Home // DBKDA 2017, The Ninth International Conference on Advances in Databases, Knowledge, and Data Applications // View article

A Causality-based Feature Selection Approach for Multivariate Time Series Forecasting

Authors:
Youssef Hmamouche
Alain Casali
Lotfi Lakhal

Keywords: Time Series Forecasting; Granger Causality; Feature selection.

Abstract:
The study of time series forecasting has progressed significantly in recent decades. The progress is partially driven by growing demand from different industry branches. Despite recent advancements, there still exist several issues that need to be addressed in order to improve the accuracy of the forecasts. One of them is how to improve forecasts by utilizing potentially extra information carried by other observed time series. This is a known problem, where we have to deal with high dimensional data and we do not necessarily know the relationship between variables. To deal with this situation, the challenge is to extract the most relevant predictors that will contribute to forecast each target time series. In this paper, we propose a feature selection algorithm specific to forecasting multivariate time series, based on (i) the notion of the Granger causality, and on (ii) a clustering strategy. Lastly, we carry out experiments on several real data sets and compare our proposed method to some of the most widely used dimension reduction and feature selection methods. Experiments illustrate that our method results in improved accuracy of forecasts compared to the evaluated methods.

Pages: 97 to 102

Copyright: Copyright (c) IARIA, 2017

Publication date: May 21, 2017

Published in: conference

ISSN: 2308-4332

ISBN: 978-1-61208-558-6

Location: Barcelona, Spain

Dates: from May 21, 2017 to May 25, 2017