Predictive Information Structure in Equity Markets: An Interpretable Machine Learning Perspective

While much of the financial prediction literature has focused on improving forecasting performance through increasingly sophisticated machine learning models, comparatively less attention has been devoted to understanding how predictive information is organized across interpretable feature families. This study investigates the hierarchical organization of predictive information in Tehran Stock Exchange equities using an interpretable financial machine learning framework. To address this gap, the study develops a unified empirical framework that combines predictive modeling, explainability analysis, feature family aggregation, out of sample ablation analysis, trading based validation, clustering based structural evaluation, and statistical inference. Variables constructed from daily price and volume data for Tehran Stock Exchange equities covering the period from March 26, 2022 to November 21, 2025 are organized into feature families representing momentum, volatility, liquidity, trend structure, and market microstructure. Predictive relevance estimated through complementary explainability methods is subsequently used to examine how predictive information is organized across these feature families. The empirical analysis evaluates four machine learning models, including Logistic Regression, Random Forest, XGBoost, and CatBoost. The results reveal a stable and interpretable hierarchy of predictive information across feature families. Across machine learning models, explainability methods, and complementary validation procedures, the microstructure feature family consistently emerges as the dominant source of predictive information, followed by momentum and liquidity related feature families. Cross method explainability analysis demonstrates strong agreement between SHAP and permutation based attribution rankings, indicating that the observed hierarchy is not driven by a single explainability framework. Out of sample ablation analysis further shows that dominant feature families retain substantial standalone predictive information, with the dominant configuration achieving predictive performance comparable to, and on average slightly stronger than, the full feature configuration. In particular, microstructure dominated equities exhibit strong standalone predictive retention, whereas momentum and liquidity dominated equities rely more heavily on complementary predictive information from other feature families. Trading based validation provides complementary evidence by showing that the identified hierarchy is also reflected in out of sample trading performance, particularly for microstructure dominated equities. Clustering and industry overlap analyses reveal weak correspondence between feature family structures and conventional industry classifications, suggesting that the identified hierarchy is broadly shared across the market rather than strongly industry specific. The study contributes to interpretable financial machine learning by introducing and validating a feature family perspective for investigating how predictive information is organized within financial feature spaces. More broadly, the proposed framework demonstrates that predictive information in Tehran Stock Exchange equities is hierarchically organized rather than uniformly distributed, providing a scalable and interpretable framework that can be applied to investigate predictive information organization across equity markets.