficient magnitude to clear the taker round trip, and the reversion strengthens with the size of the displacement and with the liquidity tier of the instrument.

The fourth row was not part of the original kill sheet. It is added here because the corrected verdict used it as supporting evidence and a reader should be able to see it applied the same way the first three were.

4. Mathematical specification

Notation per BL-D-011.
 
Displacement, in basis points, over a reference price p0 observed at or before t:
 
  disp_t   =  (p_t - p0) / p0 * 1e4
 
Volume spike ratio, trailing window w against a trailing baseline:
 
  vspike_t =  V_t(w) / V_baseline(w)          V_baseline estimated on trailing history at or before t
 
Dislocation flag:
 
  is_disloc_t  =  |disp_t| >= min_disp_bps   AND   vspike_t >= min_vol_ratio
 
Fade direction:
 
  d_t  =  +1   if disp_t < 0   (price dropped, fade by buying)
  d_t  =  -1   if disp_t >= 0  (price spiked, fade by selling)
 
Entry:
 
  taker sweep of the book, target_qty walked level by level, best price first,
  average fill price avg_fill computed from the levels actually consumed
 
Exit:
 
  fixed horizon h after entry, no discretionary exit
 
Sizing:
 
  bounded by capacity_within_bps: the quantity absorbable within a declared
  band of the reference price, read from the live ladder, never from a
  notional target assumed available
 
Cost:
 
  slippage_bps  =  d_t * (avg_fill - p_ref) / p_ref * 1e4          entry leg
  c_rt          =  2 * fee_taker  +  slippage_bps                  round trip
 
Net return:
 
  r_net(t)  =  d_t * (p_{t+h} - avg_fill) / avg_fill * 1e4  -  c_rt

Causality holds by construction for the detection terms: disp_t and vspike_t use only the reference price and volume observed at or before t, and the fade direction is a deterministic function of disp_t. The entry model is not a signal term but an execution model, and it is included here because it is the term that ultimately dominated the verdict: avg_fill is computed by walking the actual book depth rather than assumed at the touch, and it is this term, not the detection logic, that carried the finding in section 5.

5. Operating envelope

The universe selection rule is the defect this dossier exists to document, and it is stated here in full rather than abbreviated, because the rule itself is the finding: a selector built on move count actively prefers an instrument that moves constantly, which is exactly the profile of a low-grade microcap running on thin liquidity rather than of a genuine transient dislocation.

6. Instrument

The dislocation detection module, its tick-exact mirror used in the hot path, and the offline evaluation and scan tools. The hot-path mirror carries unit coverage on displacement, volume ratio, the dislocation flag, fade direction, book-walking, slippage, and absorbable capacity within a band.

The system-health check was run first and passed cleanly, which is precisely what allowed the original rejection to look sound. A healthy container was mistaken for a valid measurement; those are two different claims, and the distinction is the lesson this family produced.

7. Results

Gross reversion, before the universe was decomposed:

The sign was correct and the magnitude looked usable. Neither fact survived decomposition by instrument.

Net return under a live-exact cost model, by slippage assumption:

Per-bucket sample size after decomposition, against the pre-registered gate of 200 per bucket:

One bucket in the clean subset clears the gate. No bucket in the liquid subset does. The bootstrap confidence intervals degenerate at or above roughly 150 events per bucket in this record, which further narrows what can be read from the buckets that do clear.

Monotonicity, the supporting evidence carried forward rather than discarded:

Reversion strengthens with liquidity tier, which is what the thesis predicts and is the reason this entry is withdrawn rather than falsified.

8. Verdict, axis of death, and transferability

Verdict: WITHDRAWN. Axis: data. Transferability: not applicable.

The original verdict, REJECTED, is invalid and is not restated here as a finding. It was computed on a sample that was 87 percent a single low-grade instrument, and a rejection built on a contaminated sample is not evidence against the thesis; it is evidence against the sample.

The corrected reading is INCONCLUSIVE rather than either accepted or falsified terminology, and it is recorded as WITHDRAWN at the strategy level because the artifact that existed, the original verdict, was pulled rather than revised in place. No transferability classification applies: transferability answers whether a closed result would hold under a different envelope, and nothing here closed.

What the corrected measurement did establish, and what survives the retraction, is threefold. Forced-flow reversion is real, with the correct sign on the full undecomposed sample. It is liquidity-monotone, strengthening from +4.2 basis points on the undecomposed sample to +19.7 on the liquid, large-displacement tier. And it is latency-sensitive to a degree that dominates everything else: at zero slippage the edge is positive, and at any slippage a real fill could plausibly incur, it inverts.

9. Cost and capacity analysis

The binding cost term is not the fee. It is entry slippage from walking a book that has already been thinned by the event the strategy is trying to fade. Break-even requires entry slippage below approximately 7 percent of the displacement being captured. That bar was computed across the slippage sensitivity sweep in section 7 and is the single number that should govern any future investment in this family's execution path.

Capacity was not separately measured beyond the absorbable-quantity model in section 4. The liquid-tier bucket, where the reversion is strongest, is also the bucket with the fewest events, at n=104, 17, and 3 across the three displacement bands, which caps what any capacity estimate on that tier could claim with confidence.

10. Envelope sensitivity and reopening conditions

Binding parameter: entry slippage, expressed as a fraction of the displacement.

Reversal condition   slip_entry  <  approximately 0.07 * disp_t

This is a latency requirement stated in cost terms rather than a milliseconds figure, because the poll cadence that produced this data, 2 seconds, was itself part of what generated the slippage. A materially faster, event-driven capture path changes the number this reverses at; this record does not have that data and does not estimate what it would show.

A second, independent condition sits upstream of the cost one and must clear first: the universe selector must exclude low-grade instruments at the source. Repeating this measurement on the existing selector reproduces the contamination regardless of what is done to the cost model, because the selection rule itself prefers the contaminating profile.

Reopening conditions, stated precisely:

A universe registry with an explicit instrument list, a liquidity floor, and a hard denylist, replacing the move-count selector, run until the clean and liquid buckets each independently clear 200 events. And an entry path capable of holding realised slippage below roughly 7 percent of displacement, which on this data required event-driven capture rather than 2-second polling.

What would not reopen it: re-running the existing selector for a longer window. Volume alone does not fix a selection bias; it accumulates more of the same contaminant.

11. Reference implementation

The detection module is published as 05_INSTRUMENTS/strategies/python/src/dislocation.py, with 20 tests. The tick-exact hot-path mirror is engine code and the scan and evaluation drivers are transport; neither is reproduced.

The universe registry these instruments consumed required the correction described in section 5 and BL-IN-004 before any future measurement is trustworthy, and that is why this row is WITHDRAWN rather than falsified: 87 percent of its 39,250 events came from one symbol. The detector was correct. What fed it was not, and publishing the detector is how a reader can confirm which of the two failed.

Two properties of the published module are pinned. The trigger is an AND of displacement and volume spike, because displacement without volume is a repricing, which is informed flow and does not revert. And entry is taken a declared latency after detection at the real tick price then, with the instant displaced price reported beside it as the optimistic bound the earlier scan had used; a test confirms the realistic fill is worse.

12. Evidence register

13. Lessons extracted

This is the entry BL-IN-004 was written about, and the incident record carries the full narrative. The lesson repeated here because it belongs in the strategy's own dossier and not only in the incident book: a rich sample is not a clean sample, and the check that catches the difference, decomposing by instrument before trusting a verdict, must run before a rejection is accepted just as strictly as before an acceptance is. R-19, universe hygiene is signal quality, was produced from this entry.

A second point belongs here specifically because it did not become a numbered rule. A healthy container, zero errors, a working deduplication path, and a correct-sign gross measurement together looked like sufficient grounds for a verdict. They were sufficient grounds for trusting the system, not for trusting the sample the system had been fed.

14. Cross-references

BL-IN-004 for the full incident narrative this dossier summarises at strategy level. BL-D-009 rule R-19. BL-D-004 section 5 on verdict vocabulary and the distinction between a settled kill and an unresolved interval. Ledger row DISLOC, status WITHDRAWN.

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