In this paper, the autoregressive moving average model (ARMA) and LSTM deep neural network are first introduced, and the time series are decomposed into high volatility components and low volatility components by MA filtering method. Then the time series forecasting model ARMA and deep neural network LSTM are combined on the basis of MA filtering method to form ARMA-LSTM combination model based on MA filtering method, and the application effect of this model in financial market volatility forecasting and risk response is verified through empirical evidence. The results show that the ARMA-LSTM_t model will achieve relatively good results in predicting the GDP_IG of the current year using the data of the 12 months of the current year and the last month of the previous year, and the training and prediction sets of the ARMA-LSTM combination model proposed in this paper have the best results. In addition, there is a positive relationship between investment-related indicators and GDP_IG, and the addition of investment network search data improves the estimation accuracy of the model, obtains smaller prediction errors, and improves the prediction accuracy of the ARMA-LSTM model in the short and medium term.