Predictive Analytics System for Time Series Stock Data Using LSTM with Residual Unit and Attention Mechanism

Abstract

Stock price prediction is a challenging research topic because of non-linearity, significant noise and volatility of time series data. Deep learning techniques enable to learn complex and non-linear patterns of sequential time series data. Long Short-Term Memory (LSTM) is a technique which is designed to handle time series data. While LSTM model is used to extract temporal dependencies of stock data, the performance can be limited by noisy data and the challenge of capturing intricate patterns. In this research, LSTM-based framework with residual unit and attention mechanism is proposed to enhance the temporal dependencies and important features of stock price movements. Residual unit with skip connection captures more complex patterns and representations in stock price data and reduces the over-fitting problem to noisy time series data. LSTM with attention focuses on the significant time stamps which enhances the model prediction performance. The proposed system is experimented on five datasets: Apple (AAPL), Bitcoin, Ethereum, Litecoin and GOLD_PRICE. To prove the effectiveness of the model, the proposed system is compared with LSTM and Bidirectional LSTM (Bi-LSTM) models. Experimental results show that the proposed system outperforms baseline models such as LSTM, Bi-LSTM, LSTM+Bi-LSTM and state-of-the-art methods in term of error rates such as mean square error, root mean square error and mean absolute error.