Mitigating Class Imbalance in Pump-and-Dump Detection: A Comparative Analysis of Imbalance-Aware Algorithms

The cryptocurrency market has been subject to various forms of exploitation, with pump-and-dump schemes being a persistent problem that can lead to signif- icant price volatility and financial losses. Prior studies have focused on predicting attack targets using machine learning approaches; however, these methods are often compromised by class imbalance and biases inherent in historical data. To address this limitation, we present a novel framework for manipulation targets prediction. We employ two-step bias-minimizing data normalization techniques to mitigate the effects of class imbalance and temporal heterogeneity. We also pro- pose several innovative machine learning approaches including imbalance-aware hyperparameters optimization and ranking algorithms. Our evaluation reveals that proposed framework outperforms existing methods, achieving an accuracy rate of 5% and a top-10 accuracy rate of 45% on a highly imbalanced dataset of historical P&Ds events(imbalance ratio ≈ 270). For further validation, we con- struct a portfolio strategy based on predicted target class probabilities to assess the practical application of our constructed models.