Improving Cryptocurrency Pump-and-Dump Detection through Ensemble-Based Models and Synthetic Oversampling Techniques

Battle Count: 91

The paper is moderately relevant to quantitative trading. While it does not directly propose trading strategies, the detection of pump-and-dump events is critical for risk management in crypto trading. Quantitative traders can use such models to avoid manipulated assets, implement protective stop-loss mechanisms, and identify artificial price movements. The real-time detection capability (sub-hourly data streams) is particularly relevant for high-frequency crypto trading. However, the focus is on surveillance and regulatory compliance rather than alpha generation or portfolio optimization.

Implementation is moderately straightforward. All five ensemble models (RF, AdaBoost, GBM, XGBoost, LightGBM) are available in standard Python libraries (scikit-learn, xgboost, lightgbm). SMOTE is available via imbalanced-learn. The feature engineering pipeline (25-second chunking, 7-hour sliding window, 9 features) is well-defined. The main complexity lies in data collection from exchange APIs and ensuring proper temporal splitting to avoid data leakage. Training times are very short (3-20 minutes), making experimentation feasible.

The study uses an open dataset from La Morgia et al. (2020) collected from Binance via API. All five ensemble models are standard implementations (scikit-learn, XGBoost, LightGBM libraries). Feature engineering pipeline (25-second chunks, 7-hour sliding window) is clearly described. However, no code repository is explicitly linked, and exact hyperparameter settings are not fully detailed. The dataset split (30/70) and SMOTE application are well documented.