Research On Stock Index Prediction Based on CGAF Hybrid Deep Learning Model

Abstract

This article focuses on the prediction difficulties caused by the inherent high-frequency noise, strong nonlinearity, and complex dynamic characteristics of stock index time series, and proposes an innovative hybrid deep learning model - CGAF (CNN-GRU Attention FFN). The CGAF model integrates one-dimensional convolutional neural networks to extract multi-scale local features, gated recurrent units to capture temporal dependencies, multi head self attention mechanism to dynamically focus on key time step information, and feedforward network to deepen nonlinear feature transformation. Using the closing prices of the Shanghai Stock Exchange Composite Index from March 3, 2023 to March 31, 2025 as empirical data, the model was trained and evaluated for performance after linear interpolation, minimum maximum normalization, and sliding window processing. The research results indicate that the CGAF model is significantly superior to traditional time series models such as ARIMA, ARCH, GARCH, multiple benchmark deep learning architectures such as CNN, CNN-LSTM, CNN-LSTM Attention, CNN-GRU-MAHA, and models referenced in the literature in key evaluation indicators such as RMSE, MAE, and MAPE. The original scale RMSE, MAE, and MAPE of the CGAF model on the test set were 24.7868, 18.1145, and 0.54%, respectively, demonstrating excellent predictive performance and goodness of fit. This study not only provides an effective tool for financial market forecasting, but also verifies the progressiveness and potential of the proposed hybrid architecture in processing complex time series data, and the modular design also has good portability.