|Title||A First Approximation to the Effects of Classical Time Series Preprocessing Methods on LSTM Accuracy|
|Publication Type||Conference Paper|
|Year of Publication||2019|
|Authors||Viedma, Daniel Trujillo, Rivera-Rivas A.J., Charte Francisco, and del Jesus M. J.|
|Conference Name||International Work-Conference on Artificial Neural Networks|
|Keywords||LSTM, Preprocessing, time series|
A convenient data preprocessing has proven to be crucial in order to achieve high levels of accuracy, time series being no exception. For this kind of forecasting tasks, several specialized preprocessing methods have been described, being trend analysis and seasonal analysis some of them. Several have been formally grouped around a methodology that is always applied to state of the art time series forecasting models, like the well known ARIMA.
LSTM is a relatively novel architecture which has been specifically designed to get rid of the vanishing gradient problem. In these models, great results have been seen when applied for time series forecasting. Still, little is known about the impact of these traditional preprocessing methods on the accuracy of LSTM.
In this work an empirical analysis on how classical time series preprocessing methods influence LSTM results is conducted. That all considered ones can potentially improve LSTM performance is concluded, being the seasonal component removal the filter that achieves better, most robust accuracy gain.