Using Unevenly Spaced Time Series Data Set in a Convolutional Reconstruction Autoencoder Algorithm

Abstract

Time series has been utilized for years, especially in the financial and economic sectors. These data are usually measured over a specific period at evenly spaced intervals, and only a few have analyzed time series on unevenly spaced time series intervals. Such uneven time series are natural phenomena like storms, earthquakes, and tsunamis. The researchers used five randomly generated, unevenly spaced time series datasets with five different data averages as an input to a convolutional reconstruction autoencoder algorithm to determine its reconstruction performance using such type of time series. At the end of the study, the algorithm yielded an average mean absolute error of 0.216 with 0.16 as the lowest mean absolute error and 0.26 as the highest mean absolute error value. Therefore, the researchers can conclude that a convolutional reconstruction autoencoder algorithm can be used with unevenly spaced time series data. However, it is not without drawbacks, as the algorithm could not properly reconstruct the later parts of the dataset. Furthermore, the algorithm’s performance can be improved by extending its training period through more dataset entries.