#5 — Per-pair IronNet V3, 12 pairs

Hypothesis — what it tests

Whether per-pair specialization of the fixed-topology IronNet architecture yields a deployable 12-pair trading system.

Description

Building on the IronNet pilot finding that per-pair specialization beats a general genome (+29% on weak pairs), a fully automated pipeline (sine 30 gens → zigzag 50 gens → walk-forward P&L 150 gens, then collect → MC validate → copy to models/ → update docker-compose → deploy → Telegram alert) trained one 4-input IronNet genome per pair for all 12 pairs. Every pair passed Monte Carlo sign-shuffle at sign_p=0.0000 (10,000 shuffles), averaging 42.0 p/d OOS, from CHF_JPY at 65.5 down to EUR_GBP at 14.3. An A/B deployment ran raw genomes (account 009) against an ER-gated variant (er_norm < 0.35, account 003). Live, ironnet_v3_perpair on account 009 lost 897.5 pips over 165 trades (−5.44 p/trade) against the +42 p/d backtest — the RCA later showed all genomes trained on the compute_asi_mc pipeline had seen up to 719 future bars. This entry is the canonical example that a 12/12 MC pass has no power against a shared lookahead bug.

Key result

Indicators

Algorithms

Code