#10 — CMA-NN with sin basis

Hypothesis — what it tests

Can CMA-ES on a fixed-topology sin-activation network match NEAT's evolved-topology results at a fraction of the compute, and which extra indicator best complements the V3 core?

Description

An alternative to NEAT: a fixed 7→8→3 network trained with CMA-ES, where fixing every hidden activation to sin (weight=frequency, bias=phase, i.e. a learned Fourier basis per Tancik 2020) beat a 9-wavelet activation bank (+50.51 vs +43.15 OOS on CHF_JPY) and trained 22% faster. A 10-candidate indicator sweep found V3+macd_hist the best stack at +73.03 OOS p/d on CHF_JPY, beating IronNet V3's +65.5 in 220 seconds versus hours of NEAT — at 1/100 the compute. A striking side finding: SHAP rank did not predict trading value (tec_5, SHAP #1, ranked 9/10 here), and the strongest singleton features did not compose (adding macd_hist to a range_pos_30 base collapsed IS by 89%). The RCA later showed the +73 was lookahead; the clean causal retrain of the same configuration produced +0.11 p/d — noise.

Key result

Indicators

Algorithms

Code