Multi-Head Time Series Prediction by Decomposition and LSTM-Self-Attention-LSTM Model

Time series prediction needs advanced techniques in order to procedure a prediction result as accurate as possible. Machine learning and deep learning are known as the most sophisticated techniques in the field of time series prediction. These techniques can provide a source of tools for this prediction task. That is the reason for that in this paper we propose a new hybrid machine learning model having the structure ‘Decomposition-LSTM-Self-Attention-LSTM’ to predict the future value of time series. Our model contains three stages: analyzing, predicting, and synthesizing. For each stages, different layers and techniques were used to obtain the best prediction results. Our proposed model has been successfully applied to a real financial dataset from the Ho Chi Minh Stock Exchange. The results show that the proposed model exhibits higher forecasting accuracy than other benchmark models such as SVR, LSTM, and ConvLSTM.