#9 — IronNet V3 H1 12-pair + V7

2026-04-09

Verdict: negative · ⛔ "entirely artifact"

Full 12-pair H1 NEAT training produced spectacular MC-validated backtests (V3 avg 178.9, V7 avg 371.0, S1 747.1 p/d) and three live deployments — all of it lookahead artifact.

Hypothesis — what it tests

Does the H1-cadence NEAT edge generalize across all 12 pairs, and does SHAP-guided input selection (V7) or adding the sb_a swing-breakout feature (S1) improve it into a deployable system?

Description

The H1 V3 result was scaled to all 12 pairs on three Hetzner servers, every pair passing Monte Carlo at sign_p=0.0000 (avg 178.9 p/d, 4.3x the M5 average of 42.0). A second variant, V7, used seven SHAP-selected inputs (bb_width, stoch_d, macd_hist, range_pos_30, aroon_osc, mc_d_a + UPnL) and averaged 371.0 p/d; a shallow-retry variant S1 (V3 core + sb_a swing breakout) hit 747.1 p/d on EUR_JPY. V3, V7 and S1 were deployed live at 1 unit on accounts 002, 003 and 009 for A/B/C comparison. Live they bled (−159, −754.6 and −389 pips respectively), and the RCA declared the backtest numbers 'entirely artifact' of the compute_mc_on_series lookahead bug.

Key result

V3 H1 avg 178.9 p/d (12/12 MC p=0.0000), V7 avg 371.0, S1 747.1 p/d EUR_JPY; live: V3 −159.1p/90t, V7 −754.6p/198t, S1 −389.0p/165t — all lookahead artifact

Indicators

ASIMC(D)MC(dD)Kaufman Efficiency Ratio (er_norm)Bollinger Band widthStochastic %DMACD histogramRangePositionAroon Oscillatorsb_a (ASI swing breakout)UPnL

Algorithms

NEAT neuroevolutionLightGBM + SHAP feature selectionMonte Carlo shuffle validationwalk-forward staged fitness

Code

research/experiments/asi_mc/deploy_h1_perpair.sh research/experiments/asi_mc/collect_h1_perpair.sh research/experiments/asi_mc/train_ironnet_v7.py research/experiments/asi_mc/export_v7_training_data.py research/experiments/asi_mc/deploy_v7_h1.sh research/experiments/asi_mc/neat_config_7in_3out.ini research/experiments/asi_mc/results/ironnet_v7

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

← #8 H1 V3 (EUR_GBP + CAD_JPY)#10 CMA-NN with sin basis →

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