#15 — Deep2 mean-reversion multi-seed

Hypothesis — what it tests

Whether a deeper fixed-topology neural net can extract a deployable mean-reversion edge from an M5 z-score of price-minus-SMA10 on EUR_JPY.

Description

After the lookahead RCA wiped all pre-RCA genomes, the project probed whether a deeper fixed-topology net could find a mean-reversion edge on causal features. The feature was a percentile-scaled z-score of (close − SMA10) with a 10,000-bar rolling window, chosen over tanh because EUR_JPY z can reach ±29σ. The 'Deep2' architecture (4 inputs → 4 → 3 → 3, 54 parameters, activation-as-gene) was trained with CMA-ES across 4 seeds on EUR_JPY M5; every seed showed OOS better than IS (mean IS −1.45 vs OOS +1.15 p/d) — an unusual pattern for pure noise and the best 'clean' causal result to that date. A parallel 'Bottleneck' topology (4→4→1+skip→3) failed badly (OOS mean −5.73). The signal never passed the hard walk-forward gate (one of three chunks always unprofitable) and its magnitude sat far below EUR_JPY's 2.3-pip spread; it was later validated retrospectively as the fingerprint of the mean-reversion edge that Path D exposed architecturally.

Key result

Indicators

Algorithms

Code