Comparison of Imputation Methods for Univariate Time Series

Abstract

In order to predict and forecast with greater accuracy, handling “missing values” in “time series” information is crucial. Complete and accurate historical data are essential. There are many research studies on multivariate time series imputation, however due to the lack of associated factors, imputation in univariate time series data is rarely taken into consideration. It is natural that “missing values” could arise because almost all scientific disciplines that collect, store, and monitor data use "time series" observations. Therefore, time series characteristics must be considered in order to develop an effective and acceptable method for dealing with missing data. This work uses the statistical package R to assess and measure the effectiveness of imputation methods in the context of "univariate time series" data. The “imputation algorithms” explored are evaluated using “root mean square error”, “mean absolute error” and “mean absolute percent error”. Four types of “time series” are taken into consideration. According to experimental findings, “seasonal decomposition” performs better on the time series having seasonality characteristic, followed by “linear interpolation”, and “kalman smoothing” provides values that are more similar to the original time series data set and have lower error rates than other imputation techniques.