BL-SD-039: Conditioned Reversion at 24 Hours

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Bybit perpetuals Programme: II Verdict: CLOSED Axis of death: Cost Transferability: ENVELOPE-BOUND

1. Identification

H-R1 is the programme's third hypothesis and the first to separate characterisation from verdict as a matter of written rule rather than practice. Its pre-registration states plainly, and this dossier repeats without softening: at the time of writing, zero PnL, net, or profit code had ever been run against this signal. Only the information coefficient, a correlation statistic, had been seen.

2. Economic thesis

Sharp price moves during fragile market conditions are driven mostly by forced liquidation and liquidity withdrawal, not by new information. The party pushing price in that moment has no choice: a leveraged position closed by the margin engine, or a market maker pulling quotes because inventory risk has spiked. When the force is exhausted and liquidity returns, price reverts. In calm conditions, the same magnitude of move more often carries genuine information, and information does not revert.

Who is forced to lose. Liquidated leverage holders, and anyone who must execute while book depth is thin. Why structural. The compulsion is written into exchange margin rules and into a market maker's own risk limits, not into anyone's preference.

The prediction that distinguishes this from a pattern rather than a mechanism is sharp: reversion should be far stronger under fragile conditions than calm ones, and the separation measured at the characterisation stage, before any PnL was computed, already carried the correct sign and shape.

3. Falsifiable claim

Pre-registered before one PnL figure was computed:

Conditional on a 24-hour lookback z-score crossing 2 standard deviations against the realised move, entered contrarian, under a causally trailing fragility gate above the 80th percentile, mean net return at the 24-hour horizon is positive with a CI95 that excludes zero at the pessimistic baseline cost.

The baseline cost of 22.0 basis points is deliberately double the measured taker round trip of 11.0 basis points, adopted as a rule after the programme's second hypothesis demonstrated that a favourable assumption finds its way into a result whenever it is not locked in advance.

4. Mathematical specification

Notation per BL-D-011. All rolling statistics causal, trailing, exclusive
of bar t.
 
signal:  z4[t] = log_change(close, 4h) / (realized_sigma_1h[t] * 2)
 
fragility[t] = composite(turnover_z, realized_vol_pct, vol_of_vol_z,
                          oi_abs_chg_z), all inputs causal at t
 
condition:  trailing_percentile(fragility, 2160 bars) > 80
            (2160 bars = 90 days at 1h resolution; trailing, not
            full-sample, verified by audit not to differ materially:
            causal -0.0579 IC versus look-ahead -0.0584)
 
entry:   z4[t] <= -2  -> LONG at open[t+1]
         z4[t] >= +2  -> SHORT at open[t+1]
         (against the move, entry at next bar's open, not at signal close)
 
exit:    open[t+1+H],  H = 24h primary, 4h reported as robustness only
dedup:   one event per symbol per H hours, first wins
 
sizing:  three bases carried, none privileged in the pre-registration text
         but only one used to draft the original verdict (section 6):
           equal_weight    : capital divided evenly across however many
                             signals are live at a given timestamp
           equal_notional  : fixed notional per signal, independent of
                             how many co-fire
           capped_notional : equal_notional with a cap of 20 concurrent
                              positions
 
cost:    22.0 bps round trip, baseline (2x measured taker RT of 11.0 bps)
         11.0 bps and 33.0 bps reported as sensitivity, not as candidates
         for the verdict decision

Causality holds by construction for the signal and condition. Entry timing was itself audited rather than assumed: comparing execution at the signal bar's close against the next bar's open returned 74.4 basis points under both conventions, which is expected on a 24/7 perpetual market with no session gap, and the audit is reported here because the pre-registration required it be checked rather than presumed.

One material specification gap is recorded rather than hidden. The pre-registration does not name a sizing basis. The original verdict used equal_weight, inherited unstated from a cross-sectional module never designed for a per-symbol directional bet. Because equal_weight divides capital evenly across however many signals are live, exposure is largest precisely when the fewest signals fire, which correlates with idiosyncratic rather than market-wide events, and the sign of the headline result depends on this choice. All three sizing bases are reported in section 7 for that reason.

5. Operating envelope

This hypothesis was deliberately chosen to be testable at small capital where H-V1 was not: the loss is bounded by an explicit stop in the economic premise (though not, as section 13 records, in the tested specification), and minimum position size on a perpetual is negligible against the declared working capital.

6. Instrument

The H-R1 conditioned-reversion test, later amended to require all three sizing bases appear in every artifact rather than one. The exact-PnL backtest module carries 17 pinned tests covering direction, movement, and cost combinations, verified against a real trade: a SHORT position moving 41,140 to 39,942 returns +269.2 bps at 22 bps cost under the module, confirming sign convention.

Four audits ran before the PnL test and all four survived: bid-ask bounce (skip-1 bar), realised sigma per subset rather than a global constant, look-ahead in the fragility threshold, and entry-timing convention. None weakened the characterisation. What follows in section 7 tests whether the characterisation converts to money, which is a separate and, as it turns out, much harder question.

7. Results

Headline net return at the pre-registered baseline, equal_weight sizing, n = 11,252 collapsed events:

Cost sensitivity, the table that carries the actual reversal information:

Even at the real measured cost rather than the deliberately pessimistic baseline, the interval does not exclude zero. That is reported as strengthening the verdict, not weakening it: the death is not an artifact of an unnecessarily conservative baseline.

Sizing-basis dependency, the finding that changed how every subsequent dossier in this record reports a PnL table:

The sign flips between bases, but the verdict does not. The capped-notional basis narrows the interval fivefold, from 36.6 basis points wide to 7.3, and that is what sharpens the kill. Reading the upper confidence bound, which is the quantity that says how large an edge the measurement still permits: an edge above +9.6 basis points is excluded on the originally reported basis, and an edge above +2.5 basis points is excluded under capped notional. The earlier revision made this comparison without naming which bound it meant, and the figures it gave, 5 and 26, correspond to neither bound consistently. The honest statement is not "no edge"; it is "edge, if any, is small, and 22 basis points of cost is well outside it."

This table was published in an earlier revision with four defects, found when the evaluation core was prepared for publication in Volume V and each cell was read back against the artifact behind it. All four are corrected above and stated here rather than replaced quietly.

The two withdrawn cells are the more instructive pair. The instrument emits equal_notional as a raw per-leg mean with no interval, by construction, and it emits the cost-sensitivity sweep for equal_weight only. A table that carried a value in every cell therefore read as more complete than the measurement behind it was. Under R-25 a figure with no in-repository instrument behind it is not a result, and these two had none.

Nothing here moves the verdict. Every corrected interval still contains zero at 22 basis points, and the cost axis is unchanged.

Subperiod decomposition: +25.1, -33.0, -4.3 basis points. One of three subperiods agrees in sign with the pre-registered prediction; this fails the 2-of-3 consistency bar on its own, independent of the cost finding.

PnL concentration, the diagnostic that motivated the successor hypothesis: raw-leg mean +0.86 bps sits against a median of +40.21 bps. The worst 1 percent of trades (432 of 43,209) contribute -1,453,065 bps in aggregate; the middle 98 percent contribute a consistent +440,223 bps. Percentile spread is sharply asymmetric: p1 at -2,144.6 bps against p99 at +1,627.1 bps. Symbol concentration is clean (removing the top symbol moves the mean from +0.86 to +0.72), so the tail is not a single-instrument artifact.

The IC-to-PnL conversion tested and found unreliable: the analytical approximation IC * sigma * z predicted 74.4 basis points at the primary cell; realised gross PnL (net at 11 bps plus the 11 bps cost) measured 13.3 basis points, a 5.6-fold overstatement. Two further cells checked under the same conversion overstated by 13.7-fold and 7.3-fold respectively, an inconsistent ratio that rules out a simple correction factor.

8. Verdict, axis of death, and transferability

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

The mechanism is real by the standard this programme applies. The information coefficient survives four independent audits designed to break it, and the fragility conditioning multiplies the unconditional effect by roughly 2.5 times, in the direction and shape the mechanism memo predicted before any PnL was seen. What fails is the translation of that correlation into money at any of the three sizing conventions tested, at the pre-registered pessimistic cost.

This is a cost verdict rather than an idea verdict specifically because the point estimate is positive and the interval narrows meaningfully at the measured real cost of 11.0 basis points, even though it still does not exclude zero there. A reader should not read this as a clean "profitable below X basis points" result. It is closer to: not clearly profitable anywhere tested, but the sign and the magnitude both move in the direction cost predicts, which is different from a mechanism that reverses sign or worsens as cost falls, the signature this record reserves for a UNIVERSAL idea-axis kill.

The subperiod inconsistency (1 of 3 agreeing) is a second, independent reason this hypothesis would not have passed even at zero cost, and it is recorded rather than allowed to hide behind the cost finding. Both weaknesses are real; the cost finding is reported as primary because it is the one with a computable reversal path, addressed in section 10.

9. Cost and capacity analysis

Gross PnL, primary cell            13.3 bps   (net +2.3 at measured cost + 11.0)
Analytical prediction, IC x sigma x z  74.4 bps
Overstatement                       5.6x
Round-trip cost, measured           11.0 bps
Round-trip cost, pessimistic        22.0 bps

This is the second time in the programme an analytical shortcut proved unreliable against exact simulation, after the volatility risk premium averaging model in BL-SD-038, and it produced Rule R-11: an information coefficient may be used for characterisation only, never to stand in for a PnL verdict. Capacity was not separately measured; the small minimum position size means capacity was never the binding question for this hypothesis.

10. Envelope sensitivity and reopening conditions

Binding parameter: round-trip cost, though the reversal is partial rather than clean.

At 22.0 bps (baseline)     : mean -8.7, CI spans zero
At 11.0 bps (measured)     : mean +2.3, CI still spans zero
At 33.0 bps                : mean -19.7, CI excludes zero (negative)
Reversal condition          : point estimate crosses zero between 11 and
                               22 bps; statistical significance on the
                               positive side is not reached at any tested
                               cost level

An operator reaching a round-trip cost below approximately 11 basis points would see a positive point estimate, but this dossier does not claim that crossing produces a demonstrated edge, because the interval never excludes zero on the favourable side at any cost tested. That is the honest reading, and it is a weaker claim than the reversal computed for BL-SD-037's sibling entries elsewhere in this record, stated as such rather than rounded up.

What would reopen it: a cost structure at or below roughly 11 basis points combined with a sizing convention and subperiod behavior that has not yet been demonstrated stable, since even at measured cost the subperiod decomposition disagrees in sign on two of three windows.

What would not reopen it: trimming the worst 1 percent of trades. The diagnostic in section 7 shows this converts the interval to a significant positive result, but the trim requires knowing in advance which trades will be the worst, which is look-ahead. BL-SD-040 tests the only causal implementation of that intuition, a stop loss, directly.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/python/src/conditioned_reversion.py, with 16 tests, alongside the exact-PnL backtest module it depends on, 05_INSTRUMENTS/strategies/python/src/exact_pnl_backtest.py, with 18. The source suite for that second module carries 17 cases; all 17 are ported and one was added for the real-trade sign check quoted in section 6.

The amendment requiring equal_weight, equal_notional and capped_notional in every emitted artifact is enforced in the published module rather than noted in it. A missing basis raises. That rule exists because an unstated sizing convention had already flipped the sign of this family's headline number once, and a comment recording the rule would not have caught it a second time.

Reading each cell of section 7 back against its artifact during this publication found four defects in the sizing-basis table. They are corrected there, and in BL-SD-000 section 7.3.

12. Evidence register

13. Lessons extracted

Two rules in the programme's numbering trace directly to this hypothesis. R-10: a dead verdict is invalid if a variant implied by the hypothesis's own written economic premise was never tested. The pre-registration's section 3 named an explicit stop loss as part of the reason this hypothesis was worth testing at small capital; the tested specification in section 4 omitted it entirely. The initial closure of this branch was retracted for exactly that gap and reopened as BL-SD-040. R-11: an information coefficient characterises a phenomenon and must never stand in for a PnL verdict. The 5.6-fold overstatement here, following an order-of-magnitude overstatement in the volatility risk premium work, establishes this as a demonstrated failure mode rather than a single incident.

A third, unnumbered lesson belongs here plainly: an unstated sizing convention changed the sign of a headline result. Every PnL table in this programme now reports multiple sizing bases rather than one, specifically because of what this hypothesis's audit uncovered.

14. Cross-references

BL-SD-040 for the stop-loss successor this hypothesis's own retracted closure produced. BL-SD-038 for the prior instance of an analytical shortcut overstating realised PnL. BL-D-002 section 4 for the ENVELOPE-BOUND classification test. BL-SD-000 ledger row H-R1 and section 5's discussion of the partial cost reversal.

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