RMB exchange rate forecasting using machine learning methods: Can multimodel select powerful predictors?

Abstract

AbstractThis paper aims to study the phased influencing factors of renminbi (RMB) exchange rate (CNY against USD) and investigate the predictability of the factors selected by multimodel. We first take the time points when China's main exchange reform policies are launched as the demarcation points and divide the entire sample from July 2005 to December 2020 into three periods. Then, we select the potential predictors using several sources, including all factors (without any selection), the factors selected by each of the five commonly used machine learning methods, the significantly correlated factors selected by traditional regression analysis method, and multimodel‐driven factors. Finally, we predict the exchange rate based on the above selected factors and compare the prediction results. The research results show that the main influencing factors are different in different periods, and the influence of phase events cannot be ignored. Even if their influence on the exchange rate has decreased as a result of the “811” exchange rate reform, the money supply and foreign exchange reserves continue to be the primary drivers of RMB exchange rates during the whole period of the sample. Additionally, RMB exchange rate forward is a robust influencing factor in all periods. By comparing the forecast errors, we find that the prediction accuracy of the factors selected based on multimodel is higher than that of the factors selected based on a single method or the tradition method. The findings of this paper provide the following insights for exchange rate managers: In exchange rate risk management, it is important to pay attention to the impact of macroeconomic factors such as foreign exchange reserves and the impact of staged events, and market expectations of exchange rates are equally important. At the technical level, it is recommended to improve the forecasting accuracy by forecasting exchange rates based on common factors selected by multiple better machine learning methods simultaneously rather than those selected by a single method.