Comparing recurrent neural network with GARCH model on forecasting volatility based on SSE 50ETF

This paper compares the performance of DL models (RNN, LSTM and GRU) and GARCH model in three different train set, which with different time span. The data of empirical analysis is SSE 50ETF from 4th May 2010 to 26th Aug 2022. And the performance is compared with realized volatility. The result shows that SSE 50ETF is more relying on long historical information and pay less attention to new information. And in long periods, the DL models has better performance. However, the stability of DL models are significantly worse than GARCH model. The performance of DL models are highly relying on the selection of a train set.