Intelligent Detection of Securities Market Manipulation Using Multimodal Financial Data (Applied and Computational Engineering)

Summary

• A multimodal framework fuses one-minute market microstructure series, corporate announcements, financial news and relationships among securities to detect manipulation on China's A-share market.p.1
• The framework achieved macro-F1 of 0.872 ± 0.009, AUROC of 0.946 ± 0.006 and AUPRC of 0.821 ± 0.012 on unseen data.p.1
• In the 2025 surveillance simulation, reviewing the top 10% of ranked observations detected 0.643 ± 0.031 of manipulations with a 0.074 ± 0.009 false-positive rate.p.1
• Enforced China Securities Regulatory Commission cases from 2018–2025 were combined with unenforced stock-period observations.p.1
• Each positive instance is matched against four negative instances; the window period is 20 days before and 10 days after the event center date.p.3
• Temporal, textual and graph embeddings are projected into a common 256-dimensional space and fused by cross-modal attention.p.4

Key insights

1. Removing text information lowered macro-F1 by 0.036 ± 0.007 and removing graph information lowered it by 0.028 ± 0.006.p.1
2. Relative to the market-only TCN, the full system improved macro-F1 by 0.071 ± 0.011.p.5
3. The false-positive rate fell from 0.119 ± 0.012 for the market-only TCN to 0.074 ± 0.009 for the full system.p.5
4. Statistical differences were tested with paired bootstrap resampling using 2,000 replications.p.5
5. Manipulation windows had standardized order-book imbalance of 1.72 ± 0.38 compared with 0.61 ± 0.27 in control groups.p.6
6. Reviewing the top 10% of 2025 ranked events captured 0.643 ± 0.031 of confirmed events, compared with 0.108 ± 0.018 for random screening.p.6
7. With a 20% review budget, capture was 0.817 ± 0.024 while the false-positive rate stayed at 0.096 ± 0.011.p.6
8. In the 30-day forward shift experiment, AUROC was 0.921 ± 0.008 and AUPRC was 0.786 ± 0.015.p.6