A novel secondary decomposition method for forecasting crude oil price with twitter sentiment

Abstract

With the ubiquity of the Internet, valuable social media data have been generated, and a promising idea for using sentiment from social media has emerged in oil price forecasting. To address the complexity of oil price, this study proposed a novel secondary decomposition framework with Twitter sentiment, in which Twitter sentiment provides new information and secondary decomposition technique reduces the difficulty of oil price forecasting. This methodology involves three major steps: (1) sentiment extraction, to collect and extract Twitter sentiment via the state-of-the-art sentiment analysis technique—Bidirectional Encoder Representations from Transformers (BERT); (2) secondary decomposition, to extract scale-aligned components from crude oil price and Twitter sentiment using bivariate empirical mode decomposition (BEMD) first and then decomposing the residual terms through GA-VMD; and (3) oil price prediction, including individual prediction at each intrinsic mode function (IMF) and ensemble prediction across different IMFs. With WTI oil price as a sample, the empirical study results indicate that the proposed novel learning paradigms statistically outperform their corresponding original techniques (without Twitter sentiment and secondary decomposition), semi-improved variants (with either Twitter sentiment or secondary decomposition), and similar counterparts (with one-time decomposition analysis) in terms of prediction accuracy.