As an important component of the economic market, the stock market has been concerned by many researchers. How to get the trend of the stock market and predict the stock price is a problem that many researchers are studying. In previous works, the prediction methods are mainly focused on statistical models and traditional neural network models which are relatively popular in recent years. Deep learning is not often used in the field of financial time series, but it has a strong learning ability and is suitable for complex time series such as financial time series. In particular, the LSTM network has the function of long-term memory because of its cyclic structure, so it is very suitable for financial time series prediction in theory. In the study, a novel stock closing price forecasting framework is proposed, which has a higher prediction than traditional models. The data processing part, the deep learning predictor part, and the predictor optimization method are the components of this deep hybrid framework. Data processing includes empirical wavelet transform (EWT) based preprocessing and outlier robust extreme learning machine (ORELM) model based post-processing. Long short-term memory (LSTM) network based deep learning network predictor, as the main part of the mixed frame, is jointly optimized by dropout strategy and particle swarm optimization (PSO) algorithm. Each algorithm in the hybrid framework can give full play to its own functions to achieve better prediction accuracy. In order to verify the performance of the model, three challenging datasets are selected for forecasting experiments. Some comparative models are also selected to prove the effectiveness of the proposed framework. Experimental results show that the hybrid framework proposed in the study has the best prediction accuracy and can be applied to stock market monitoring or financial data analysis and research.

股票市场作为经济市场的重要组成部分，已经引起许多研究者的关注。如何获得股市趋势并预测股价是许多研究者正在研究的问题。在以前的工作中，预测方法主要集中在近年来相对流行的统计模型和传统神经网络模型上。深度学习在金融时间序列领域中并不经常使用，但是它具有强大的学习能力，并且适合于诸如金融时间序列之类的复杂时间序列。特别是LSTM网络由于具有循环结构，因此具有长期记忆的功能，因此从理论上讲非常适合金融时间序列的预测。在研究中，提出了一种新颖的股票收盘价预测框架，比传统模型有更高的预测。数据处理部分，深度学习预测器部分和预测器优化方法是此深度混合框架的组成部分。数据处理包括基于经验小波变换（EWT）的预处理和基于异常鲁棒极限学习机（ORELM）模型的后处理。基于长期短期记忆（LSTM）网络的深度学习网络预测器作为混合框架的主要部分，是通过辍学策略和粒子群优化（PSO）算法共同优化的。混合框架中的每种算法都可以充分发挥其功能，以实现更好的预测精度。为了验证模型的性能，选择了三个具有挑战性的数据集进行预测实验。还选择了一些比较模型来证明所提出框架的有效性。实验结果表明，本文提出的混合框架具有最佳的预测精度，可用于股票市场监测或金融数据分析研究。