A Novel Non-Ferrous Metals Price Forecast Model Based on LSTM and Multivariate Mode Decomposition

Abstract

Non-ferrous metals are important bulk commodities and play a significant part in the development of society. Their price forecast is of great reference value for investors and policymakers. However, developing a robust price forecast model is tricky due to the price’s drastic fluctuations. In this work, a novel fusion model based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Singular Spectrum Analysis (SSA), and Long Short-Term Memory (LSTM) is constructed for non-ferrous metals price forecast. Considering the complexity of their price change, the dual-stage signal preprocessing which combines CEEMDAN and SSA is utilized. Firstly, we use the CEEMDAN algorithm to decompose the original nonlinear price sequence into multiple Intrinsic Mode Functions (IMFs) and a residual. Secondly, the component with maximum sample entropy is decomposed by SSA; this is the so-called Multivariate Mode Decomposition (MMD). A series of experimental results show that the proposed MMD-LSTM method is more stable and robust than the other seven benchmark models, providing a more reasonable scheme for the price forecast of non-ferrous metals.