基于ARIMA-CNN-LSTM的股票价格预测

Stock Price Prediction Based on ARIMA-CNN-LSTM

摘要: 股票市场的波动会影响人们生活的各个方面，因此准确预测股票价格具有重要意义。然而，传统的时间序列预测模型(如ARIMA)在处理股票价格中的非线性特征时表现不佳，难以获得令人满意的预测效果。鉴于深度学习在处理非线性问题上的优越能力，它在股票价格预测中展现出巨大的潜力。本文提出了一种ARIMA-CNN-LSTM混合模型来预测股票价格，该模型充分挖掘了股票市场的历史信息，从而实现更高精度的预测。首先，使用ARIMA对股票数据进行预处理，将处理得到的残差序列输入CNN，通过卷积提取股票数据的深层特征；随后，通过LSTM挖掘股票数据的长期时间序列特征。本文以中国银行2009年1月1日至2023年12月31日的股票数据为研究对象，构建混合模型，并与ARIMA-LSTM模型、CNN-LSTM模型的预测结果进行对比。实验结果表明，本文提出的混合模型预测精度更高，相较于ARIMA-LSTM与CNN-LSTM模型，本模型在均方误差(MSE)指标上分别降低了25%与16.7%，在平均绝对误差(MAE)指标上分别降低了14.1%与9.4%，展示了更好的预测效果。

Abstract: The volatility of the stock market affects various aspects of people’s lives, making accurate stock price predictions highly significant. However, traditional time series forecasting models, such as ARIMA, perform poorly when dealing with the nonlinear characteristics of stock prices, leading to unsatisfactory prediction outcomes. Given the superior capability of deep learning in handling nonlinear problems, it shows great potential in stock price prediction. This paper proposes a hybrid ARIMA-CNN-LSTM model for stock price prediction, which fully exploits historical market information to achieve higher prediction accuracy. First, ARIMA is used to preprocess the stock data, with the resulting residual series fed into a CNN, which extracts deep features of the stock data through convolution. Subsequently, LSTM is employed to capture the long-term temporal characteristics of the stock data. Using the stock data of the Bank of China from January 1, 2009, to December 31, 2023, as the research object, the hybrid model is constructed and compared with the prediction results of the ARIMA-LSTM model and CNN-LSTM model. The experimental results show that the proposed hybrid model achieves higher prediction accuracy. Compared to the ARIMA-LSTM and CNN-LSTM models, this model reduces the mean squared error (MSE) by 25% and 16.7%, respectively, and decreases the mean absolute error (MAE) by 14.1% and 9.4%, respectively, demonstrating better forecasting performance.