Intelligent Detection of Securities Market Manipulation Using Multimodal Financial Data

Market manipulation through securities involves increasingly sophisticated mechanisms such as irregular trading, selective information disclosures, and joint activities in different securities, posing threats to surveillance relying solely on a particular information source due to both false positives and late identification of manipulations. We propose a multimodal framework for the detection of market manipulation on China's A-share market by fusing one-minute market microstructure time series, corporate announcements, financial news, and evolving relationships among securities. Enforced cases by China Securities Regulatory Commission from 2018–2025 were combined with unenforced stock-period observations and temporal, textual, and relational features are extracted using Temporal Convolutional Network, MacBERT encoder, and Graph Attention Network, respectively, followed by cross-modal attention to generate a risk score. Our framework achieved macro-F1 of 0.872 ± 0.009, AUROC of 0.946 ± 0.006, and AUPRC of 0.821 ± 0.012 on the unseen data, outperforming the TCN-only model in market only by a wide margin of 0.036 ± 0.007 and 0.028 ± 0.006 for removing text and graph information, respectively. In the 2025 market surveillance simulation, reviewing the top 10% of ranked observations led to detecting 0.643 ± 0.031 of manipulations with 0.074 ± 0.009 false-positive rate.