BL-IN-001: The Impossible Win Rate

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Incident date: 2026-05-24 Class: Cost arithmetic never performed Rule produced: R-01, and the practice that became layer 0 of the falsification chain

1. What was believed

The taker strategy was a market-order execution model driven by an order-flow and toxicity signal stack. Work on it was proceeding normally: threshold tuning, signal calibration, universe selection. The working assumption was that a win rate somewhere in the sixty to seventy percent range would make the strategy commercially viable once it scaled.

Nobody had asked what win rate the strategy would need.

2. The trigger

A 25-hour demo run accumulated 356 fills. The analytics were run to check viability against the venue's fee schedule, which was ordinary housekeeping rather than a suspicion that something was wrong.

3. The investigation

Four steps, all on the same day, all arithmetic.

Compute the round-trip fee. Compute the win rate required to cover that fee given the observed average win and average loss. Compare it against the theoretical ceiling. Cross-check against what the demo actually produced.

4. The finding

Taker fee            0.06 pct per side, 0.12 pct round trip
Average win          0.08 pct per winning trade
Average loss         0.10 pct per losing trade
 
Breakeven win rate   (0.10 + 0.12) / (0.08 + 0.10)  =  1.222
Theoretical ceiling  1.000

The strategy required a win rate of 122.2 percent to break even.

Not difficult. Not ambitious. Arithmetically unreachable, by a margin of twenty-two percentage points beyond winning every single trade.

The demo confirmed the shape of it. Across 356 fills the observed win rate was 64 percent and net profit and loss was -12.30 USD.

5. What made it invisible

The mechanism is worth stating precisely, because it is not a mistake anyone would defend and it still took twenty-five hours of live data to surface.

Every component of that calculation was known. The fee schedule was public. The average win and average loss were in the analytics. The formula is one line of secondary-school algebra.

What was missing was the habit of putting them together before the work started. The strategy had a signal, the signal had a measurable edge in mid-price terms, and the natural next question felt like "can we improve the win rate" rather than "what win rate would we need".

The first question generates months of work. The second closes the line in an afternoon.

6. What it cost

The demo lost 12.30 USD, which is the entire realised cost.

The real cost is not in the ledger. It is the engineering time spent tuning thresholds on a strategy whose fee structure had already ruled it out, and that figure is not recorded anywhere in the archive.

7. The fix

The programme pivoted from taker execution to passive maker execution, which changes the fee basis:

Maker fee            0.01 pct per side, 0.02 pct round trip
Breakeven win rate   (0.10 + 0.02) / (0.08 + 0.10)  =  0.667

A 67 percent requirement against an expected 70 percent leaves a positive margin. The signal stack was not discarded; the order-flow and toxicity components were repurposed as quote-skew and toxicity inputs to the maker engine.

That pivot did not save the strategy either. The maker line ran for eight and a half days, accumulated 28,119 shadow trades across three variants, and died because median win magnitude of 1.68 to 3.99 basis points sits below the fee floor at every horizon tested. But it died for a different reason, on measured evidence, which is progress.

8. The rule

Two things came out of this incident.

The immediate one was procedural, recorded in the decision document itself: a major pivot must carry a decision record stating rationale, alternatives, refinements, and acceptance criteria. That convention produced the 58 pre-registration and decision documents this record is built from.

The durable one took longer to formalise and became layer 0 of the falsification chain. Before a hypothesis is written, compare the plausible signal magnitude against the cheapest available round-trip cost. If the first is smaller, there is no test to run.

BL-D-008 sets that doctrine. This incident is why it sits ahead of every other layer.

9. What the source does not settle

The decision record does not state how the 25-hour window was chosen, whether the sample spans one instrument or several, or whether the average win and average loss are means or medians.

More significantly, this calculation is itself incomplete by the standard the programme adopted two weeks later. It contains a fee term and no spread term. The breakeven figure of 122.2 percent is therefore optimistic, and the true requirement was higher. The next incident in this book is the discovery of that omission.

10. Cross-references

BL-D-008 cost-first doctrine. BL-D-001 section 3, layer 0. BL-IN-002, the missing cost term. Ledger rows T0 and VAC.

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