BL-IN-003: The Look-Ahead Beta

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Incident date: 2026-06-25 Class: K5, point-in-time violation Rule produced: the walk-forward harness as a permanent, mandatory instrument

1. What was believed

Strategy B was a pairwise mean-reversion model on cointegrated instruments. Entry at two sigma of the spread, exit at half a sigma, stop at three and a half, four-hour bars, hedge ratio by ordinary least squares.

It was believed viable on the strength of one number. Backtested on a clean instrument universe, it reported +180 basis points net at an 80 percent win rate. That figure is described in the decision record as the claim that justified the strategy.

Live shadow, on the same clean universe, was returning -6.4 basis points.

Both numbers were on the table at the same time and the contradiction had been absorbed rather than resolved.

2. The trigger

A demand for a precise, evidence-based decision on a strategy that had been persistently troublesome, computed from a verified clean sample rather than from assumption.

3. The investigation

1. Pull all 105 live shadow trades. Split into the full set and the clean crypto-only subset of 79.
2. Compare the clean subset's live result against the backtest's claim on the same universe.
3. Read the backtest's pair-selection and beta-fitting code to find the source of the gap.
4. Write a walk-forward harness that fits on trailing data only.
5. Re-run the same clean universe through it.
6. Test the inverse hypothesis, spread momentum, through the same harness.

4. The finding

The defect was a single line. Pair selection computed the hedge ratio as an ordinary least squares fit across the entire price panel, and the resulting coefficient was then used to simulate trading over that same panel.

beta, alpha = ols(logP[:, i], logP[:, j])       fitted on the full window
                                                 then traded on the full window

The backtest did contain an out-of-sample check. It gated which pairs were selected, using a seventy-thirty split to screen for drift. It never touched the profit and loss simulation. The presence of a split in the code created the appearance of out-of-sample discipline while the quantity that mattered, the hedge ratio, was fitted on the future.

5. The correction, in code

The walk-forward harness differs in three places, and the differences are worth reading because they generalise to any parameter-fitting strategy.

Selection takes an explicit training bound, so fitting can never see beyond it:

Ptr = P if train is None else P[:train]

The in-sample simulator, whose own comment acknowledged that beta was fitted on the full window and traded on that same window, was replaced. The walk-forward simulator refits at every bar on the strictly trailing window [t-L, t-1], freezes beta, mu and sigma at entry for the life of the position, and only then evaluates exit, stop and timeout.

The cost model was aligned to live economics: a two-leg round trip at 11 * (1 + beta) basis points plus per-leg spread.

6. What the numbers say

The backtest missed the live outcome by more than two hundred basis points. The walk-forward figure missed it by roughly thirty to forty, and erred toward pessimism.

That gap is the entire value of the correction. The walk-forward harness did not find an edge the backtest had missed. It produced an honest negative where the backtest had produced a confident positive, and an honest negative is what stops capital from being committed.

It is worth being precise about the residual. Walk-forward predicted -55 to -68. The full live set returned -25.5 and the clean live subset -6.4. The harness was conservative rather than accurate, and this record does not claim that walk-forward reproduces live results. It claims that walk-forward gets the sign right, which is the decision the number is used for.

7. The inverse hypothesis

The same harness was pointed at the opposite claim: that at extreme spread divergence the spread trends rather than reverts. It returned +20.7 basis points median, +40.7 mean, at a t-statistic of +1.80.

That is not significant, and it was recorded as a lead rather than an edge and never deployed. It appears in the Kill Ledger as SPR-MOM, verdict NOT PROVEN.

Recording it that way was the correct outcome and it is worth noting because the temptation ran the other way. The programme had just lost a strategy and had a promising-looking replacement produced by the very instrument that killed the first one. The threshold held.

8. The fix

Strategy B disabled by configuration, brain restarted with a configuration-only mount and no rebuild, zero subsequent activity confirmed. The 105 live shadow trades frozen for the archive.

9. The rule

The walk-forward harness was declared a permanent validation instrument, and mandatory for validating any future mean-reversion or momentum strategy.

The generalisation is BL-D-011 section 3: parameters are part of the signal. A hedge ratio, a mean, a standard deviation, and a rank threshold are all estimated quantities, and all of them must be measurable at decision time. Written as an equation the defect in this incident is a visible subscript. Written in prose it survived review, implementation, and a live deployment.

10. What the source does not settle

The record does not date the first run of the flawed backtest or the start of the live shadow accumulation, so the total time the +180 figure was trusted cannot be reconstructed. It also does not state whether the spread-momentum lead was ever picked up by later research; the Kill Ledger indicates it was not.

11. Cross-references

BL-D-011 section 3. BL-D-010 section 2.5. BL-D-009 rule R-06 on the distinction between a lead and an edge. Ledger rows COINT and SPR-MOM.

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