Downloads: 0 | Views: 35
Informative Article | Science and Technology | India | Volume 9 Issue 11, November 2020
Anomaly Detection Techniques in Time Series Forecasting: Identifying Outliers
Abstract: Time series forecasting, a linchpin in data science, faces challenges due to the inherent fluctuations, trends, and patterns in time series data, making it susceptible to anomalies. Detecting outliers becomes crucial to preserve the accuracy of forecasting models, especially in domains like finance, supply chain, healthcare, manufacturing, and cybersecurity. This paper explores the significance of anomaly detection in time series forecasting, emphasizing both statistical and machine learning - based approaches. Statistical methods like the Z - Score, Grubbs' Test, and Modified Z - Score provide foundational techniques, while machine learning algorithms like Isolation Forests, One - Class SVM, and Autoencoders offer advanced anomaly identification. Additionally, time series - specific methods such as Seasonal Hybrid ESD, the Prophet Algorithm, and Dynamic Time Warping address challenges unique to time series data. These techniques cater to seasonal patterns, varying speeds, and the dynamic nature of evolving patterns. The paper concludes by addressing challenges in anomaly detection, emphasizing real - time detection, balancing false positives and negatives, and managing imbalanced datasets. As technology advances, integrating sophisticated anomaly detection remains critical for resilient and effective time series forecasting in the evolving landscape of data science.
Keywords: anomaly detection, time - series forecasting, z - score, isolation forests, one - class SVM, dynamic time warping, real - time detection
Edition: Volume 9 Issue 11, November 2020,
Pages: 1707 - 1709