As an emerging trading market, the crude oil futures market has exhibited substantial uncertainty since its inception. Influenced by macroeconomic and geopolitical factors, its price movements are highly nonlinear and nonstationary, making accurate forecasting challenging. Therefore, it is vital to develop a powerful forecasting model for crude oil futures prices. However, conventional forecasting models rely solely on historical data and fail to capture the intrinsic patterns of complex sequences. This work presents a hybrid deep learning framework that incorporates multi-source features and a state-of-the-art attention mechanism. Specifically, search engine data were collected and integrated into the explanatory variables. By using lagged historical prices and search engine data to forecast future crude oil futures closing prices, the proposed framework effectively avoids lookahead bias. To reduce forecasting difficulty, the initial time series were then decomposed and reconstructed into several sub-sequences. Thereafter, traditional time series models (ARIMA) and attention-enhanced deep learning models were selected to forecast the reconstructed sub-sequences based on their distinct data features. The empirical study conducted on the INE crude oil futures price proves that the proposed model outperforms other benchmark models. The findings help fill the gap in the quantitative literature on crude oil futures price forecasting and offer valuable theoretical insights for affiliated policymakers, enterprises, and investors.
Liu et al. (Sun,) studied this question.