[Paper Review] Early Rug Pull Warning for BSC Meme Tokens via Multi-Granularity Wash-Trading Pattern Profiling

Dingding Cao, Bianbian Jiao+3 more|arXiv (Cornell University)|Mar 14, 2026

Financial Markets and Investment Strategies0 citations

TL;DR

This paper presents an end-to-end framework that uses multi-granularity wash-trading patterns to build token-level risk features and applies supervised models to provide early rug-pull warnings for BSC meme tokens, with Random Forest outperforming Logistic Regression under weak supervision.

ABSTRACT

The high-frequency issuance and short-cycle speculation of meme tokens in decentralized finance (DeFi) have significantly amplified rug-pull risk. Existing approaches still struggle to provide stable early warning under scarce anomalies, incomplete labels, and limited interpretability. To address this issue, an end-to-end warning framework is proposed for BSC meme tokens, consisting of four stages: dataset construction and labeling, wash-trading pattern feature modeling, risk prediction, and error analysis. Methodologically, 12 token-level behavioral features are constructed based on three wash-trading patterns (Self, Matched, and Circular), unifying transaction-, address-, and flow-level signals into risk vectors. Supervised models are then employed to output warning scores and alert decisions. Under the current setting (7 tokens, 33,242 records), Random Forest outperforms Logistic Regression on core metrics, achieving AUC=0.9098, PR-AUC=0.9185, and F1=0.7429. Ablation results show that trade-level features are the primary performance driver (Delta PR-AUC=-0.1843 when removed), while address-level features provide stable complementary gain (Delta PR-AUC=-0.0573). The model also demonstrates actionable early-warning potential for a subset of samples, with a mean Lead Time (v1) of 3.8133 hours. The error profile (FP=1, FN=8) indicates that the current system is better positioned as a high-precision screener rather than a high-recall automatic alarm engine. The main contributions are threefold: an executable and reproducible rug-pull warning pipeline, empirical validation of multi-granularity wash-trading features under weak supervision, and deployment-oriented evidence through lead-time and error-bound analysis.

Motivation & Objective

• Address the challenge of detecting rug-pull risk in high-frequency BSC meme tokens under weak supervision.
• Develop a reproducible pipeline linking wash-trading patterns to actionable warning scores.
• Identify which feature groups (trade-level, address-level, contract-level) drive predictive performance.
• Provide deployment-oriented metrics such as lead time and error boundaries to inform risk triage.

Proposed method

• Construct 12 token-level behavioral features from three wash-trading patterns (Self, Matched, Circular).
• Aggregate transaction-, address-, and flow-level signals into a token-level risk feature vector.
• Train supervised models (Logistic Regression and Random Forest) to output warning scores and lead-time estimates.
• Evaluate using accuracy, precision, recall, F1, AUC, PR-AUC, and Lead Time, with PR-AUC treated as a key ranking metric.
• Perform ablation to assess contributions of trade-, address-, and contract-level features.

Figure 1: BSC Meme Token Rug Pull Early-warning Framework. The overall pipeline contains four stages (E1–E4): data construction and labeling, wash-trading pattern profiling, early-warning modeling, and ablation/error analysis.

Experimental results

Research questions

• RQ1 Can multi-granularity wash-trading features provide stable early-warning signals for rug pulls under weak supervision?
• RQ2 Which feature groups contribute most to predictive performance and how do they affect recall and precision?
• RQ3 What are the lead times and error profiles when deploying an early-warning model for BSC meme tokens?
• RQ4 How does a non-linear model compare to a linear baseline in capturing on-chain risk patterns?

Key findings

• Random Forest outperformed Logistic Regression on core metrics (AUC 0.9098, PR-AUC 0.9185, F1 0.7429).
• Trade-level features are the primary performance driver (PR-AUC drops by 0.1843 when removed).
• Address-level features provide stable complementary gain (PR-AUC drops by 0.0573 when removed).
• Current contract-level feature shows limited discriminative value (PR-AUC increases by 0.0077 when removed).
• Lead Time (v1) has a mean of 3.8133 hours for a subset of cases, with a median around 1.0331 hours.
• Error profile shows FP=1 and FN=8, indicating high-precision screening rather than high-recall automation.

Figure 2: Data Collection and Labeling Pipeline. The workflow includes export, token-wise merging, deduplication, normalization, window capping, rule-based labeling, and quality-control checks.