Attention LSTM : Application in Stock Price Prediction on Single Companies

Abstract

According to the development of data-related techniques, aimed at exploring the largest value of data, price prediction has been seen as more vital for quantifying and pricing stock. To solve this problem, the learning based algorithm became popular during modern computing techniques development. LSTM (Long Short-Term Memory) techniques attract most within the new techniques for their interesting architecture and excellent performance. However, this leads to a new idea that maybe it can achieve better performance and also avoid recurrent message-passing steps with Attention LSTM, a multi-head Attention model. The main topic of this paper is the performance of Attention LSTM and LSTM models in the stock price prediction of single companies. To evaluate the performance of Attention LSTM in stock price prediction, the researcher designed several tests with different input lengths and two ways of splitting data: sliding window and feed-forward input. Based on the comparison, the result could be seen clearly that Attention LSTM could have better performance than LSTM with suitable test procedure 'sliding window'. LSTM could performed better with 'feed-forward input' approach. However, Attention LSTM could still run faster in this approach so it could help in tasks with large sequence length required high speed.