A Causality Based Feature Selection Approach for Multivariate Time Series Forecasting

Youssef Hmamouche (1) , Alain Casali (1, 2) , Lotfi Lakhal (1, 2)

1 LIF - Laboratoire d'informatique Fondamentale de Marseille
2 DANA - Data Mining at scale

—The field of time series forecasting has progressed significantly in recent decades, specially in regards to the need of forecasting economic data. That said, some issues still arise. In particular when we are working with a set of time series that have a large number of variables. Hence, a selection step is usually needed in order to reduce the number of variables that will contribute to forecast each target time series. In this paper, we propose a feature selection and / or dimension reduction algorithm for forecasting multivariate time series, based on (i) the notion of the Granger causality, and (ii) on a selection step based on a clustering strategy. Finally, we carry out experiments on different real data sets, by comparing our proposal and some of the most used feature selection methods. Experiments show that we improved the forecasting accuracy compared with the evaluated methods.

Mots clés: Granger Causality; –Multivariate Time Series Forecasting; Feature selection; Multivariate Time Series Forecasting; Forecasting using Feature Selection

Domaines: Base de données [cs.DB]; Algorithme et structure de données [cs.DS]; Machine Learning [stat.ML]

Youssef Hmamouche, Alain Casali, Lotfi Lakhal. A Causality Based Feature Selection Approach for Multivariate Time Series Forecasting. DBKDA 2017, The Ninth International Conference on Advances in Databases, Knowledge, and Data Applications, May 2017, Barcelone, Spain. ⟨hal-01467523⟩