Time-Series Outliers Detection Algorithm with Clustering Approach on Non-Linear Trends

Abstract

It has been found that the existence of outliers, particularly in time-series data, can be significantly influenced the modelling and analysis results that are conducted on the data, which is further may lead to improper decision making. Nevertheless, the task of time-series outlier detection can be quite challenging when dealing with collection of data that retain non-linear trends over time as the progression of series may shifted and would be infer as possible outliers. In this study, an algorithm for time-series outlier detection that makes use of a clustering approach on time-series data to construct a set of localized trend models that is capable to identify anomalous data in a collection of non-linear trends is proposed. Decisively, results from conducted experiments confirm that the procedure performs prompt, incremental valuation of information as soon as it becomes accessible, able to handle significant amount of data, and does not need any pre-classification of anomalies. Furthermore, trials with real-world data from insurance field confirm that the proposed method is able to correctly identify abnormal data and can be of help to increase decision making process.