Benchmarking Nonstationary Time Series Prediction

Abstract

The prediction of time series has gained increasingly more attention among researchers since it is a crucial aspect of decision-making activities. Unfortunately, most time series prediction methods assume the property of stationarity, i.e., statistical properties do not change over time. In practice, it is the exception and not the rule in most real datasets. Several transformation methods were designed to treat nonstationarity in time series. In this context, nonstationary time series prediction is challenging since it demands knowledge of both data transformation and prediction methods. Since there are no silver bullets, it leads to exploring a large number of data transformation and prediction method combinations for building prediction setups. However, selecting a prediction setup that is appropriate to a particular time series and application is not a simple task. Benchmarking of different candidate combinations helps this selection. This work contributes by providing a review and experimental analysis of transformation methods and a systematic framework (TSPred) for benchmarking and selecting prediction setups for nonstationary time series. Suitable nonstationary time series transformation methods provided improvements of more than 30% in prediction accuracy for half of the evaluated time series. They improved the prediction by more than 95% for 10% of the time series. The features provided by TSPred are also shown to be competitive regarding prediction accuracy. Furthermore, the adoption of a validation phase during model training enables the selection of suitable transformation methods.