Using Multi-Transformer Architecture Hybrid Models to Predict Stock Market Volatility

Abstract Forecasts of stock volatility are crucial for estimating equity risk and, therefore, for the management decisions made by financial institutions. Therefore, this work aims to provide new machine and deep learning techniques-based stock volatility models that are more accurate. This study presents the Multi-Transformer architecture based on neural networks that integrate Transformer and Multi-Transformer (M.T.) layers with the GARCH and the LSTM model and compares their performance with the traditional GARCH-type models. The empirical findings based on the daily returns of NIFTY-50 data suggest that compared to previous autoregressive algorithms, hybrid models based on Multi-Transformer and Transformer layers provide more accuracy and, therefore, are more appropriate for risk assessments. Moreover, even in highly variable conditions like the COVID-19 pandemic, the hybrid Neural-network models beat individual traditional GARCH-type models. The result shows that the bagging mechanism added to the Multi-Transformer architecture helped to reduce the error in the variance in the noisy data and, therefore, reduced the RMSE of hybrid Multi-Transformer models to almost a tenth of the RMSE observed in traditional GARCH-type models.