Time Series Analysis for Big Data: Evaluating Bayesian Structural Time Series Using Electricity Prices

In the era of Big Data, time series analysis faces new challenges. The incredible amount of data, which becomes available every day, is accompanied by a similarly huge amount of external variables that could be used to explain or predict data. To cope with those possible covariates, a robust selection of relevant variables is necessary to efficiently predict or explain data. However, a careful distinction between predictive and explanatory analysis is crucial as the two tasks can be tackled differently well by the same method. In this paper, we evaluate new research methods for the analysis of Big Data time series. More specifically, we analyze the performance of Bayesian Structural Time Series and classical machine learning when performing explanatory or predictive tasks. We find that the Bayesian structural time series approach is well suited for the explanatory variable selection, whereas a support vector regression performs better in the prediction task.