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 first §14-bar kill: Phase 2’s MFD (Macro-Filtered Donchian) post-mortem. The substance of the document below is unchanged from the internal version; only this framing block and personal/identifying content scrubs are added.

ADR-0011: MFD §14 FAIL post-mortem — no live ship, no re-tuning

Context

Phase 2e ran the full §14 evaluation protocol (Phase 2e spec §4): widened-atr_mult smoke (seed 9001), branch-decision, OOS candidate-union evaluation across three seeds, 4-window walk-forward per ADR-0008, and the formal scripts/evaluate-pass-fail.py harness. The harness output:

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

Phase 2d already flagged this as the modal expected outcome (Phase 2e spec §7.6). The OOS Sharpe gap from Phase 2d (best 0.27 vs threshold > 1.0) was ~4×; walk-forward + a third seed + atr_mult widening could plausibly close some of that gap but not the full 4×.

Decision

Macro Donchian Filter (MFD) does NOT ship. The Phase 2 final tag is phase-2-mfd-killed. Phase 7 (paper-trading) is NOT entered for MFD. Phase 3 (new strategy thesis) becomes the immediate next phase.

What bound, in plain English

Sharpe (crit 3, threshold > 1.0) was the binding constraint across every single OOS candidate (15 of 15, best 0.314 at seed=42 epoch=4 with params (36, 7, 171, 2.3)) and every walk-forward window (0 of 4 consistent, best W3 Sharpe 0.553 testing Jan–Jun 2024). The best OOS Sharpe of 0.314 was a marginal +0.04 improvement over Phase 2d’s 0.27 — directionally consistent with the predicted “some gap closure but not the full 4×” but nowhere near the threshold. The diagnostic shift confirmed by the stoploss-firing breakdown is decisive: ~100% of exits fire as exit_signal (the donchian_exit lower-band crossing), with near-zero stop_loss and zero roi exits — meaning the strategy’s edge depends on exit-signal timing rather than stoploss tightness, and ADR-0010’s atr_mult widening to 1.5–6.0 was therefore a methodological no-op for §14 outcomes (which is itself useful evidence: the per-trade edge is structurally too small relative to trade frequency, regardless of seed or regime).

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 third widening, or alternative hyperopt-loss functions would be “kept tuning until it worked” — the exact antipattern §15 principle 2 forbids. Future strategies will run the same protocol; if the protocol fails three strategies in a row, that’s evidence the protocol is too strict (revise PRD), not that the strategies were close to working (re-tune them).

What we keep

The full Phase 2 infrastructure remains usable for the next strategy:

The §14 harness in particular is strategy-agnostic — Phase 3’s strategy will reuse it unchanged.

What we don’t keep

freqtrade/user_data/strategies/macro_donchian.py stays in-tree but is no longer the live strategy. Phase 3 produces a new strategy file alongside it; the runtime config selects which is active.

References