Abstract: As a core component of the capital market, the price trends and future trend prediction in the stock market are among the most important concerns for investors. Accurate prediction of stock prices can provide strong technical support for investors. For the past few years, deep learning has been widely used for stock prediction to explore more effective information. However, the nonlinearity, complexity, and multiscale characteristics of stock data make it relatively difficult to extract hidden information, and deep learning prediction models suffer from issues such as vanishing gradients and time delays, making it challenging to fit when the sequence exhibits significant fluctuations. Thus, there is an urgent need to improve prediction accuracy.
Long Short-Term Memory (LSTM) networks have been widely applied to stock prediction in recent years, but their structural characteristics can easily fall into local optima, thereby affecting prediction accuracy. Drawing on the good approximation ability of the Broad Learning System (BLS) in time series forecasting, this study attempts to combine width learning with deep learning. It constructs an LSTM-BLS model by utilizing LSTM for feature extraction, feeding the extracted features to the mapping nodes of BLS through fully connected layers, generating enhanced nodes, and calculating the prediction values. Additionally, to address the non-stationary nature of stock sequences, Complementary Ensemble Empirical Mode Decomposition (CEEMD) is introduced for denoising. The adaptive decomposition characteristic of CEEMD does not excessively increase the complexity of the model. Therefore, the proposed CEEMD-LSTM-BLS stock prediction model is presented. The model is implemented using the Keras framework in the Python language, and empirical research is conducted using closing price data of the Agriculture, Forestry, Animal Husbandry, and Fishery Index (399231) from the China Stock Market & Accounting Research Database (CSMAR).
Three evaluation metrics, namely mean absolute error, root mean square error, and coefficient of determination, are selected to assess the performance of the model. By comparing the LSTM-BLS model with baseline models and existing stock prediction models, it is demonstrated that the fusion of Deep Learning and Broad Learning System shows significant improvement in multiple accuracy indicators. Particularly, when comparing the CEEMD-LSTM-BLS model with the CEEMD-LSTM model without BLS integration, it is found that the LSTM model exhibits certain errors in prediction, especially at turning points where the volatility is higher, leading to more pronounced prediction errors. The BLS module in the CEEMD-LSTM-BLS model can address such issues. When the data exhibits significant fluctuations, the proposed new model in this study outperforms existing models in terms of fitting discrepancies and time delays. Therefore, the proposed CEEMD-LSTM-BLS stock closing price prediction model can accurately forecast the market’s ups and downs, providing valuable reference for investors.
However, this study only considers the single factor of closing price, which has certain limitations. Therefore, the future experimental direction will be based on this study and consider multiple-factor input prediction to prevent information compression issues caused by single-factor input for highly volatile data, which can hinder the achievement of the desired prediction accuracy. The aim is to construct a stock prediction model that integrates deep learning and width learning systems under the influence of multiple variables and factors. Finally, the authors express their gratitude to the mentors for their guidance and support throughout this research.

Key words: stock forecasting, Complementary Ensemble Empirical Mode Decomposition, Long Short-Term Memory, Broad Learning System

摘要： 长短时记忆网络(LSTM)近年来广泛应用于股票预测中,其结构特点易陷入局部最优,从而影响预测精度。借鉴宽度学习系统(BLS)在时间序列预测上良好的逼近能力,本文尝试宽度学习与深度学习相结合。进一步地,针对股票序列不平稳特点,引入互补集成经验模态分解(CEEMD)进行降噪处理,提出CEEMD-LSTM-BLS(C-L-B)股票预测模型。选取农林牧渔行业股票价格,对新提出的模型进行实证研究。通过与基线模型、现有股票预测模型对比,证明了新模型在多个精度指标上都有明显提升。特别地,通过分别将C-L-B模型与不融入BLS的CEEMD-LSTM模型,对CEEMD分解后的分量预测结果进行对比发现:LSTM模型预测存在一定的误差,且越是拐点处,越是高频波动,预测误差越明显。而C-L-B模型中的BLS模块能够解决这类问题。当数据出现较大波动时,本文提出的新模型与现有模型相比,可以很好的解决拟合差、时滞等问题。

关键词: 股票预测, 互补集成经验模态分解, 长短时记忆网络, 宽度学习系统