Robust Stock Opening Price Forecasting Based on LSTM-Stack Ensembles

Stock data is renowned for its complexity and nonlinearity. In the task of stock forecasting, when only relying on a single model, it often leads to unstable forecasting results.This instability arises from factors such as initialization variability and randomness during the training process.In addition, due to the specificity of the training dataset, singular models may fall into local optima. To address these challenges, a LSTM-Stack model has been introduced to this research.This is an ensemble consisting of multiple Long Short Term Memory (LSTM) networks specifically designed for predicting stock opening prices. This research used the LSTM-Stack model’s predictions as new features to train a linear regression model. Compared to single LSTM, the predictive performance of the model using stacked LSTM has been greatly improved. When conducting comparative experiments using the same dataset from CATL, the $\mathbf{R}^{\mathbf{2}}$ score improved by $6.32 \%$. This provides a more accurate solution for stock forecasting and other similar tasks. The promising results presented in this work offer fundamental grounds for stock price forecasting applications to be used for predicting the opening prices and providing advices for the investors.