Machine Learning in Time Series Modeling

Detecting anomalies in time series data is an important task with applications in fields such as finance, industry, and IT monitoring. In this study, an analysis of existing machine learning models for anomaly detection was conducted using five benchmarks and synthetically generated data. Models such as Isolation Forest, Local Outlier Factor (LOF),one-class support vector machine (OCSVM), clustering methods, and LSTM autoencoders were examined. Additionally, new hybrid models were proposed, combining OCSVM with architectures of LSTM, transformers, and temporal convolutional networks, as well as isolation forest with temporal convolutional networks. This work aims to improve anomaly detection by combining the advantages of nonlinear transformations and anomaly detection algorithms. The results show that the obtained hybrid models are effective for certain tasks and have potential for further research.