A decomposition ensemble based deep learning approach for crude oil price forecasting

As the price of crude oil has nonlinearity, instability, and randomness, capturing its behavior precisely is significantly challenging and leads to difficulties in forecasting. This study combines a decomposition-ensemble approach, optimized by the seagull algorithm, with a sentiment analysis to handle the problem. First, the cumulative sentiment score sequence is obtained by a sentiment analysis of news headlines data. Second, an adaptive signal decomposition method, namely, ensemble empirical mode decomposition (EEMD), is employed to decompose the historical crude oil future prices data into several intrinsic mode functions and one residual component to reduce the impact of noise. Third, a seagull optimization algorithm (SOA) is introduced to tune the hyperparameters of gated recurrent units (GRUs). The optimized GRU model is established to acquire the predicting values of each component integrated with the cumulative sentiment score sequence. Subsequently, multiple linear regression (MLR) is then introduced as the ensemble approach that integrates the forecasting results of each component. The empirical forecasting results of daily West Texas Intermediate (WTI) crude oil future and news headlines data validate our proposed decomposition-ensemble approach with different forecasting horizons. This approach significantly outperforms some other comparison models by means of forecasting accuracy and hypothesis tests. In addition, the EEMD components of WTI crude oil future price are reconstructed to analyze the impact of black-swan events on crude oil prices fluctuations from January 4, 2010 to September 17, 2019.