Financial Forecasting Method for Generative Adversarial Networks based on Multi-model Fusion

Abstract

<p> To some extent, stock prices can reflect national economic development and residents’ living standards. However, current stock forecasts are mainly analyzed using the stock’s own price. It is found that the exchange rate and technical indicators are closely related to the fluctuations of stocks and stock indexes. In this study, a generative adversarial network (GAN) financial forecasting method based on multi-model fusion is proposed, which introduces the exchange rate and technical indexes of stock and stock index into the input data, and combines the characteristics of its own highest and lowest price, the closing price of stock or stock index is predicted by using the generated antagonistic network. The methods are as follows: firstly, the exchange rate features are de-noised by wavelet transform technology, the overall trend of exchange rate features is extracted, and the dimensionality of technical indicators is reduced by Principal Component Analysis (PCA). Then, a convolutional neural network (CNN) is used in the generator to extract the local features of the input data, and the attention mechanism is used to improve the prediction effect of the model. Finally, the prediction was made by long short-term memory network (LSTM). Convolutional neural network and multilayer fully connected neural network are used in the discriminant, and gelu is used as activation function in the hidden layer. This study selects the data of Zhuhai Port (000507.SZ), Xiamen International Trade (600755.SH), Orient Venture (600278.SH) and Shanghai Stock Exchange Index (SSE, 000001.SH) in the export industry from January 1, 2013 to December 31, 2019 as the experimental data set. The exchange rate price between RMB and US dollar at the corresponding time is selected as the exchange rate feature. Compared with the best in-depth learning financial prediction comparison model, the results show that the generation confrontation network proposed in this paper has improved by 3% to 15% in each experimental data set.</p> <p> </p>