#11 — CMA-NN 12-pair grid + mixed activations

Hypothesis — what it tests

Does the CHF_JPY CMA-NN result generalize across all 12 pairs and hold up across random seeds, timeframes, and deeper mixed-activation architectures?

Description

The +73 p/d CMA-NN config (V3+macd_hist, 7→8(sin)→3) was trained on all 12 pairs for ~3 min/pair ($0.11 of Hetzner compute): every pair positive OOS, avg +55.2 p/d, all Monte Carlo sign_p=0.0000, beating IronNet V3 M5 on all 12 pairs with the edge coming from win/loss asymmetry (avg win 3-4x avg loss) rather than win rate. A 4-seed x 12-pair x 2-TF study showed remarkable stability (M5 mean +54.6 within 1 p/d of the single seed; H1 +40.2 with near-zero seed variance), and a mixed-activation architecture study found a 4-layer sin→mexican-hat→tanh→tanh net with skip connections marginally beating the flat baseline (+74.26 vs +73.03). None of it was deployed: the same-day RCA showed the underlying features leaked future data, hollowing out the entire grid.

Key result

Indicators

Algorithms

Code