#17 — Slot-4 indicator sweep (43 indicators)

Hypothesis — what it tests

Whether some 4th-input indicator, swapped into the fixed V3 CMA-NN core, carries enough signal to produce a positive OOS edge net of spread.

Description

With the V3 core fixed (mc_d_a, mc_dd_a, Kaufman ER, UPnL), the project asked whether ANY single additional indicator carries deployable signal, by porting 43 candidates to the causal incremental-feature builder and running each through an identical CMA-ES pipeline (tier-1: 2 pairs × 2 seeds × 100 generations each). None crossed 0 p/d OOS; the best average was CCI at −1.62 p/d, and 7 candidates cleared the −2.0 promotion threshold to tier-2. Pattern analysis showed trend-strength features cluster at the top while oscillators, position-in-range, and seasonality features consistently fail. A side finding: SHAP importance from the earlier LightGBM zigzag study had no predictive power for actual trading edge (SHAP #1 TEC_5 ranked #36 of 43 in the sweep). The session shipped the 43-feature causal port of lib/incremental_features.py plus a reusable parametric sweep harness.

Key result

Indicators

Algorithms

Code