基于神经网络分位数回归的人民币汇率概率密度预测

Probability density prediction of RMB exchange rate via neural network quantile regression

中国执行以外汇市场供求为基础的有管理浮动汇率制度,人民币汇率形成机制较为复杂,其影响因素可能存在非线性效应。为此,将人民币汇率作为输出变量,其影响因素作为输入变量,考虑从输入到输出的非线性效应,构建神经网络分位数回归模型((quantile regression neural network,QRNN),有助于理解人民币汇率决定机制。选取人民币兑美元汇率为研究对象,建立了神经网络分位数回归模型并进行概率密度预测,将其预测效果同线性分位数回归、BP神经网络、线性均值回归等方法进行比较,实证研究结果表明:第一,QRNN模型通过神经网络的非线性处理能力,显著提高了预测准确程度;第二,QRNN模型通过分位数回归得到的概率密度预测结果,能够预测人民币汇率完整条件分布信息,便于科学决策。

China implements a managed floating exchange rate system based on supply and demand of foreign exchange market.In this regard,the formation mechanism of the RMB exchange rate is relatively complicated with obvious nonlinear effects.Using the RMB exchange rate as an output variable and its influencing factors as input variables,we consider the nonlinear nexus between the inputs and the output to construct a neural network quantile regression(QRNN)model,which helps to understand the RMB exchange rate determination mechanism.In the empirical analysis,we consider the exchange rate of RMB against the US dollar.A QRNN model is established and the probability density prediction is performed.The QRNN model is compared with several competing models,such as linear quantile regression,BP neural network,and linear mean regression.The empirical results bring us at least two conclusions.First,the QRNN model significantly improves the accuracy of prediction by exploiting the ability of neural network in handling nonlinear nexus.Second,the QRNN model provides the probability density prediction results of the RMB exchange rate,which is convenient for scientific decision-making.