Stock price prediction is one among the complex machine learning problems due to its chaotic nature which makes it difficult to predict. Predicting a stock price is an ability to uncover hidden patterns in data and focusing on this to make prediction. Some analyst argue that history repeat itself and that historical patterns do repeat and finding this long dependencies patterns can be sometimes difficult.
Long short-term memory (LSTM), a tool in the deep learning framework and with its design nature to capture long term dependencies, is more appropriate for problems involving time series such as speech recognition, language translation and stock prediction. In this paper, we propose a novel model incorporating a sequence-to-sequence model that consists two LSTMs, one encoder and one decoder.
The encoder LSTM accepts input time series of fixed lengths, extracts information from the raw data and based on which the decoder LSTM constructs fixed length sequences that can be regarded as discriminatory features. For better utilization of the raw data (Some being redundant), we also introduce the attention mechanism into our model so that the output prediction generation process can peek at the input data, learn the correlations among the data and focus on the part of the input data that is most relevant to the current prediction at a time step t. We compared our results with traditional LSTM and evaluated their performance using various measures. We observed our model was superior to the traditional LSTM which was supported by the error metric used for the evaluation.

http://www.riss.kr/search/detail/DetailView.do?p_mat_type=be54d9b8bc7cdb09&control_no=14525fce0fb98995ffe0bdc3ef48d419