A combined architecture of multivariate LSTM with Mahalanobis and Z-Score transformations for oil price forecasting

Abstract

Oil price plays a vital role in a country's economy. Oil price forecasting helps in making better economic planning and decisions. The fluctuation in the oil price occurs due to several factors that make oil price forecasting a challenging task. Deep learning models such as Recurrent Neural Networks (RNN) have emerged as a successful method in solving many nonlinear and complex problems. In this research, a combined architecture of Multivariate Long Short Term Memory (MLSTM) is proposed with Mahalanobis and Z-score transformations. These transformations improve the data to uncorrelated and standardized variance, thus making data more suitable for regression analysis. The available historical time-series data on the West Texas Intermediate (WTI) oil prices and the factors affecting the oil prices are collected to form the data set. The feature selection process using selectKBest and correlation analysis is carried out. We develop six variants of MLSTM models with different features, transformations, and outlier elimination. The MLSTM model with Z-score transformation on the selected features and outlier eliminated data gives an RMSE of 0.212 and an R2 score of 0.954. This indicates that our model performed well.