#16 — CMA-ES + regime gates

Hypothesis — what it tests

Whether regime inputs (trending / high-vol binaries) let a small evolved net gate its entries to profitable environments and clear the retail spread on M5.

Description

Following a null result on raw close-difference inputs, the project pivoted from NEAT to cheap CMA-ES fixed-topology searches to probe which inputs and constraints could produce a post-RCA edge on EUR_JPY M5. A parametric script added M5-native regime inputs (ATR14 + EMA24-slope binaries thresholded by a 1-week rolling median), a causal validator that perturbs future bars, latching entries, and stacked hard gates (1,000+ trades minimum, direction ratio ≥ 0.15) so 'don't trade' could not win. Eleven configurations across 3 topologies, 2 pairs and 3 fitness modes produced zero profitable OOS genomes; the best was USD_JPY h4+skip 6-input at −2.38 p/d. An ablation showed the regime binaries were load-bearing filters (removing them made OOS 6× worse, −17.07 p/d), and adding an H1-slope input made every topology worse. Verdict: apparatus validated, the 'M5 momentum + binary regime + position state' input class carries only ~10% of the spread in raw edge.

Key result

Indicators

Algorithms

Code