On replacement of outliers and missing values in time series

Abstract

Presence of missing values and occurrence of outliers in time series cause many hindrances in the analysis of data. Several methods are proposed for determining estimates to replace the missing values and outliers. Mean, median, the largest order statistic and time series model based forecast values are used as the estimates for replacing missing values and outliers. But, no recommendations have been made so far for selection of the estimation methods. This paper attempts to compare the performance of six such estimation methods. Among them, time series models are fitted applying the autoregressive moving average method, long short-term memory method and Facebook’s Prophet method. Models are validated using the test data. Time series of Air Quality Index is used for carrying out for comparative study.

Keywords

• missing value
• outlier
• autoregressive moving average
• facebook’s prophet
• long short-term memory
• mean and median imputation.