For external readers: This document is one of 10 curated case-study artefacts from Project Aṣe, a quantitative crypto-trading research project that produced a documented negative result with mechanism across four strategy attempts plus a literature audit, all under pre-committed criteria (Sharpe > 1.0 AND profit-factor > 1.4 OOS, called “§14” internally) that held without bar erosion across 19 ADRs. The full project repo is private (scope B of the Phase 5e shipping spec). Start with ./00-readme.md for the reading guide. Read order context: The second §14-bar kill: Phase 3’s BMR (Bollinger Mean Revert) post-mortem. Different failure mode from Phase 2. The substance of the document below is unchanged from the internal version; only this framing block and personal/identifying content scrubs are added.

ADR-0013: BMR §14 FAIL post-mortem — no live ship, no re-tuning

Context

Phase 3 ran the §14 evaluation protocol for BMR (Bollinger Mean-Reversion) — the second strategy attempt of Project Aṣe after MFD’s Phase 2e FAIL (ADR-0011). The harness output:

VERDICT: FAIL
**Binding constraint:** no OOS candidate passes crit 1-5

The Phase 3 protocol was abbreviated per Plan Option B (recorded in the verdict report): walk-forward (crit 6) was skipped because OOS evidence (15 candidates across 3 seeds × top-5 epochs each) was decisive enough that walk-forward would only have confirmed the result.

Decision

BollingerMeanRevert does NOT ship. The Phase 3 final tag is phase-3-bmr-killed. Phase 7 (paper-trading) is NOT entered for BMR. Phase 4 (new strategy thesis — Project Aṣe’s third strategy attempt) becomes the immediate next phase.

What bound, in plain English

Both Sharpe (crit 3, threshold > 1.0) AND profit factor (crit 2, threshold > 1.4) failed across every OOS candidate (15 of 15). The best OOS Sharpe across the 15 candidates was −0.083 (seed=42 epoch=51), with the worst being −0.361 (seed=9001 epoch=35). Mean OOS Sharpe across the 15: −0.180. Mean PF: 0.59 (the best was 0.85, the worst 0.31). Every candidate was both unprofitable AND low-conviction.

This is a structurally worse failure shape than MFD’s. MFD’s OOS run produced Sharpe ≈ 0.31 with profitable PF > 1.4 on multiple candidates — MFD was profitable-but-inconsistent. BMR is unprofitable-AND-inconsistent.

Diagnostic evidence from the hyperopt phase reinforces the kill:

The deeper finding: the BMR entry signal is dominated by ONE binding indicator (likely the daily EMA-stack macro filter, which is very restrictive on trades). The Bollinger band parameters do not materially shift WHICH candles fire — they only affect the per-trade exit timing within a macro-filter-selected trade set. This is a fundamental thesis-level defect, not a tuning failure.

Why walk-forward was skipped

Plan Option B (recorded 2026-05-11) authorized skipping walk-forward (Plan Task 14) when OOS evidence is decisive. The criteria for triggering Option B were:

Under these conditions, walk-forward (35 min Hetzner CPU + ~5 min of analysis) is statistically guaranteed to confirm rather than refute the OOS-only verdict. Running it would be methodologically thorough but operationally redundant.

This precedent applies to future Phase N strategies IF the same Option B trigger conditions hold. If any of the three triggers fails (e.g. OOS Sharpe is positive but below 1.0; only one crit binds; non-zero overlap exists), walk-forward is REQUIRED — the protocol defaults to belt-and-braces.

The harness’s manifest validator was relaxed in commit 8ac7a4e to accept walk_forward_windows count of either 4 (normal) OR 0 (Option B skip); any other count is still rejected as malformed.

Why not re-tune

PRD §15 principle 2 forbids “kept tuning until it worked.” The protocol made every methodologically allowed move:

A fourth seed, a fifth tuned parameter, or a different exit mechanism (fixed-percent take-profit, time-stop tiebreaker — deferred per spec §9.7) would be re-tuning. The thesis is wrong; we don’t lower the bar.

Cross-strategy comparison: MFD vs BMR failure modes

Dimension MFD (Phase 2e FAIL) BMR (Phase 3 FAIL)
Best OOS Sharpe 0.314 (positive, ~3× below threshold) −0.083 (negative)
Best OOS PF > 1.4 on multiple candidates < 1.0 on all 15 candidates
Hit rate 35-45% (trend-following profile) 66% (mean-reversion as predicted)
Profitability Profitable but inconsistent Unprofitable AND inconsistent
Binding §14 crit Sharpe only Sharpe AND PF
Hyperopt landscape Distinct candidates per seed (no overlap) Degenerate plateaus (638/1000 tied for seed-9001)
Diagnostic Per-trade edge too small for trade frequency Wrong thesis: macro-filter dominates entry signal

Two consecutive strategies have failed §14 in two different ways — the protocol is doing its job, and the project’s discipline is intact (the strategies were killed; the bar was not lowered). The pattern across the two failures suggests that the universe + timeframe (BTC/ETH spot, 4h primary + 1d informative) may be too narrow to support strategies with both high enough Sharpe AND high enough profitability to clear §14. Phase 4’s brainstorm should consider that hypothesis explicitly.

What we keep (carry-over to Phase 4)

The full Phase 1-3 infrastructure remains usable for the next strategy:

What we don’t keep

freqtrade/user_data/strategies/bollinger_mean_revert.py stays in-tree but is no longer the live strategy. Phase 4 produces a new strategy file alongside it; the runtime selector (now in the run scripts as a --strategy flag default) selects which is active.

ADR-0010 (atr_mult range widening) remains MFD-specific. ADR-0012’s decision-criteria lock (BMR-specific) is historical record only.

References