The Edge That Wasn't
All 81 experiments › #13

#13 — Causal retrain verdict

2026-04-12
Verdict: negative · ⛔ zero causal edge
Retraining the best architecture on fixed, causal features collapsed the edge to +0.11 p/d at best — proving the entire pre-RCA edge was lookahead and none of it survives.

Hypothesis — what it tests

Does any of the V3/macd_hist feature-family edge survive once the lookahead bug is fixed and all features are computed causally?

Description

The decisive control experiment after the RCA: CMA-NN V3+macd_hist was retrained on causal features from the new FXFeatureBuilder (Kalman10 smoother, no lookahead) across 4 pairs x 4 seeds = 16 runs. Best of 16: +0.11 p/d on 13 trades over 410 OOS days — statistically nothing; worst: −12.23 p/d over 2,586 trades. The optimizer either found a barely-trades 'safe' solution or a trade-active losing one, with no middle ground. A parallel causal sweep of upstream smoothers (sma5/ema3/ema5/rma5) failed on every variant, and removing lookahead alone collapsed CHF_JPY from +73 to −1.56 OOS on the identical architecture. All trading was stopped, Hetzner servers deleted, and the project's rule 'do NOT restart any pre-RCA genome' was set. The write-up explicitly weighed accepting the null result: a year of backtest optimization had mostly found lookahead.

Key result

Best of 16 causal runs +0.11 p/d (13 trades / 410 OOS days); worst −12.23 p/d (2,586 trades); CHF_JPY +73 → −1.56 OOS with lookahead removed; 4/4 causal AMA smoother variants FAIL

Indicators

MC(D)MC(dD)Kaufman Efficiency Ratio (er_norm)MACD histogramKalman filter smoother (Kalman10)

Algorithms

CMA-EScausal incremental feature pipelinemulti-seed control studyKalman filter

Code

lib/incremental_features.pyresearch/experiments/cma_5in/build_causal_parquets.pyresearch/experiments/cma_5in/adaptive_ma_sweep_causal.pyresearch/experiments/cma_5in/train_cma_v2.pyresearch/experiments/cma_5in/test_cma_causal.py

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

← #12 RCA — lookahead + MACD bug#14 S5 momentum NEAT + PnF →
This page is one experiment from the audited record behind the book The Edge That Wasn't — get the book · about the project.