#7 — LightGBM/SHAP indicator ranking

Hypothesis — what it tests

Whether SHAP importance from a LightGBM classifier on zigzag swing labels can cheaply rank indicators by their usefulness as NEAT trading inputs, avoiding brute-force combinatorial training.

Description

With 11+ candidate indicators and ~3h of NEAT training per combination, testing all 165 triplets was impossible, so a LightGBM multiclass classifier was trained on zigzag BUY/SELL/FLATTEN labels and SHAP values used to rank features, pairs and triplets. The first pass (12 indicators including a newly added Aroon Oscillator, 12 pairs) put aroon_osc #1 at M5 (SHAP 0.430) and er_norm #1 at H1, and showed M5 and H1 are different signal regimes (swing structure vs momentum). A comprehensive follow-up (33 classic indicators x 12 pairs x 393K M5 bars over 5 years) ranked bb_width #1 (0.427) with classic oscillators beating the ASI-MC derivatives. But the ranking failed as a guide to trading: TEC_5, SHAP #1 (0.490), scored fitness −10 as a NEAT input and ranked #36 of 43 in the later CMA slot-4 sweep, while sb_a — SHAP dead last at #33 (0.004) — was the key feature in the (lookahead-era) 747 p/d S1 run. Verdict: SHAP measures classification power on labels, not entry-timing profitability — the two are orthogonal.

Key result

Indicators

Algorithms

Code