A New Deep Network Model for Stock Price Prediction

Abstract

AbstractOver the recent past, many stock price prediction models that rely on deep neural networks have been developed. However, each has unique characteristics that cause variations in performance between models. With the existing deep neural network models, we propose a novel deep neural network-based stock price prediction model in this paper in order to predict stock prices more accurately. Specifically, this paper presents a method for extracting stock price series data based on the auto-encoder (AE) technique, which has strong non-smoothness and non-linear characteristics. Furthermore, the bi-directional long short-term memory (BiLSTM) module is imported as the primary unit structure in AE so that the historical and future important information of stock price series data can be sufficiently mined. Attention mechanisms are also investigated to make the extracted features more valuable for predicting stock prices. Lastly, the prediction is implemented by multi-layer, fully connected network work. The prediction results of the proposed method on two stock datasets are more prominent than other methods.KeywordsStock price predictionAuto-encoderBi-directional LSTMAttention mechanism