Abstract: With the diversity of applications scenarios and rapid growth of data,the stock prediction models based on classical statistical methods are unable to meet the requirements for high prediction accuracy.But traditional stock reco-mmendation models either never consider the time factor or just consider the unidirectional relationship over time.However,existing stock recommendation models based on deep learning rarely consider the time factor.This paper proposed a deep bidirectional LSTM model for stock prediction,which makes full use of context relationship in the forward direction and backward direction of time series.The problem of vanishing gradient and exploding gradient are solved by introducing LSTM when dealing with long-term sequence.The proposed model can learn information which has long-term dependence on time.At the same time,dropout strategy is introduced to prevent over-fitting caused by deep network model and speed up the training.Experiments on S&P500 dataset show that the neural network prediction model based on the deep bidirectional LSTM outperforms the existing prediction models,the error is about 5% lower,and the coefficient of determination (r²) is increased by 10%.

Key words: Bidirectional long short-term memory, Deep recurrent neural networks, Recommendation system, Stock forecast