Hybrid Deep Learning for Financial Volatility Forecasting: An LSTM-CNN-Transformer Model

Abstract

 This paper proposes a deep learning model for stock market volatility prediction. The model integrates LSTM, CNN, and Transformer structures. It is designed to enhance the ability to model complex dynamic features in financial time series. CNN is used to extract local price movement patterns. LSTM captures long-term temporal dependencies. The Transformer's self-attention mechanism is introduced to improve global feature learning. This enables multi-level and multi-scale information fusion. The experiments are conducted on S&P 500 Index. The model performance is evaluated using three metrics: Mean Squared Error, Mean Absolute Error, and R². The results demonstrate that the proposed model surpasses conventional methods such as Support Vector Machines (SVM) and Random Forests, as well as mainstream deep learning models like Gated Recurrent Units (GRUs) and Multi-Layer Perceptrons (MLPs). It achieves higher prediction accuracy and stability. In addition, the study compares the effects of different optimizers on training performance. The results further confirm the effectiveness of the AdamW optimizer in improving model convergence and fitting ability. This research demonstrates the potential of multi-structure fusion models in financial time series modeling. It provides a new approach for fine-grained stock market volatility forecasting.