Price forecasts of ten steel products using Gaussian process regressions

Abstract

Addressing price forecasting problems is an important exercise to policymakers and market participants in the resource business sector. In this work, we build Gaussian process regression models through cross validation and Bayesian optimizations over different kernels and basis functions for daily price index forecasting of ten major steel products in the Chinese market during July 20, 2011–April 15, 2021. This study aims to be the first attempt of exploring potential of Gaussian process regressions for price forecasting exercises with the coverage of all of the ten most important steel products that carry enormous economic significance in China as the largest steel consumer and producer in the world. The models offer accurate out-of-sample forecasts for the two-year period of April 16, 2019–April 15, 2021 with relative root mean square errors ranging from 0.07404% to 0.22379% across the ten price indices and correlation coefficients above 99.9%. They also lead to better forecast performance than some traditional econometric models and some other machine learning models as benchmarks. The models constructed here could be utilized by policymakers as part of policy design and implementations and by market participants as part of market assessments and decision making.