LEAF: A Leakage-Aware, Explainable Framework for Market-Abuse Surveillance in the Saudi Capital Market with an LLM-Based Proof of Concept for Arabic Enforcement-Decision Labeling

This repository is the reproducibility package for the study “LEAF: A Leakage-Aware, Explainable Framework for Market-Abuse Surveillance in the Saudi Capital Market (Tadawul)”. It contains the full analysis pipeline supporting the paper’s empirical results: causal feature construction, unsupervised and weakly supervised detection (Isolation Forest, gradient boosting, autoencoders), SHAP-based explanation, and the leakage-aware validation harness — including the walk-forward protocol, the data-leakage audit, and the label-deconfounding test that constitute the framework’s methodological contribution. Contents: the analysis notebook; feature-engineering, detection, explanation and validation code; figures; and the aggregated result tables (per-company AUC, naive-vs-controlled leakage comparison, and operating-point precision/recall/FPR). Data. The study uses publicly available daily OHLCV data for the fifteen most-traded Tadawul companies (2001–2020), obtained from a public Kaggle dataset, and public enforcement decisions from the Committees for the Resolution of Securities Disputes (CRSD). No confidential or market-restricted data is included in this package. Phase-A labels are a transparent heuristic proxy, documented in the code; confirmed-case labels and tick-level data (Phase B) are not part of this release. Reproducibility. The notebook runs on the public dataset and reproduces the reported tables and figures. Environment and dependencies are specified in the repository. The deconfounding and leakage-audit procedures can be applied to other datasets as described in the paper.