For financial volatility and market risk, starting the empirical analysis by neural network model based on four-factor data of more than 70 stocks in smart industry sector in the A-share market for the past ten years. In this model, stochastic gradient algorithm for closing price predictionwas used, then the model error and loss function of predicted and actual values was compared , and optimizing from factor selection, algorithm improvement and indicator merit. The results show that closing price is best fitted when neural network model at batch = 2 and epoch = 4150, MSE, MAPE and MAE are 60.1911, 30.7326, 4.8032 respectively. The neural network model with these parameters can be used to explore the stock market price trend and provide some reference basis for investors or financial institutions.