Anomaly Detection Analysis on Ethereum’s Historical Price Data Using an Unsupervised Machine Learning Approach

The price movement of cryptocurrency assets such as Ethereum exhibits highly volatile and dynamic characteristics, making it prone to anomalies that can affect trend analysis and investment strategies. This study aims to analyze anomaly detection in Ethereum’s historical daily price data using three unsupervised machine learning approaches: Isolation Forest, Local Outlier Factor (LOF), and Autoencoder. The dataset used consists of 1,438 daily closing price data points collected over multiple years to represent the dynamics of the crypto market. The results show that the Isolation Forest detected approximately 32 anomalies, LOF found 29 anomalies, and the Autoencoder identified 26 anomalies. All three methods consistently detected extreme price peaks between December 2017 and February 2018, corresponding to Ethereum’s highest surge before a major correction occurred. Based on the comparative results, the Isolation Forest excelled in efficiency and detection speed, LOF was more sensitive to local variations, and the Autoencoder demonstrated greater adaptability to non-linear patterns and complex volatility. This study concludes that combining these three methods has the potential to produce a more accurate and adaptive hybrid anomaly detection system capable of responding to the ever-changing dynamics of the cryptocurrency market.