#3 — ASI-MC V3 efficiency-ratio gate

2026-04-01 (trained/deployed) — refuted by RCA 2026-04-12

Verdict: negative · ⛔ lookahead artifact

Adding Kaufman Efficiency Ratio as a 4th NEAT input to gate ASI-MC entries to trending regimes produced +84,861p OOS (499 p/d) that was later shown to be entirely a lookahead artifact.

Hypothesis — what it tests

Whether a trend-regime gate (Kaufman ER as a 4th input) lets NEAT trade ASI-MC momentum signals only in trending markets and hold flat in chop.

Description

V3 added the Kaufman Efficiency Ratio (60-bar window, normalized via arctan(ER/0.3)/(pi/2)) as a 4th input alongside MC(D), MC(dD) and unrealized P&L, hypothesizing NEAT would learn to trade momentum-confluence signals only when price is trending and stay flat in chop. Trained on 4 Hetzner servers with different seeds, the best genome (s23, gen 113) scored +84,861 pips OOS over a 170-day window (499.1 p/d, 135.9 trades/day), within 2.3% of the V2 baseline, and passed Monte Carlo sign-shuffle with sign_p=0.0000 across all 12 pairs. Deployed live on account 009, the general genome collapsed to 100% LONG and lost 791 pips over 60 trades in 10 days. The 2026-04-12 RCA found the root cause: compute_mc_on_series in lib/asi_indicator.py leaks up to 719 future bars through a timeframe downsample/index mismatch, so every genome trained on compute_asi_mc saw the future — the entire backtest edge was lookahead.

Key result

OOS +84,861p / 499.1 p/d / 135.9 trades/day (170-day window; V2 baseline +86,833p / 510.7 p/d); MC sign_p=0.0000 x12 pairs; live −791p over 60 trades in 10 days (100% LONG collapse); RCA: up to 719 future bars leaked

Indicators

ASIMC(D)MC(dD)Kaufman Efficiency Ratiounrealized P&L (UPnL)

Algorithms

NEAT neuroevolutionMonte Carlo sign-shuffle validation (10,000 shuffles)

Code

research/experiments/asi_mc/export_d_indicators.py research/experiments/asi_mc/train_from_indicators.py research/experiments/asi_mc/deploy_v3.sh lib/asi_indicator.py services/strategy_neat/main.py

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

← #2 JPY exposure bug fix #4 IronNet fixed-topology NEAT →

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