#18 — Path B (new core) + Path C (H1 cadence)

Hypothesis — what it tests

Whether a new input core (Path B) or a slower H1 decision cadence (Path C) can lift the slot-4 CMA-NN family out of its sub-zero OOS ceiling.

Description

After the slot-4 sweep exhausted single-indicator swaps on M5, two structural escapes were run in parallel on Hetzner: Path B replaced the fixed core with the sweep's top-4 winners (CCI + BB width + EMA21 ratio + ATR ratio) at M5 cadence, and Path C kept the V3 core but evaluated the net's decisions only at H1 boundaries (every 12 M5 bars, ~24 decisions/day instead of 288, cutting spread burn 12×). Path B was killed at 102/184 jobs with zero positive runs. Path C tier-1 produced 9 positive single runs out of 200 (4.5% vs 0% on M5), and its tier-2 (top 10 candidates × 12 pairs × 4 seeds × 200 gens = 480 jobs) improved the ceiling to −1.82 p/d (two_bar_momentum, n=44) versus M5's −2.21 — an 18% improvement, but still negative with zero candidates averaging above 0. This closed the entire 'causal-CMA on V3-class core + one slot + small NN' paradigm across ~2,500 runs: the per-trade edge is simply too small relative to the spread at 5-pip-class features, and pointed the campaign toward the mean-reversion architecture (Path D).

Key result

Indicators

Algorithms

Code