Disentangling trend and microstructure: A context-aware framework for stock manipulation detection

Detecting market manipulation, such as spoofing and ramping, is increasingly challenging in high-frequency trading environments due to the low signal-to-noise ratio of limit order book (LOB) data. Existing unsupervised anomaly detection models often struggle in this context; by minimizing global reconstruction errors without structural constraints, they tend to indiscriminately approximate both legitimate stochastic market noise and manipulative spikes. To address this limitation, this paper proposes WTM-CAD, a context-aware unsupervised framework that embeds structural inductive biases into the sequence reconstruction process. The model applies a Discrete Wavelet Transform (DWT) to explicitly decompose raw order book sequences into macroscopic trends and microscopic fluctuations. These components are then processed by a dual-stream encoder: a contextual temporal stream validates the consistency between short-term fluctuations and the underlying trend, while a relational interaction stream captures concurrent inter-variable dependencies, such as price-volume dynamics. Evaluated on a real-world dataset constructed from China Securities Regulatory Commission (CSRC) administrative penalty cases, WTM-CAD achieves the highest mean Average Precision (AP) and Recall@Top5% among the compared methods on this benchmark. By constraining representation learning through structural market dynamics, the framework is designed to distinguish candidate manipulation windows from stochastic high-frequency noise.