BL-SD-037: Crowded Leverage, Directional Deleveraging

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-P1 opened the programme's second research budget line. It consumed the first of twelve pre-registered hypothesis slots and was closed inside two days, including the time spent building and validating the substrate underneath it.

2. Economic thesis

When leveraged positions accumulate one-sided on a symbol while price fails to pay that side, funding stays costly for the crowded side, price stagnates or drifts the wrong way, margin on those positions thins. The holders do not choose when they exit. The exchange's liquidation engine chooses for them, and it executes without regard to price. Liquidation itself presses price further, which margins the next cohort, which liquidates in turn.

Who is forced to lose. The crowded leverage holders. The force is written into the exchange's margin rules rather than into any participant's judgment, and it cannot be escaped by being faster or better informed, because the mechanism does not consult the position holder at all once the margin threshold is crossed.

3. Falsifiable claim

Pre-registered before any forward return was computed:

Conditional on LONG_CROWDED, the forward 24-hour return distribution carries (a) more negative skewness and (b) a higher left-tail probability than a matched baseline, on the same symbol and the same time regime.

The mirror test on SHORT_CROWDED is written into the pre-registration as mandatory rather than optional, on the grounds that a genuine deleveraging mechanism predicts a skew of the opposite sign on the opposite side.

4. Mathematical specification

Notation per BL-D-011. Base bar = 1 hour. All rolling statistics trailing
and exclusive of bar t.
 
State inputs, per symbol i, bar t:
 
  oi_z[t]     = (oi_chg[t] - mean_roll(oi_chg, W)) / sd_roll(oi_chg, W)
                oi_chg[t] = log(OI[t] / OI[t-L]),  W = 720 bars trailing
 
  fund_pct[t] = midrank_percentile(funding[t], trailing 90 days)
                midrank: 100 * (n_less + 0.5 * n_equal) / n_valid
                (amendment 27a: ECDF alone misplaces the clamp-baseline
                atom, which sits at 32 to 68 percent of bars on several
                symbols, near the median rather than near zero)
 
  ret_pct[t]  = percentile(ret_L[t], trailing 90 days)
                ret_L[t] = log(close[t] / close[t-L])
 
State, primary side:
 
  LONG_CROWDED[t]  = (oi_z[t] >= theta_oi)
                 AND (fund_pct[t] >= theta_fund)
                 AND (ret_pct[t] <= theta_ret)
 
State, mirror side (mandatory confirmation, not a separate hypothesis):
 
  SHORT_CROWDED[t] = (oi_z[t] >= theta_oi)
                 AND (fund_pct[t] <= 100 - theta_fund)
                 AND (ret_pct[t] >= 100 - theta_ret)
 
Persistence: state must hold for k consecutive bars; events within
H_primary of a prior event on the same symbol collapse to one (first wins).
 
Grid, locked before any outcome was computed:
  theta_oi in {1.5, 2.0, 2.5}, theta_fund in {80, 90, 95} pct,
  theta_ret in {30, 50} pct, L in {24, 72} h, k in {1, 3} bars.
Selected configuration (mechanical, outcome-blind selection rule):
  theta_oi=1.5, theta_fund=80, theta_ret=50, L=24h, k=3.
 
signal:  fwd[t]     = log(close[t+H] / close[t]),  H = 24 h fixed
         sigma_local = realised 1h sd over trailing 720 bars, scaled to H
entry:   characterisation only. No position is opened; the test compares
         the forward-return distribution conditional on the state against
         a matched baseline.
exit:    not applicable. H = 24h defines the measurement window, not a
         trade holding period.
sizing:  not applicable, no trading construction exists for this test.
cost:    not applicable. Rule 4 of the pre-registration prohibits
         computing PnL, entry, or bracket rules before this test clears
         its own kill criteria. It did not clear them, so no cost model
         was ever specified and none is reconstructed here.
 
Test statistics:
  delta_skew  = skew(fwd | crowded)           - skew(fwd | matched baseline)
  delta_Ptail = P(fwd < -2*sigma_local | crowded) - P(fwd < -2*sigma_local | baseline)
 
Matched baseline: for each event on symbol i at bar t, the baseline
population is bars of symbol i within a 30-day window of t that are not
in a crowded state, controlling for both symbol and regime.

Causality holds by construction: oi_z, fund_pct, and ret_pct at bar t are computed from rolling windows ending at t-1; the state at t uses only information at or before t; the forward window [t, t+H] is the outcome being measured, never an input to the state.

No trading construction is published because none was ever authorised. The pre-registration's rule 4 explicitly bars computing PnL or an execution rule before the distributional test passes, which is a stronger discipline than most of this record's entries carry and is preserved here rather than retrofitted with an entry rule that was never run.

5. Operating envelope

The cost pre-check in section 2 of the pre-registration found cost was, for the first time in the record, not the binding constraint: at a targeted event magnitude of 300 basis points or more, an 11.0 basis point round trip consumes under 4 percent of the move. That check never mattered, because the hypothesis died on the distributional test before a trading construction existed to price.

6. Instrument

The crowding characterization instrument performs outcome-blind threshold calibration; select_config is structurally prevented from receiving a forward-return column, pinned by a bytecode-level test (test_select_config_signature_has_no_outcome_argument). The causal state module carries anti-look-ahead property tests. The G3 skew and tail test ran on the calibrated configuration only, after calibration was locked.

A caveat inherited from the data gate that preceded this test: the exchange's forced-deleveraging proxy, built to validate that the crowding state actually corresponds to liquidation-driven flow, reaches Spearman rho of 0.067 at 1-hour resolution and 0.454 at 5-minute resolution against realised liquidation notional on the 9-day validation window. The 5-minute figure clears the pre-registered floor of 0.40 but not the pass threshold of 0.50, and amendment 27b locks its status at WEAK permanently, inherited by every verdict in this family. The caveat concerns the causal story linking crowding to forced liquidation flow; it does not touch the skew and tail test itself, which is computed directly from open-interest, funding, and price state rather than from the liquidation proxy.

7. Results

Primary side, LONG_CROWDED, n = 904, against baseline n = 45,200:

The predicted direction was more negative skewness. The measured direction is more positive. This is not a marginal miss; the sign itself is wrong.

Mandatory mirror side, SHORT_CROWDED, n = 3,050, against baseline n = 152,500:

A directional deleveraging mechanism predicts opposite-signed skew on the two sides: negative for LONG_CROWDED, positive for SHORT_CROWDED. Both sides instead skew positive.

The audit that decides the axis measures both tails on both sides, not one tail per side as the pre-registration originally specified:

Both tails thicken on both sides. Median absolute z rises 15 percent on LONG and 31 percent on SHORT. The state predicts a wider distribution, not a directional one.

8. Verdict, axis of death, and transferability

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

The kill criterion in section 3 fired exactly as written: n_event = 904 is well above the 150-event power floor, and both confidence intervals contain zero. That alone would be a clean but unremarkable falsification. What makes the verdict decisive rather than merely underpowered is the sign and the mirror test together: skewness moves in the wrong direction on the primary side, and it moves in the same wrong direction on the side that should have mirrored it. A mechanism that produces the same-signed distortion regardless of which side is crowded is not a directional mechanism. It is a volatility mechanism wearing a directional hypothesis's clothes.

This is a sign reversal rather than a magnitude shortfall, which places it in the UNIVERSAL class under the test in BL-D-002 section 4: no envelope substitution, cheaper cost, faster reaction, more capital, converts a wrong-signed effect into a right-signed one.

Two caveats are carried forward honestly rather than left to qualify the verdict quietly. The deleveraging proxy that motivates the causal story is permanently rated WEAK, which limits how much of the mechanism narrative in section 2 can be taken as confirmed rather than as a plausible account. And the panel is unbalanced, with only half the symbol universe carrying full research-era history, which the programme's survivorship note in the source report flags as a conservative rather than an inflationary bias for this particular mechanism.

9. Cost and capacity analysis

No PnL, gross or net, was ever computed. The pre-registration barred pricing a trading construction before the distributional test passed, and it did not pass. The only cost figure attached to this hypothesis is the pre-check in section 5: an 11.0 basis point round trip against a targeted event magnitude of 300 basis points, a ratio of 3.7 percent, confirming that cost was not expected to be the binding constraint if the mechanism had held. Capacity was never measured for the same reason.

10. Envelope sensitivity and reopening conditions

No binding parameter exists to name, because the finding is a sign reversal rather than a magnitude shortfall.

Reversal condition: none computable.

A cost, latency, or capital substitution cannot reverse this result. The measured effect is a symmetric widening of the forward-return distribution on both sides of the crowding state, and no envelope parameter changes a distribution's third moment from positive to negative. Setting every cost and capacity term to its most favourable value leaves the sign exactly where it was measured.

What would reopen it: a demonstration, on a substrate this measurement never touched, that the crowding state predicts direction at a horizon other than 24 hours. That is a testable claim and it is untested here; testing it on the data this verdict was drawn from would be horizon shopping under the programme's own prohibition.

What would not reopen it: any repair to the deleveraging proxy. The proxy's WEAK rating concerns the causal story, not the skew and tail test, which does not depend on the proxy at all.

The hypothesis leaves behind a measurement worth more than the one it failed to produce: crowding state predicts a volatility expansion, not a directional move, at p_perm = 0.001 and consistent sign on both sides. That finding seeded the volatility risk premium work in BL-SD-038 and the defensive regime gate now in production, and it is recorded here as the honest yield of a hypothesis that died on its own premise.

11. Reference implementation

The causal crowding state module is published as 05_INSTRUMENTS/strategies/python/src/crowding_state.py, with 21 tests, on the trailing percentile already published in 05_INSTRUMENTS/strategies/python/src/fragility_conditioning.py. The mechanical selection module, the characterisation instrument and the stage-three tail test remain scheduled. No trading construction exists to publish, for the reason given in section 4.

Three properties are pinned. The state bounds are inclusive exactly as pre-registered, and the boundary values themselves are tested, because tightening a bound after seeing the data is the cheapest form of threshold shopping. A NaN input maps to no event, since not-yet-known is not no-event and the only thing that matters on the selection side is that it does not fire. And the trailing z-score never reads the bar it labels; a test moves a late bar and confirms no earlier score shifts.

12. Evidence register

13. Lessons extracted

Two instrument defects surfaced in the verdict document and neither reversed the finding, but both changed how later hypotheses in the programme were built. Subperiod consistency was originally scored as passing when either delta skew was negative or delta P_tail was positive, which let a subperiod with the wrong-signed skew count as agreeing with the prediction. And a block bootstrap sensitivity table printed a point estimate that does not depend on block length by construction, which is not evidence of stability; only the interval width carries that information. Neither defect is reused.

The substantive lesson is now load-bearing across the record: measure both tails from the start, run the mirror test as a two-sided requirement rather than a single anticipated failure mode, and treat a same-signed distortion on both sides of a conditioning variable as proof of a width effect rather than a direction effect, regardless of significance. Volatility expansion is easier to predict than direction, and this hypothesis is the programme's first clean demonstration of that asymmetry, which is why it seeded the fragility-conditioned volatility work rather than simply closing a line item.

14. Cross-references

BL-SD-038 for the volatility risk premium work this falsification seeded. BL-P2-002 section 4 for the gate narrative. BL-D-002 section 4 for the UNIVERSAL classification test. BL-SD-000 ledger row H-P1.

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