Abstract: Aiming at the prediction challenges caused by the inherent nonlinearity and low signal-to-noise ratio of financial volatility series, we proposed an adaptive gated hybrid model  SCLA-GARCH that fused structural priors with deep learning. The model utilized the generalized autoregressive conditional heteroskedasticity (GARCH)  to capture linear structures, and constructed a parallel CNN-LSTM-Attention network to learn nonlinear patterns in its residuals, which were dynamically fused by an adaptive gating mechanism. Empirical analysis results show  that the prediction error of this model is significantly lower than that of multiple benchmark models, with a maximum reduction of 59% in mean squared error.

Key words: volatility modeling, GARCH model, neural network, financial volatility, forecast