A Study on Stock Risk Prediction Methods Based on Deep Learning

Abstract

With the increasing complexity of stock markets and the nonlinear nature of stock price fluctuations, traditional financial forecasting methods often fail to achieve satisfactory results. This study proposes a hybrid neural network model that integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks to enhance the accuracy of stock closing price prediction. The model leverages CNN to extract spatial features from historical financial indicators such as opening price, highest price, and trading volume, and then uses LSTM to capture temporal dependencies within the time series data. Experimental validation is conducted using a dataset of the CSI 300 Index from 1992 to 2021, demonstrating the proposed model’s superior performance in comparison with CNN-only, LSTM- only, and CNN+RNN configurations. Evaluation metrics including Mean Relative Error (MRE) and Mean Absolute Error (MAE) indicate that the CNN-LSTM hy- brid network significantly improves prediction precision. The results highlight the potential of deep learning in modeling complex financial dynamics and offer insights into data-driven approaches for stock risk forecasting.