Outliers detection in multivariate time series by independent component analysis

Area of Science:

• Statistics
• Data Science
• Time Series Analysis

Background:

• Multivariate time series data frequently contain outliers that deviate from the common pattern.
• Undetected outliers can significantly distort time series analyses, including model building, seasonality assessment, and forecasting.
• The inherent structure of multivariate time series can cause smearing and masking phenomena, complicating outlier identification.

Purpose of the Study:

• To investigate the utility of Independent Component Analysis (ICA) for detecting outliers in multivariate time series.
• To model outliers as non-Gaussian components superimposed on Gaussian multivariate time series.
• To compare the effectiveness of different ICA algorithms in identifying various outlier types.

Main Methods:

• Modeling outliers as non-Gaussian components within a Gaussian multivariate time series framework.
• Applying Independent Component Analysis (ICA) to separate observable time series into regular and outlying unobservable components.
• Utilizing factor models to further support ICA's role in outlier detection.
• Comparing the performance of multiple ICA algorithms.

Main Results:

• Independent Component Analysis (ICA) proves effective in analyzing multivariate observable time series and separating regular from outlying components.
• All tested ICA algorithms demonstrated effectiveness in detecting diverse outlier types, including patches, level shifts, and isolated outliers.
• Outlier detection was successful even when outliers occurred at the beginning or end of the observation period.
• No significant differences were observed in the ability of various ICA algorithms to reveal outlier patterns.

Conclusions:

• Independent Component Analysis (ICA) is a valuable tool for the identification and analysis of outliers in multivariate time series.
• ICA successfully addresses challenges posed by smearing and masking phenomena common in multivariate time series outlier detection.
• The findings support the use of ICA for enhancing the accuracy of time series modeling, seasonality assessment, and forecasting by accounting for outliers.