Optimizing Bitcoin Anomaly Detection: A Comparative Analysis of Feature Selection Techniques

Abstract

This paper investigates the efficiency of feature selection methods in the context of enhancing unsupervised anomaly detection in cryptocurrency markets. By taking into account historical Bitcoin data from 2013 to 2024, we make use of a process involving data normalization, feature importance selection, and model evaluation to do research. We preprocess several features comprising price returns, trading volume, and market capitalization and also generate technical indicators such as volatility and momentum metrics. We perform feature importance by two methods: Principal Component Analysis (PCA) and Random Forest feature importance that are mutually supportive. Later on, we measure the detection performance of Isolation Forest and Local Outlier Factor (LOF) algorithms for the different feature subsets. From our experiments, we see that the features selected by the Random Forest reached the highest model agreement (95.90%) and anomaly overlap (54.73%) that is far better than that of both the complete feature set and PCA-selected features. Short-term volatility indicator, price momentum as well as the number of trades appeared as very informative features for the detection of market anomalies. Theadetected anomalies were in line with Bitcoin market major events such as the period of 2017-2018 bubble burst, March 2020 market crash, and 2021's bull run. These results show that if feature subsets are precisely chosen, they can improve anomaly detection consistency in cryptocurrency markets and can even supply market surveillance and risk management systems with new insights.