Paper: arXiv 2508.17086
Authors: Yushi Lin, Peng Yang
Abstract
Trade-based manipulation (TBM) undermines the fairness and stability of financial markets drastically. Spoofing, one of the most covert and deceptive TBM strategies, exhibits complex anomaly patterns across multilevel prices, while often being simplified as a single-level manipulation. These patterns are usually concealed within the rich, hierarchical information of the Limit Order Book (LOB), which is challenging to leverage due to high dimensionality and noise. To address this, we propose a representation learning framework combining a cascaded LOB representation architecture with supervised contrastive learning. Extensive experiments demonstrate that our framework consistently improves detection performance across diverse models, with Transformer-based architectures achieving state-of-the-art results. In addition, we conduct systematic analyses and ablation studies to investigate multilevel manipulation and the contributions of key components for detection, offering broader insights into representation learning and anomaly detection for complex time series data.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced deep learning architectures (Transformers) and contrastive learning with formalized loss functions, indicating high mathematical complexity. It also includes extensive experiments, ablation studies, and mentions real-world datasets, demonstrating strong empirical rigor.
Research Flowchart
flowchart TD A["Research Goal:<br>Detect Multilevel Trade-Based<br>Manipulation in LOB"] --> B["Input: Limit Order Book<br>Time-Series Data"] B --> C["Key Methodology:<br>Cascaded LOB Representation<br>+ Supervised Contrastive Learning"] C --> D["Computational Process:<br>Transformer Architecture<br>Processing Hierarchical Features"] D --> E["Outcome 1:<br>State-of-the-Art<br>Spoofing Detection Performance"] C --> F["Outcome 2:<br>Ablation Studies &<br>Multilevel Pattern Analysis"] E --> G["Final Contribution:<br>Generalizable Framework<br>for Complex Time Series"] F --> G
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a cascaded LOB representation architecture combined with supervised contrastive learning to detect multilevel spoofing, achieving state‑of‑the‑art results with Transformer models.
Source sentence
We propose a novel LOB-based representation learning framework that enhances multilevel manipulation detection across diverse models, achieving state-of-the-art performance with Transformer-based architectures.
Assumes 2
- Anomalous (manipulation) samples are extremely rare, leading to severe data imbalance. data
Source sentence
In practice, a key challenge is the severe data imbalance, where anomalous samples are extremely rare.
- The selected LOBSTER data contain no real market manipulation events, so injected anomalies are used for evaluation. market
Source sentence
Through careful examinations, all selected data do not contain any reported market manipulation events.
Data
- LOBSTER dataset for CSCO, TSLA, INTC order-book public
Source sentence
For our study, we selected three stocks representing different industries and liquidity characteristics: Cisco Systems (CSCO), Tesla (TSLA), and Intel (INTC) on January 2, 2015, with millions of entries providing sufficient data for our study.
Costs
Not applicable
Finds 5
- The proposed framework consistently improves detection performance across diverse models, with Transformer‑based architectures achieving state‑of‑the‑art results. performance
Source sentence
Experimental results demonstrate the effectiveness of our approach: our framework consistently improves detection performance across diverse models, with Transformer-based architectures achieving state-of-the-art results.
- Multilevel manipulation is harder to detect than single‑level, yielding lower AUC‑PR scores for models trained on multilevel insertions. performance
Source sentence
Figure 2 reveals a consistent performance gap between models trained on single-level versus multilevel insertions, with the former achieving higher AUC-PR scores.
- Removing either the contrastive loss or the MSE loss sharply degrades performance, confirming their complementarity. performance
Source sentence
Performance drops sharply when either loss is removed (α = 0 or 1), confirming their complementarity.
- LOB’s hierarchical information can be effectively leveraged through representation learning, improving detection. performance
Source sentence
We empirically show that LOB’s hierarchical information can be effectively leveraged through representation learning.
- Ablation studies clarify that both the cascaded LOB architecture and the combined loss contribute complementary gains to detection. performance
Source sentence
Ablation studies further clarify the complementary contributions of the cascaded LOB architecture and the combined loss with limited oversampling.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 29 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.