AI-Driven Predictive Modeling for Detecting Suspicious Trading Patterns, Anomalous Order Activity, and Market Manipulation in U.S. Equity Markets

Abstract

Modern U.S. equity surveillance increasingly depends on predictive systems that flag suspicious order placement, cross-venue quote behavior, and abnormal execution patterns before manipulative episodes fully unfold. This paper develops an AI-driven research framework for detecting suspicious trading patterns, anomalous order activity, and market manipulation in U.S. equity markets by combining market microstructure theory, public regulatory evidence, and contemporary machine-learning methods. Drawing on more than forty professional sources, the study positions surveillance as a multimodal prediction problem spanning limit-order-book dynamics, order-routing behavior, venue fragmentation, trader interaction networks, news and disclosure context, and enforcement-informed labels. Market-structure evidence shows that U.S. equity average daily volume rose from an estimated 11.1 billion shares in 2023 to 12.2 billion in 2024 and 17.6 billion in 2025, while off-exchange Trade Reporting Facility share increased from about 45% in April 2023 to 47.0% in 2024 and 50.6% in 2025, underscoring the scale and fragmentation challenge facing surveillance teams. The paper proposes hybrid architecture that integrates gradient-boosted baselines, temporal transformers, and graph neural networks with explainable AI overlays for analyst review. Methodologically, the framework emphasizes severe class imbalance, regime shifts, weak and delayed labels, and cost-sensitive evaluation anchored in precision at alert budgets, time-to-detection, and economic materiality rather than headline accuracy alone. The paper argues that trustworthy surveillance requires predictive power, governance discipline, auditable features, and human escalation aligned with U.S. market-structure reforms. The result is a blueprint for more adaptive and useful market-abuse detection.