BL-SD-012: Cross-Sectional Funding Carry, Market-Neutral

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

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

1. Identification

Designed and settled the same day. The measurement is a walk-forward out-of-sample test on data already streaming to the platform, so no new instrumentation was required between design and verdict.

2. Economic thesis

Perpetual funding is a real cashflow paid every eight hours between the long and short side of a contract. Where funding disperses widely across a cross-section of symbols, a basket can be built that shorts the richest-funding names, collecting the payment, and longs the cheapest or most negative, also collecting a payment, while remaining dollar and beta neutral. Two edges were claimed to stack: the carry itself, and a belief that funding extremes mark crowded positioning that should mean-revert, adding a price-reversion return in the same direction as the carry.

Who is forced to lose. The design record does not name a structural forced loser for this thesis. It argues from a statistical property, funding dispersion, rather than from a mechanism that compels one side of the trade to transact regardless of price. That absence is a retrospective weakness of the design and is treated as one in section 13, not smoothed over here.

3. Falsifiable claim

The source record does not preserve a pre-registration sentence written before data was seen; the design document states the decision rule in narrative form rather than as a quoted claim. Reproduced as closely as the source allows:

Cross-sectional funding carry, dollar and beta neutral, produces a positive net return out of sample, beating a random-entry null, under a correlation-aware walk-forward test.

4. Mathematical specification

Notation per BL-D-011.
 
Rebalance at funding stamp t, using only funding realised at or before t:
 
  legs_t = select_funding_legs( funding_rate_{.,t}, k )
    ascending sort of funding_rate_{i,t} across the universe
    long the k lowest (cheapest or most negative funding)
    short the k highest (richest funding)
    direction_i in {+1 long, -1 short}
 
Funding carry collected while the position is open, Bybit sign convention
(positive rate: longs pay shorts):
 
  carry_i = sum over settlements s in (t, t+hold] of  -direction_i * funding_rate_{i,s} * 1e4
 
Price leg, basket-relative spread (market neutral):
 
  spread_{i,t} = ret_{i,t}(lookback) - mean_j ret_{j,t}(lookback)
  price_i      = direction_i * ( spread_{i,t+hold} - spread_{i,t} ) * 1e4
 
entry:  at stamp t, non-overlapping rebalance every `hold` stamps
exit:   at stamp t + hold
sizing: equal notional per leg, 2k legs, dollar neutral by construction
cost:   c_rt = 2 * fee_taker_per_fill + spread_cost_bps    (per leg, round trip)
 
net per leg:
  r_net,i(t) = price_i + carry_i - c_rt

Causality holds for the selection term: funding_rate_{i,t} is the settled rate observed at the decision stamp, not a predicted rate, so legs_t is measurable with respect to F_t. carry_i sums only settlements that occur after entry and before exit, which is a realised cashflow rather than a forecast. spread_{i,t+hold} is used only to close the position at the exit timestamp and is never referenced before it occurs.

5. Operating envelope

6. Instrument

The funding-carry primitive, pure functions with 13 unit tests, pinning the Bybit funding sign convention explicitly as the identified primary risk of the module. The price leg reuses the tested market-neutral spread primitives from the cross-sectional reversion engine rather than re-deriving them. Harness: a dedicated out-of-sample runner, paginated funding history plus four-hour closes aligned to eight-hour funding stamps, walk-forward out-of-sample with 5 folds, random-entry null.

A correlation-aware caveat applies to the inference method and is stated rather than hidden. Per-leg statistics assume independent trades, but same-timestamp legs and overlapping windows are correlated, which overstates significance; a matched null recomputed under that caveat came in near -57 basis points against a naively expected null near -15. The finding here is not reversed by that caveat, because the point estimate is roughly six times the naive null's shift, but the record states the caveat rather than presenting the original interval as if it were free of it.

7. Results

The carry component is real and matches its predicted sign. It is a small fraction of the outcome. The price component, which the thesis required to add in the same direction as the carry, instead runs against it by more than an order of magnitude.

8. Verdict, axis of death, and transferability

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

At zero cost, the carry leg alone clears the null. The full position does not, because the price leg moves in the direction that continues the pre-existing funding-driven trend rather than in the direction that would correct it. A coin at funding extremes continues rather than reverts at this horizon. No fee substitution, rebate tier, or execution channel repairs a return whose sign is set by the direction of an ongoing price move rather than by the margin available to capture a carry payment.

This places the result in the same family as the market's general behaviour recorded across this programme: crypto perpetual prices at these horizons are continuation-shaped rather than reversion-shaped, and a thesis built on the second without evidence of the first fails on the mechanism.

9. Cost and capacity analysis

Carry component, mean            +15.2 bps
Price component, mean           -344.9 bps
Net, out-of-sample                -345.0 bps

Cost is not the binding term here. A round trip of roughly 15 basis points is immaterial against a 345 basis point loss driven by the price leg. Capacity was not measured; the universe of 26 symbols was already understood at design time to be a constraint on cross-sectional power rather than a constraint that would have been reached, since the hypothesis failed before frequency or size became the relevant question.

10. Envelope sensitivity and reopening conditions

No binding parameter exists to name, because the finding is a sign reversal on the price leg, not a magnitude shortfall against cost.

Reversal condition:  none computable.

Setting the round-trip cost to zero and crediting it back to the observed net figure still leaves the position far short of breakeven, because the defect is in the direction of the price move rather than in the fee paid to capture it. This is the same reasoning BL-D-002 section 4 applies elsewhere in this record: a hypothesis that would still fail under arbitrarily favourable envelope parameters is falsified on the idea, not closed on the envelope.

What would reopen it: evidence, on data this programme has not seen, that funding extremes revert at some horizon other than the one tested here. That is a testable claim and it was not tested, because testing it on the same data the verdict was drawn from would be horizon shopping.

What would not reopen it: re-running the same horizon with a different fee schedule. The price leg, not the fee, decided this result.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/python/src/funding_carry.py, with 24 tests. It imports the market-neutral spread primitives from cross_sectional_reversion rather than re-deriving them, exactly as the source did: the price component of this strategy is the same basket-relative move BL-SD-010 measures, and a second definition of that quantity would let the two dossiers drift apart.

Six of the twenty-four tests do nothing but pin the funding sign convention, one for each combination of position direction and rate sign. On a standard perpetual a positive rate means longs pay shorts, so per settlement stamp

carry_bps  =  -direction * funding_rate * 1e4

Getting that backwards turns every loss into a gain of the same size and would not fail loudly anywhere else in the pipeline. BL-IS-006 carries the same warning about markout for the same reason.

The module reports price_bps and carry_bps as separate fields rather than only their sum, and that is a deliberate choice about what this row found. Carry came in at +15.2 basis points, which is real and is what the cashflow argument predicted. Net came in at -345.0. A single net figure would have hidden the one part of the thesis that held, and a reader with a different envelope needs to see that the carry edge exists and that this selection rule buys a price exposure more than twenty times its size.

One guard was added that the source did not have. The price and funding matrices must be aligned on the same settlement grid, and the published module raises if they are not. A misalignment there produces a plausible number rather than an error, which is the failure mode most worth stopping in a strategy whose entire premise is that two series share a stamp.

12. Evidence register

13. Lessons extracted

This hypothesis is one of four in the same design cycle that reinforced a single conclusion: the programme's data substrate at this point rewarded continuation, not reversion, at every horizon then tested. It contributed directly to design principle 11 in the later Bybit strategy redesign, stated plainly as a prior against fading a flow that still has fuel.

The absence of a named forced loser in the original design, noted in section 2, is itself a lesson this record preserves rather than retrofits: a hypothesis argued from a statistical property of dispersion, without a mechanism compelling one side to transact regardless of price, is structurally closer to a pattern than to a thesis.

14. Cross-references

BL-D-011 section 5 for the correlation-aware portfolio-collapse method this measurement's caveat anticipates but predates in full rigor. BL-SD-000 ledger row XSEC-CAR, and sibling same-day rows XSEC-REV and XSEC-MOM for the companion approaches tested against the same substrate.

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