#8 — H1 V3 (EUR_GBP + CAD_JPY)

2026-04-08

Verdict: negative · ⛔ lookahead-contaminated

Running the IronNet V3 NEAT genome at H1 cadence instead of M5 tripled-to-quintupled backtest pips per day on EUR_GBP and CAD_JPY — numbers later shown to be entirely lookahead artifact.

Hypothesis — what it tests

Does evaluating the same NEAT genome at H1 cadence (24 decisions/day instead of 288) reduce noise-driven false entries and spread cost enough to multiply the edge?

Description

IronNet V3 was a NEAT-evolved network trading on the project's ASI-derived momentum-consensus features. This experiment retrained it on H1 bars instead of M5, on the hypothesis that hourly cadence cuts intraday-noise false entries and spread burn (~12x fewer decisions per day). EUR_GBP improved from 14.3 to 41.2 OOS pips/day (+188%) and CAD_JPY from 41.6 to 218.1 (+424%), both passing Monte Carlo with sign_p=0.0000 on 10,000 shuffles. The 2026-04-12 RCA later proved the underlying compute_mc_on_series indicator leaked up to 719 future bars, so the entire result was lookahead artifact.

Key result

EUR_GBP 14.3→41.2 p/d OOS (+188%), CAD_JPY 41.6→218.1 p/d (+424%); MC sign_p=0.0000 both pairs; 13.6K H1 bars — all later proven lookahead artifact

Indicators

ASIMC(D)MC(dD)Kaufman Efficiency Ratio (er_norm)UPnL

Algorithms

NEAT neuroevolutionMonte Carlo shuffle validation

Code

research/experiments/asi_mc/train_ironnet_perpair.py research/experiments/asi_mc/mc_validate_ironnet.py research/experiments/asi_mc/results/ironnet_h1 lib/asi_indicator.py

Interactive version (search, filters, figures): the experiment explorer.

← #7 LightGBM/SHAP indicator ranking #9 IronNet V3 H1 12-pair + V7 →

This page is one experiment from the audited record behind the book The Edge That Wasn't — get the book · about the project.