Depth feature extraction-based deep ensemble learning framework for high frequency futures price forecasting

Abstract

Whether the change trend of futures price can be accurately analyzed and predicted is the key to the success or failure of futures trading. This paper constructs a new deep ensemble learning framework combining signal decomposition and exogenous variable feature mining for high-frequency futures price prediction, which consists of depth feature extraction (DFE), long short-term memory optimized by attention mechanism (ALSTM) and Light gradient boosting machine (LightGBM). In the depth feature extraction stage, based on multi-scale entropy (MSE) and Savitzky-Golay filter (SG filter), an improved denoising variational mode decomposition (VMD) is proposed to extract the fluctuation characteristics of futures price signal and eliminate the interference of complex components. To avoid the collinearity redundancy of high-dimensional exogenous variables, an enhanced dimensionality reduction method combining Spearman correlation analysis and stacked autoencoder (SAE) is designed to ensure the simplicity and correlation of input factors. In the prediction phase, ALSTM is adopted as a base predictor for constructing point prediction model by the DFE results, which can focus on learning more important data features. Finally, LightGBM, which has excellent effect in the field of ensemble learning, is used to integrate the base prediction results to obtain the final results. The actual closing price data of three representative futures varieties in China's futures market are selected to verify the accuracy of the proposed framework. Compared with other benchmark models, this developed framework has better futures closing price prediction performance.