BL-SD-008: Taker Wider Edge on a Rolling Universe

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Venue: Bybit perpetuals Programme: I Verdict: FALSIFIED Axis of death: Cost Transferability: ENVELOPE-BOUND

1. Identification

Two measurements are reported under this one code, taken 24 hours apart on growing samples of the same live shadow run: an interim read at n=662 and a fuller evaluation at n=995. Both are reported because the interim figure looked like a pass and the fuller one did not, and collapsing them into a single number would hide that reversal.

2. Economic thesis

Unchanged in mechanism from BL-SD-007: order-flow, VPIN, and spread-based triggers on a wide, actively rotated universe of high-movement altcoins should supply enough 100-basis-point-class dislocations to clear the venue's round-trip cost, where a fixed 8-symbol universe could not. The universe itself becomes a parameter: a movement-spec filter rotates in the top 15 to 25 symbols by qualifying-move frequency every 24 hours, replacing the static Tier 1.5 set.

Who is forced to lose. Unchanged: the passive counterparty on the side being swept by the triggering order flow.

3. Falsifiable claim

Pre-registered before the shadow window began:

Over a rolling universe, the strategy accumulates at least 30 shadow trades, produces an EV per trade above 0 basis points after a 12 basis point taker-to-taker worst-case fee, holds an aggregate p_tp at or above 0.45, shows no per-signal-type EV below -10 basis points, and completes at least one clean universe rotation with no infrastructure regression.

4. Mathematical specification

Notation per BL-D-011.
 
Signal, symbol-agnostic, three independent triggers, any one sufficient:
 
  ofi_3sigma:    |ofi_l5_norm(t)| >= 3.0
  vpin_extreme:  vpin(t) >= 0.85  AND  |ofi_l5_norm(t)| >= 1.0
  spread_spike:  spread_bps(t) >= 2.0 AND |ofi_l5_norm(t)| >= 0.5
 
  each gated on is_warmed_up(t) = true
 
Universe, U(t):
 
  U(t) = top_n_by_movement_spec_score(t), n in [15, 25]
  rotation cadence: 24 h
  symbol_protection_hours = 12       (no drop within first 12h of entry)
  eligible only if symbol in U(t) at signal time
 
Entry, fire at t when any trigger fires for symbol in U(t):
 
  cooldown: 60,000 ms per (symbol, signal_type)
  direction  d_t = sign(ofi_l5_norm(t))
  order type: Market IOC (taker)
 
Exit, path-dependent, first touch wins (identical bracket to BL-SD-007):
 
  tp_price  =  entry_mid + d_t * entry_mid * 100 / 10000
  sl_price  =  entry_mid - d_t * entry_mid * 50  / 10000
  timeout: 4 hours (empirically found never binding; 99.9 pct of positions
           resolved within 2 hours, observed max hold 7,430 s)
 
Sizing:
 
  notional_usd = 5.0
 
Cost, per outcome branch:
 
  c_rt(tp_hit)              =  6.0 (taker entry) + 1.0 (maker exit) = 7.0 bps
  c_rt(sl_hit or timeout)   =  6.0 (taker) + 6.0 (taker)            = 12.0 bps
 
Breakeven win rate, from the branch payoffs:
 
  WR_breakeven = (50 + 12) / ((100 - 7) + (50 + 12)) = 62 / 155 = 40.0 pct

Causality holds by construction: U(t) is recomputed on movement-spec scores from state at or before t, each trigger reads current snapshot state, and the cooldown check reads only the prior exit timestamp per (symbol, signal) pair.

5. Operating envelope

6. Instrument

The taker wider-edge strategy implementation, redesigned from a per-combo hardcoded struct to the symbol-agnostic signal specification shown in section 4, with a movement-spec filter supplying the rotating universe. A shadow-run aggregation tool and a what-if signal-exclusion simulator performed the offline aggregation for the fuller read.

The two reported samples used measurably different costing rigor. The interim read (n=662) computed EV as mean raw realised bps less a flat 7 basis point assumption, applied uniformly regardless of outcome. The fuller read (n=995) applied the branch-conditional cost model shown in section 4, which correctly assigns the higher 12 basis point cost to stop-loss and timeout exits. That refinement, not a change in the underlying market, is part of why the two reads disagree.

7. Results

Dropping spread_spike alone, holding the other two signals fixed, would have lifted aggregate win rate to approximately 41.4 percent and produced a positive aggregate EV, per the what-if filter simulation run on the same sample.

8. Verdict, axis of death, and transferability

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

The strategy as deployed, all three signals firing on the full rotating universe, produced an observed win rate of 39.5 percent against a breakeven requirement of 40.0 percent, a gap of half a percentage point that closes the line on the pre-registered EV criterion. The gross magnitude of the signal, a mean raw move around 8 basis points, is real and an order of magnitude larger than anything found in BL-SD-003 through BL-SD-006; the finding here is that the asymmetric branch cost, 7 basis points on the take-profit path against 12 on the stop-loss and timeout paths, converts a positive-looking gross figure into a fee-adjusted loss.

Applying the test in BL-D-002 section 4: the interim read at n=662 was positive under a simplified flat-fee assumption, which is the concrete demonstration that a lower or more favourable cost structure produces a positive expectation from the same underlying signal. The classification is therefore envelope-bound, not universal.

9. Cost and capacity analysis

Mean raw realised move (gross), full read     +8.05 bps
EV, interim read, flat 7 bps assumption       +1.38 bps
EV, full read, branch-conditional cost        -1.97 bps
Cumulative loss, 23h, $5 notional              -0.98 USD
Toxic signal contribution to net loss          spread_spike: -2,363.6 bps
                                                cumulative, 700 of 995 trades

Capacity was not separately assessed; frequency was ample (995 trades in roughly 23 hours across a 15-25 symbol universe) and the closure is on unit economics rather than on throughput.

10. Envelope sensitivity and reopening conditions

Binding parameter: round-trip cost on the stop-loss and timeout exit leg.

Fixed: TP-path net at observed WR         100 - 7  = 93 bps (weight 39.5 pct)
Solve for c_sl at which observed WR (39.5 pct) clears breakeven:
 
  0.395 = (50 + c_sl) / (93 + 50 + c_sl)
  0.395 * (143 + c_sl) = 50 + c_sl
  56.485 + 0.395 c_sl = 50 + c_sl
  c_sl = 6.485 / 0.605 = 10.72 bps
 
Reversal condition:  c_sl < 10.72 bps    (assumed 12.0 bps in this
                                            measurement; declared programme
                                            figure 11.0 bps taker-to-taker)

This is a genuinely narrow margin. The currently declared taker-to-taker round trip of 11.0 basis points still exceeds the 10.72 basis point reversal point, so the strategy remains marginally unprofitable even under the updated schedule; the gap closed from 2.0 basis points under the original 12.0 basis point assumption to roughly 0.3 basis points under the declared figure, but did not cross zero.

A second, cheaper reopening path does not require touching cost at all: the per-signal breakdown in section 7 shows that disabling spread_spike alone, which contributes 70 percent of volume and the entirety of the net loss while the other two signals are independently marginal-positive, would have produced a positive aggregate EV on the same sample without any change to the fee schedule. That configuration change, not a cost substitution, is what the successor decision (BL-SD-009) attempted, though it changed the signal logic further rather than simply dropping the toxic trigger.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/rust/src/taker_wider_edge.rs, with the 15 test cases that pinned it in the source. No code defect was found; the finding is a cost-and-composition result on a correctly implemented signal set. The async run loop and its transport are not reproduced.

12. Evidence register

This measurement is recorded in narrative decision documents; registration as a standalone document-evidence artifact is scheduled per BL-SD-000 section 8.

13. Lessons extracted

Two lessons, both methodological rather than about the market. First, a flat blended fee assumption applied uniformly across outcome types understated the true cost of this strategy's asymmetric TP/SL cost structure by enough to flip an interim read from marginal pass to marginal fail once corrected; any strategy with outcome-dependent execution channels needs the branch-conditional cost model from the start, not as a later refinement. Second, a signal-agnostic OR-fan-out across three independently reasonable triggers concealed a single toxic signal responsible for the entire net loss; aggregate reporting without a per-signal breakdown would have closed this line as a flat failure rather than as a mostly-working strategy carrying one bad component. Both lessons feed directly into the orthogonality and attribution requirements later formalised for Hyperliquid multi-confirmation work in BL-D-007.

14. Cross-references

BL-SD-007 for the static-universe predecessor this design widened past. BL-SD-009 for the successor confluence redesign this result's toxic-signal finding motivated. BL-D-002 section 5 for the reversal-point method applied in section 10. Ledger row T1.5.

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