GFSM: a Feature Selection Method for Improving Time Series Forecasting

Youssef Hmamouche (1) , Piotr Przymus , Alain Casali (2, 3) , Lotfi Lakhal (4, 2)

1 LIF - Laboratoire d'informatique Fondamentale de Marseille
2 DANA - Data Mining at scale
3 LIS - Laboratoire d'Informatique et des Systèmes (LIS) (Marseille, Toulon)
4 LIM - Laboratoire d'Informatique de Marseille

Handling time series forecasting with many predictors is a popular topic in the era of "Big data", where wast amounts of observed variables are stored and used in analytic processes. Classical prediction models face some limitations when applied to large-scale data. Using all the existing predictors increases the computational time and does not necessarily improve the forecast accuracy. The challenge is to extract the most relevant predictors contributing to the forecast of each target time series. We propose a causal-feature selection algorithm specific to multiple time series forecasting based on a clustering approach. Experiments are conducted on US and Australia macroeconomic datasets using different prediction models. We compare our method to some widely used dimension reduction and feature selection methods including principal component analysis PCA, Kernel PCA and factor analysis. The proposed algorithm improves the forecast accuracy compared to the evaluated methods on the tested datasets.

Mots clés: Time Series Forecasting; Feature Selection; Multi- variate prediction models; Artificial Neural Networks

Domaines: Algorithme et structure de données [cs.DS]; Calcul parallèle, distribué et partagé [cs.DC]; Base de données [cs.DB]; Recherche d'information [cs.IR]

Youssef Hmamouche, Piotr Przymus, Alain Casali, Lotfi Lakhal. GFSM: a Feature Selection Method for Improving Time Series Forecasting. International Journal On Advances in Systems and Measurements, 2017. ⟨hal-02448277⟩