#69 — ESCMA exit-learner (shock-entry + CMA-NN exit)

Hypothesis — what it tests

Whether a learned exit on automatic (shock-triggered or random) entries can manufacture a trading edge by itself — i.e. whether the exit, rather than the entry, can be the lever.

Description

Instead of learning when and which way to trade (the hardest problem), this arc pinned the entry with a deterministic momentum-shock gate (σ-normalized: ≥2 of the 5m/15m/1h windows beyond |2σ|, 9,751 USD_JPY S5 events over 5.5 years) and trained a small 8-hidden-unit neural network via CMA-ES to learn ONLY the exit, rewarded on AMDDP5 (PnL minus λ times the area under the per-trade drawdown curve). The pipeline was rigorous — index-only causal features with a bit-identical R7 lookahead proof, multi-seed validation, JIT-parallel evaluation — and it correctly caught its own mirage: the single-seed 'star' did not replicate, and at convergence the learned exit loses to the trivial exit-at-first-negative-tick baseline (+22/+167 vs +437 AMDDP5). The load-bearing result is the Phase-0 sine positive control: on a synthetic sine market the same learner finds the provably optimal exit exactly (holds 181 bars vs optimum 180, 100% WR, OOS +14,440p), survives 10x-spread noise, and per-trade PnL tracks amplitude−spread to the pip — so the apparatus works and the deficiency is entirely the entry signal. Phase 1 (momentum-continuation entries, multi-seed) and Phase 2 (a LightGBM screen finding ~zero forward predictability in the features, best decile move 0.02–0.08p vs the 1.7p spread) closed the arc definitively: random/auto-entry exit-alpha is a NO, and a bonus structural finding is that bump activations (gauss/morlet/ricker) are necessary for exit-at-the-extreme decisions where sin/cos/tanh fail entirely.

Key result

Indicators

Algorithms

Code