BL-SD-041: Positioning Drift, Directional

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Bybit perpetuals Programme: II Verdict: FALSIFIED Axis of death: Idea Transferability: UNIVERSAL

1. Identification

H-P2 exists to correct a closure rather than to open a new question from scratch. An independent audit found that the analytical conversion used to close a leading directional signal, referred to internally as Door 1, had applied the standard deviation of a single instrument to an information coefficient averaged across 154 symbols, a 2.6-fold error that reversed the door's status from closed to unresolved. This dossier is the exact-PnL test that settles what the corrected arithmetic alone could not.

2. Economic thesis

When open interest accumulates rapidly, the marginal positions are opened by the most recent entrants, at the worst price, with the thinnest margin buffer, closest to the liquidation threshold. They do not choose when they exit; the exchange's margin engine chooses for them, without regard to price. The mirror side carries its own claim: when open interest contracts rapidly, the forced selling has already finished, the coerced sellers are already out of the book, and the pressure that depressed price departs with them.

Who is forced to lose. Marginal leverage holders in the accumulation phase. Why structural. The compulsion is written into the exchange's margin rules, not into anyone's preference.

The author's prior, stated before any PnL was computed and reproduced here in full: H-P2 is likely dead, for two reasons stateable without reference to data. H-P1 died on a closely neighbouring axis, finding that crowding predicts width rather than direction with the same positive sign on both sides; if the same underlying phenomenon drives oi_z, its information coefficient is more likely an asymmetry artifact than a harvestable direction. And the measured information-coefficient-to-PnL degradation factor of 5.6 to 13.7 times, if it applies here, would pull an apparently above-cost signal back below it. The hypothesis was tested anyway because the cost-to-noise ratio at the 72-hour horizon was the best the programme had yet measured, and because a door could not remain recorded as closed on arithmetic already shown to be wrong.

3. Falsifiable claim

Pre-registered before one PnL figure was computed:

A single-signal directional bet on the 72-hour open-interest z-score, shorting extreme accumulation and buying extreme contraction, produces a mean net return with a CI95 excluding zero at 11.0 basis points cost, a positive point estimate at 22.0 basis points, and a mirror test whose two sides carry opposite-signed raw forward returns.

The mirror-test criterion is written directly from the H-P1 lesson: a same sign on both sides of a conditioning variable means width or common drift is being measured, not direction, whatever the significance level.

4. Mathematical specification

Notation per BL-D-011. All rolling statistics causal, trailing, exclusive
of bar t.
 
signal:  oi_z_24h[t] = trailing_zscore(log_change(OI, 24h), 720 bars)
                       (720 bars = 30 days at 1h resolution)
 
entry:   oi_z_24h[t] >= +2  -> SHORT at open[t+1]   (against accumulation)
         oi_z_24h[t] <= -2  -> LONG  at open[t+1]   (coercion already spent)
 
exit:    open[t+1+H],  H = 72h primary, 24h reported as robustness only
dedup:   one event per symbol per H hours, first wins
 
sizing:  three bases, all reported, capped_notional (cap 20 concurrent
         positions) designated primary in advance on the finding that
         equal_weight is an undeclared convention that concentrates
         exposure on the fewest concurrent signals (BL-SD-039 section 4);
         cap of 20 fixed from the data's own p90 concurrent-leg count,
         not from any H-P2 outcome
         PASS requires a positive point estimate on all three bases
 
cost:    11.0 bps round trip, baseline (measured taker RT, not the 2x
         pessimistic multiple H-R1 used, because entry here is not
         conditioned on volatility and does not systematically fall on a
         widened-spread bar); 22.0 and 33.0 bps as sensitivity

Causality holds by construction: the z-score at bar t is computed on a trailing 720-bar window exclusive of t, and entry executes at the next bar's open rather than the signal bar's close.

The cost baseline departs from BL-SD-039's 22 basis points, and the pre-registration states the mechanical reason before any result was seen: H-R1's entry falls, by construction, on the most fragile bars, where spread widens; H-P2's entry is a slow-moving positioning state, unconditioned on volatility, so doubling the round trip would penalise a risk this signal does not carry. To close the gap this leniency opens, the pass criterion requires a positive result at both cost levels rather than one.

5. Operating envelope

The cost-to-noise ratio at this horizon, 11 basis points against a measured panel standard deviation of approximately 960 basis points at 72 hours, is 1.1 percent, the best the programme had measured to that point. That favourable ratio is why the hypothesis was worth testing despite a stated negative prior; it does not appear in the verdict, because the mechanism failed before the ratio became relevant.

6. Instrument

The positioning-drift test, built on the same tested direction-selection and PnL modules as H-R1 and H-R2 (rndv2_backtest.select_reversion_events, net_bps), with no sign logic rewritten for this hypothesis.

The confidence interval width is worth stating on its own. At plus or minus 2.3 basis points, this instrument can exclude an edge as small as 4 basis points, sharper than H-R1's roughly plus or minus 18.3 basis points under its originally reported sizing basis. The death recorded in section 7 is not a power failure; the interval is narrow and it is centred near zero.

7. Results

Primary horizon, 72 hours, n = 12,329 collapsed events:

Cost sensitivity, primary basis:

Subperiods: -0.2, -0.3, +3.1 basis points, only 1 of 3 positive. Robustness at 24 hours, primary basis: +0.08, CI [-1.24, +1.36], closer still to zero.

The mirror test, the pre-registration's most diagnostic requirement, fails:

Both sides are positive. The mechanism requires opposite signs; what is measured instead is a magnitude difference on the same sign, +1.4 against +6.6, a gap of roughly 5.3 basis points, which is what the information coefficient of -0.0188 was actually detecting. A rank-based, unsigned statistic cannot distinguish "opposite direction" from "same direction, different size," and this is the concrete case where that limitation produced a false impression of a directional signal.

Per-leg PnL confirms the same defect from the trading side:

The entire result is carried by the LONG leg; the SHORT leg, which carries the premise that accumulation forces a directional decline, loses money. Selecting the LONG leg alone after seeing this would be the same after-the-fact selection the record already flags as illegitimate in the liquidity-tertile diagnostic under H-R1, and the LONG leg's own interval spans zero regardless.

8. Verdict, axis of death, and transferability

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

The kill criteria fire on cost (CI95 contains zero at the sharpest interval the programme had produced) and, more decisively, on mechanism: the mirror test fails outright, and the leg carrying the hypothesis's own stated cause loses money on its own. Both sides of the conditioning variable show a positive raw forward return, which is the same signature H-P1 produced with skewness, now reproduced with mean return on a different signal, a different horizon, and a different statistic. The common explanation is a shared positive drift component in the research-era sample that a per-symbol directional bet cannot separate from the signal itself: the LONG leg benefits from that drift, the SHORT leg is penalised by it, and the signal drowns between them.

This is a construction failure rather than a magnitude failure, which is why it lands UNIVERSAL rather than ENVELOPE-BOUND: no cost substitution changes which leg of a directional bet a shared market drift helps and which it hurts. The failure mode is specific enough to be named and, unlike the prior two entries in this family, points to a concrete alternative construction, cross-sectional and market-neutral, that would separate the signal from common drift by design.

9. Cost and capacity analysis

Mean net, primary basis, 72h, 11 bps    +0.94 bps, CI [-1.36, +3.33]
CI half-width                            approximately 2.3 bps
Edge excludable at this power             above approximately 4 bps

The interval is the narrowest in the programme to this point, which makes the null result informative rather than inconclusive. Capacity was not separately measured; the hypothesis never reached a stage where capacity was the binding question.

10. Envelope sensitivity and reopening conditions

No binding parameter exists to name, because the mirror test fails outright rather than the magnitude falling short.

Reversal condition: none computable.

A cost substitution cannot repair a mirror-test failure. The measured defect is that a shared positive drift component enters both legs of a directional construction and cannot be separated from the signal by adjusting price paid per trade. Setting cost to zero leaves both raw forward returns positive and the SHORT leg still losing money relative to the LONG leg's drift exposure.

What would reopen it: a cross-sectional, market-neutral construction that differences the SHORT and LONG legs against each other rather than against zero, removing the shared drift by design. This is not a speculative suggestion; it is the axis the programme's own decisions document opened formally and had not yet run at the time of this verdict, and it is the construction BL-SD-042 through BL-SD-045 test directly.

What would not reopen it: selecting the LONG leg alone, or re-running the same construction at a different horizon. The mirror test's failure is a construction defect, not a horizon-specific artifact, and the 24-hour robustness check moves the point estimate closer to zero rather than away from it.

11. Reference implementation

The direction-selection and PnL modules this row shares with BL-SD-039 and BL-SD-040, without modification, are published: 05_INSTRUMENTS/strategies/python/src/conditioned_reversion.py, 05_INSTRUMENTS/strategies/python/src/exact_pnl_backtest.py, 05_INSTRUMENTS/strategies/python/src/validation_stats.py, 05_INSTRUMENTS/src/causal_stop.py and 05_INSTRUMENTS/strategies/python/src/exit_geometry.py. The positioning-drift test driver itself remains scheduled.

The sharing claim is the reason that split is defensible rather than convenient. Every quantity this row's verdict rests on -- how a direction is chosen, how a leg is priced, how legs are collapsed into a portfolio return, how an interval is put around it -- comes from code that is published and tested, and none of it was modified for this hypothesis. What is scheduled is the loop that selects this row's events and calls those functions in order.

A reader who disagrees with this verdict will disagree about one of the published pieces.

12. Evidence register

13. Lessons extracted

This is the third instrument, after H-P1's skew test and H-R1's mean-return test, to reach the same diagnosis by a different statistical route: a per-symbol directional construction cannot separate a genuine signal from a shared, common drift entering both sides of its own conditioning variable. An information coefficient, being rank-based and unsigned, is structurally unable to detect this defect on its own, since it does not distinguish "opposite direction" from "same direction, different magnitude." Both findings sharpen Rule R-11's justification: verdicts must come from signed, exact PnL, never from a correlation statistic, and this dossier is a case where the correlation statistic would have reported a real number computed from an artifact.

The pre-registration's cost-baseline reasoning, that a signal not conditioned on volatility does not need the 2x pessimistic multiple a fragility-conditioned entry does, held up under sensitivity testing at 22 and 33 basis points without changing the verdict, and is retained as a transferable principle for future hypotheses that similarly avoid entering into widened-spread conditions by construction.

14. Cross-references

BL-SD-037 for the H-P1 mirror-test failure this hypothesis's own prior was built against. BL-SD-039 and BL-SD-040 for the shared instrument lineage. BL-SD-042 for the cross-sectional, market-neutral construction this failure mode motivates directly. BL-D-002 section 4 for the UNIVERSAL classification test. BL-SD-000 ledger row H-P2.

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