BL-SD-051: Time-Series Momentum, Market Neutral

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Bybit USDT perpetuals Programme: II Verdict: NOT PROVEN Blocking axis: Power Transferability: -

1. Identification

NOT PROVEN is not a kill and this row is not recorded as one. It is included in full, with both its published figure and its corrected figure, because the correction is the most instructive part of the record.

2. Economic thesis

An instrument that has trended tends to keep trending, over horizons measured in weeks rather than minutes, because information diffuses into price slowly and because rebalancing and hedging flows push in the direction a trend has already established. The classic term structure, negative at very short lookbacks where the effect is reversion, turning positive and growing through roughly 30 days, then decaying, is documented across asset classes for decades and is not a shape noise produces by itself.

Who is forced to lose, if the mechanism works as stated. Participants who must trade against an established trend for reasons unrelated to its information content: index rebalancers, forced hedgers, and market participants exiting a position on a schedule rather than on a view. A trend follower is on the other side of that forced flow, not predicting a fundamental outcome.

3. Falsifiable claim

No claim was pre-registered. The characterisation swept a grid of lookback and holding-period combinations and reported every cell, which the source material treats as the discipline that substitutes for a single pre-committed specification when no fresh pre-registration quota remains:

4. Mathematical specification

Notation per BL-D-011.
 
Time-series signal, instrument i, lookback L hours, at rebalance time t:
 
  s_i(t) = sign( r_i(t - L, t) )        trailing return sign, own history only
 
Position, holding period H hours, equal risk across the active universe:
 
  w_i(t) = s_i(t) / N(t)                N(t) = number of instruments with a
                                         valid signal at t
 
Realised portfolio return over the holding window:
 
  R(t) = sum_i w_i(t) * r_i(t, t + H)
 
Beta-neutral decomposition, against the average net exposure the
construction carries:
 
  expo(t)   = mean_i( w_i(t) )          net directional tilt of the book
  R_beta(t) = expo(t) * r_mkt(t, t+H)   r_mkt = equal-weight universe return
  R_resid(t) = R(t) - R_beta(t)         momentum component net of beta
 
Cost:  c_rt = 8.85 bps, measured hybrid round trip, `BL-FM-002` section 3.3,
       scaled by realised turnover at holding period H
 
Sizing: equal risk per active instrument, gross exposure normalised to 1.0

Causality holds for the signal: s_i(t) depends only on the trailing window ending at t. L and H are fixed a priori across the grid and never fit to the data being evaluated, which the source material treats as the reason this class differs from a fitted-parameter result: there is nothing here to overfit.

One parameter was not recognised as free at the time of the original measurement, and this dossier states it plainly rather than reconstructing a version of events in which it was controlled from the start. The rebalance anchor, the specific hour of the day at which the grid resets, was fixed by an accident of the era boundary rather than chosen or tested, across the original run of this measurement. Section 7 reports both the single-anchor figure as originally measured and the full 24-anchor sweep that corrected it.

5. Operating envelope

6. Instrument

The momentum cell map produced the 24-cell lookback and holding-period grid. The rebalance-hour sweep, BL-IS-004, built for an unrelated reversion question, was the instrument that first exposed the rebalance-anchor problem and was then applied to this row's headline cell as a direct audit.

7. Results

Term structure across the lookback and holding-period grid, TSMOM, holding period 168 hours (1 week), net of cost:

The classic shape is present: negative at the shortest lookback, growing through 30 days, decaying by 60. Cost falls from 4.70 to 0.50 bps as lookback lengthens, because turnover falls with it.

Beta-neutral decomposition, best cell, L=720, H=168, single-anchor measurement as originally reported:

Capacity check, same cell, full universe against a liquid-half subset:

The beta-neutral component is stronger, not weaker, on the more liquid half, which is the opposite of a result concentrated in illiquid names.

Per-year decomposition, beta-neutral net, same cell:

Six of six years positive, including 2026, which falls inside held-out B.

The anchor-sensitivity correction. The figures above for the headline cell were measured at a single rebalance anchor that was never chosen; it was inherited from an era-boundary accident, the same defect documented in BL-IN-006. Re-running the identical specification at all 24 possible hourly anchors gives:

The figure originally published was the best of 24 alternatives that had never been disclosed as existing, because nobody had recognised the rebalance hour as a parameter at all.

8. Verdict, blocking axis, and transferability

Verdict: NOT PROVEN. Axis: power. Transferability: not applicable.

Two things are true at once and both belong in the verdict rather than either alone.

The direction is robust in a way few results in this record can claim: 24 of 24 rebalance anchors positive, 6 of 6 calendar years positive including a stretch inside held-out B, and a beta-neutral residual that strengthens rather than weakens on the more liquid half of the universe. Compare the sibling reversion construction audited in the same pass, whose sign flips across 13 of 24 anchors: this is a materially different class of evidence.

The magnitude is not established at conventional significance once the anchor-selection bias is removed. Median t across the 24 anchors is approximately 1.4, and only 5 of 24 individually clear 1.96. The interval on the median-anchor figure does not exclude a much smaller true effect, and no sealed held-out era remains on this substrate to certify it further: both held-out A and held-out B had already been opened for other hypotheses before this characterisation was run. Per BL-D-005, a result on the power axis is not a kill; it is an honest statement that the sample, in the sense of independent statistical evidence rather than calendar length, was not enough to settle the magnitude.

NOT PROVEN is the correct label precisely because the sign question and the magnitude question point in different directions, and collapsing them into a single verdict in either direction would misstate one of the two.

9. Cost and capacity analysis

Beta-neutral net, full universe, best anchor      +14.11 bps/week
Beta-neutral net, full universe, median anchor     +8.72 bps/week
Beta-neutral net, liquid-half subset, best anchor  +18.73 bps/week
Turnover-scaled cost, best cell                     1.84 bps/week

The capacity direction is favourable: the beta-neutral component is stronger on the liquid half of the universe than on the full universe, the opposite of the pattern in constructions where the edge concentrates in thin names. This was not a fitted result; lookback and holding period were fixed a priori across the entire grid before either the capacity split or the anchor sweep was run. Total leveraged return, before the beta-neutral adjustment, sits at roughly 78 percent beta exposure to a net-short-altcoin tilt and should not be reported or sold as the strategy's return; the beta-neutral figure is the only one this dossier treats as attributable to the mechanism itself.

10. Envelope sensitivity and reopening conditions

Binding parameter: statistical power, specifically the absence of a clean held-out era for this substrate.

Reversal condition (toward a settled verdict, either direction):
  A held-out era not previously opened for another hypothesis on this
  substrate, sized to distinguish a median effect near t=1.4 from zero
  at conventional power.

This is not a cost-reversal arithmetic in the usual sense, because the finding is not that a cost floor consumes an otherwise-adequate edge; the median-anchor net of +8.72 bps/week already sits above the 1.84 to 2.9 bps cost measured at this lookback. The binding constraint is that neither held-out era on this substrate remains unopened, so no fresh test can independently confirm the median-anchor magnitude without reusing data already spent on H-X4 and on the carry characterisation in BL-SD-050.

What would reopen it toward VALIDATED: a new, sealed out-of-sample era, pre-registered before being opened, on which the median-anchor specification is run once, anchor-swept from the start, and clears conventional significance.

What would close it toward FALSIFIED: the same test, run cleanly, with a result that reverses sign or fails to distinguish from zero at the median anchor. Given 24 of 24 anchors currently agree in sign, this outcome would itself be a notable finding.

What would not change the verdict: further characterisation on the existing substrate. Both held-out eras are spent, and additional grid sweeps on data already examined would only compound the multiple-testing exposure this row's own correction exists to prevent.

11. Reference implementation

The anchor-robustness instrument specified in BL-IS-004 is published as anchor_robustness in 05_INSTRUMENTS/src/inference.py. The momentum cell map and the rebalance-hour sweep that feed it remain scheduled.

That instrument exists because of this row, and the correction it produced went against us. The rebalance-hour sweep on this family was the measurement that found nineteen pre-registrations had left the phase of a periodic rebalance uncontrolled, and publishing the instrument later found a second defect in the same place: significance had been counted on a signed statistic where the specification requires the two-sided form. What had been reported as no phase being significantly positive is, corrected, seven of twenty-four phases significantly negative. BL-IS-004 section 8 and BL-IN-006 carry it.

12. Evidence register

13. Lessons extracted

The rebalance anchor was never chosen for this measurement, in the same way and for the same structural reason documented for the reversion family in BL-IN-006: the grid builder inherited midnight, or in this family's case a different accidental hour, from an era-boundary timestamp that had been set for an unrelated purpose. Nineteen pre-registrations across the programme carried this same uncontrolled parameter, and this row is the one where the damage ran in the flattering direction: the figure originally reported was the maximum of 24 alternatives nobody knew existed.

What survived the correction is worth stating as plainly as what did not. The direction of the effect, 24 of 24 anchors positive, is qualitatively different evidence from a result that flips sign across the same sweep, and the correction strengthened confidence in that direction even as it cut the reported magnitude by roughly forty percent. An audit that reduces a published number and increases confidence in the underlying mechanism at the same time is the audit behaving correctly, and the record states this without treating the reduced number as a disappointment to be minimised.

R-09, requiring every periodic-rebalance measurement to report anchor sensitivity rather than a single phase, exists because of this row and its reversion sibling together. This dossier applies it retroactively rather than reporting the single favourable anchor a second time.

14. Cross-references

BL-IN-006 for the incident this row's correction belongs to. BL-IS-004 for the anchor-robustness instrument built in response, which names this row directly as one of the two results it corrected. BL-D-009 rule R-09. BL-SD-050 for the sibling characterisation sharing this row's held-out- exhaustion caveat. Ledger row TSMOM.

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