Forecasting the Dynamic Correlation of Stock Indices Based on Deep Learning Method

With the economic globalization, it is important to accurately estimate and predict the correlation between global stock markets for the efficient asset allocation around the globe. While practitioners may resort to various GARCH-type models for correlation forecasting, the forecasting accuracy is still not very satisfactory. For the first time, this paper introduces a deep learning method into the research of stock market correlation. A hybrid DCDNN model is developed based on the RDNN (recurrent deep neural network) and DCC-GARCH models. Deep learning is then devised to forecast the prediction error of the DCC-GARCH model in order to improve the prediction accuracy of stock market correlation. An autoencoder is also introduced in the empirical study to extract useful features of the stock index data. Then, the correlations among the stock markets in China, Hong Kong, the United States and Europe are predicted and tested. We show that the accuracy of the DCDNN model is significantly higher than that of the DCC-GARCH model. The results indicate that the introduction of deep learning can help improve the efficacy of existing correlation forecasting methods.