A Hybrid Model of MEMD and PSO-LSSVR for Steel Price Forecasting

Abstract

Herein, we propose a novel hybrid method for forecasting steel prices by modeling nonlinearity and time variations together to enhance forecasting adaptability. The multivariate empirical mode decomposition (MEMD)–ensemble-EMD (EEMD) approach was employed for preprocessing to separate the nonlinear and time variation components of a hot-rolled coil (HRC) price return series, and a particle swarm optimization (PSO)-based least squares support vector regression (LSSVR) approach and a generalized autoregressive conditional heteroskedasticity (GARCH) model were applied to capture the nonlinear and time variation characteristics of steel returns, respectively. The empirical results revealed that compared with the traditional models, the proposed hybrid method yields superior forecasting performance for HRC returns. The evidence also suggested that in capturing the price dynamics of HRC during the COVID-19 pandemic period, the asymmetric GARCH model with MEMD–LSSVR outperformed not only standard GARCH models but also the EEMD-LSSVR models. The proposed MEMD–LSSVR–GARCH model for steel price forecasting provides a useful decision support tool for steelmakers and consumers to evaluate steel price trends.