Interval-Valued Time Series Prediction for Vietnam Stock Indicators Based on Ensemble Long Short-Term Memory Networks

Abstract

The prediction of interval-valued data has many advantages over point prediction in many cases because it takes into account the data’s volatility and uncertainty. However, the number of studies related to interval time series has been still limited so far. This paper pioneers a new framework for applying deep learning and ensemble learning to forecast for interval-valued data. Firstly, we proposed a new two-stage Monte Carlo algorithm to generate N interval time series from an original one. Secondly, for each interval time series in the generated ones, two Long Short-Term Memory Network models are trained to predict the middle point and the range. Finally, the prediction for the middle point and range could be calculated by the average of the forecast results. The proposed method is applied to forecast VN-Index as well as some other financial indicators. The results show that the proposed method has higher accuracy and stability than alternative approaches.