Exploring machine learning-based methods for anomalies detection: evidence from cryptocurrencies

Studies on forecasting cryptocurrency price anomalies have been narrowly adopted, with inconclusive results. This paper addresses anomalies detection in cryptocurrency returns by exploring three advanced machine learning methods: Local Outlier Factor, Isolation Forest and One-Class Support Vector Machine. The study spans over the period ranging from January 2013 until September 2024. To assess the accuracy of these methods, we adopt five metrics: Mean Absolute Error, Explained Variance Score, Root Mean Squared Error, recall, and precision. The results reveal that Local Outlier Factor (LOF) emerges as the most effective method for detecting outliers in cryptocurrency returns, surpassing Isolation Forest (IF) and One-Class Support Vector Machine (OCSVM) in both directional and predictive accuracy, and across all metrics. Local Outlier Factor accurately identifies true anomalies, while minimizing false positives, performing consistently well with both stable (Bitcoin) and volatile (Dogecoin) assets. While Isolation Forest exhibits moderate effectiveness, it lacks the reliability of Local Outlier Factor. One-Class Support Vector Machine, however, struggles with the lowest recall and precision, particularly in volatile markets, making it the least effective method. These results provide valuable insights for investors and regulators to effectively manage risk and navigate the dynamic landscape of the cryptocurrency market.