When a 1978 Indicator Outperforms Custom Neural Features
ADX values as direct L1 XGBoost features, ranking in the top 5 by SHAP importance in 4 of 6 asset clusters.
ADX in the Feature Vector
S08 is about using ADX not as a filter (that is S04) but as a raw feature in the L1 XGBoost models. The 14-period ADX value is included in the 38-feature vector alongside more sophisticated engineered features like autocorrelation coefficients, Hurst exponents, and microstructure ratios.
What surprised me is where ADX ranked. Across SHAP importance analysis of all six asset clusters, ADX landed in the top 5 most important features for four clusters: FOREX, METALS, INDEX, and COMMODITY. It outranked custom features that took weeks to engineer.
Why Simple Features Win
XGBoost does not care about feature sophistication. It cares about information content. ADX captures trend strength in a single number that interacts well with tree-based splits. When XGBoost sees ADX above 25 combined with RSI above 60, it has identified a strong-trend-with-momentum condition that is highly predictive for certain clusters.
Complex engineered features often encode similar information with more noise. The bid_ask_imbalance feature took two weeks to build and ranked 28th in SHAP importance for FOREX. ADX took five minutes to add and ranked 3rd. Engineering effort does not correlate with predictive value.
Lessons from Feature Humility
The V7 system uses 38 features per bar. Some are sophisticated (Hurst, autocorrelation). Some are classical (RSI, ADX, ATR). The lesson from S08 is that classical indicators survive because they work. A 47-year-old indicator ranking in the top 5 by SHAP importance alongside modern features should make any quant developer humble about the value of complexity. Build the fancy features, test them rigorously, but never dismiss the simple ones because they lack novelty. The market does not care about your engineering effort. It cares about information content.