Multi-step-ahead stock price index forecasting using long short-term memory model with multivariate empirical mode decomposition

Abstract

Accurate and reliable multi-step-ahead forecasting of stock price indexes over long-term future trends is challenging for capital investors and decision-makers. This study developed a hybrid stock price index forecasting modelling framework using Long Short-Term Memory (LSTM) with Multivariate Empirical Mode Decomposition (MEMD), which can capture the inherent features of the complex dynamics of stock price index time series. In conjunction with time–frequency analysis and deep learning algorithms, the proposed modelling framework implemented multi-step-ahead forecasting for stock price indexes using a multiple-input multiple-output (MIMO) strategy, where MEMD was first employed to simultaneously decompose the relevant features of the stock price index. Then LSTM was used to train the forecasting model by using the components extracted by MEMD and performing multi-step-ahead forecasting of the closing price of the stock price index. The hyperparameters of the LSTM model were optimized using an orthogonal array tuning method (OATM) based on the Taguchi design of experiments for enhancing the performance of prediction. Three real-world datasets were used for model validation from three exchange markets including Standard & Poor 500 index (SPX), Shanghai Stock Exchange (SSE), and Hang Seng Index (HSI). The results of the experiments suggested that the proposed hybrid model outperforms the benchmark models and improves the accuracy of multi-step-ahead forecasting.