Effect of stationarity on traditional machine learning models: Time series analysis

Abstract

Recently, researchers have started the analysis of time series data. In time series data, it is difficult to apply prediction and forecasting techniques effectively. This research work examines how the nature of stationarity of time series data affects the accuracy and forecasting errors. Here, we first categorize the datasets into their stationarity type. Then some state-of- art models are applied to these datasets. Results show that traditional model accuracy and error in the case of forecasting become extremely vulnerable when datasets belong to the non-stationary category. Stationarity tests and experiments are performed on different kinds of benchmark datasets and results are analyzed.