Abstract: The prediction of financial time series has been a very challenging and meaningful work. An effective prediction model should reflect the complex features such as nonlinearity, non-stationary and sequential correlation that exists in financial time series and conclude the dynamic non-linear interaction effects among the financial economic variables. So, the Long-Short Term Memory (LSTM) deep neural network is used to predict financial time series data. In order to improve the generalization ability of the LSTM model, wavelet analysis is adopted to preprocess the financial time series data to eliminate the noise components of high frequency. That is, wavelet analysis and LSTM deep neural network are combined to forecast financial time series data. At the same, taking the daily closing price of Dow Jones Industrial Average as an example, the prediction ability of LSTM neural network for actual financial data is explored. And this result is compared with the prediction results of Multilayer Perceptron, Support Vector Machine, K-nearest Neighbors and GARCH model. Results show that LSTM neural network can balance the prediction effect of the training set, validation set and test set. And LSTM shows a better prediction effect than shallow machine learning models and GARCH model and better generalization ability. Also, the wavelet decomposition and reconstruction of the financial time series data can effectively improve the generalization ability of the LSTM neural network and can better predict the long-term dynamic trend of the financial time series data. It proves the applicability and effectiveness of LSTM neural network in the area of financial time series prediction, and it is of great significance to monitor the risk of securities market and provide investors with investment suggestions.

Key words: long-short term memory neural network, wavelets, deep learning, financial time series prediction