#14 — S5 momentum NEAT + PnF

Hypothesis — what it tests

Does multi-lag 5-second momentum encode causal microstructure signal invisible at M5 cadence, and can NEAT's free topology exploit it where a fixed-topology CMA-ES net cannot?

Description

After 200+ negative causal tests on the day of the lookahead verdict, a NEAT network fed multi-lag S5 momentum features (price differences at 30-second, 5-minute and 1-hour lags plus a 5-minute second difference, and UPnL/MAE/MFE) went positive on EUR_JPY: IS +2.08 p/d, OOS +2.74 p/d, perfectly bidirectional (9 long / 9 short), +5.75 pips/trade against a 2.3-pip spread, with a tiny 5-node / 8-connection winning topology and OOS>IS as an anti-overfit signal. The same features failed on CMA-NN (gate FAIL, learned not to trade), and Point & Figure variants at 5- and 10-pip boxes either overfit catastrophically (b=5: IS +1.68 → OOS −12.13) or barely traded (b=10). Caveats were huge — single pair (only EUR_JPY had local S5 data, 6 months), single seed, 18 OOS trades, and the genome pickle failed to save on Hetzner — so the result stayed a documented thread rather than a deployment.

Key result

Indicators

Algorithms

Code