Stock price prediction using multi-scale nonlinear ensemble of deep learning and evolutionary weighted support vector regression

Abstract Stock price prediction has become a focal topic for relevant investors and scholars in these years. However, owning to the non-stationarity and complexity of stock price data, it is challenging to predict stock price accurately. This research develops a novel multi-scale nonlinear ensemble learning framework for stock price prediction, which consists of variational mode decomposition (VMD), evolutionary weighted support vector regression (EWSVR) and long short-term memory network (LSTM). The VMD is utilized to extract the basic features from an original stock price signal and eliminate the disturbance of illusive components. The EWSVR is utilized to predict each sub-signal with corresponding features, whose penalty weights are determined according to the time order and whose parameters are optimized by tree-structured Parzen estimator (TPE). The LSTM-based nonlinear ensemble learning paradigm is employed to integrate the predicted value of each sub-signal into the final prediction result of stock price. Four real prediction cases are utilized to test the proposed model. The proposed model’s prediction results of multiple evaluation metrics are significantly improved compared to other benchmark models both in stock market closing price forecasting.