A Stacking Ensemble Deep Learning Model for Stock Price Forecasting

Due to the low signal-to-noise ratio and high volatility characteristics inherent in stock time-series, stock price prediction has long been a highly challenging problem. Existing research primarily focuses on constructing complex deep learning models to capture the intrinsic autocorrelation within stocks and to learn intricate nonlinear relationships from stock features. Despite the significant efficacy of deep learning in this domain, two limitations persist. Firstly, individual deep learning models lack diversity, which could hinder their ability to comprehensively capture factors influencing market dynamics. Secondly, deep learning models tend to overfit and are sensitive to noise in stock time-series, exhibiting weak generalization capability. This may engender heightened levels of uncertainty in predictive outcomes. To address these limitations, this paper proposes a Dual Base Learner Decision Neural Network (DBLDNN) framework, comprising two base learners dedicated to learning stock features and assisting the meta learner in final prediction. DBLDNN addresses overfitting and lack of diversity issues by integrating two distinct deep learning models. Each model contributes independently, allowing for balanced predictions and reduced decision errors. Experimental results demonstrate that the proposed method outperforms previous approaches on two public datasets.