Abstract: Fluctuations in the foreign exchange market have a wide-ranging impact on the entire economy. The exchange rate is one of the important variables in the foreign exchange market. It is expressed as the exchange rate between the two currencies, which represents the purchasing power of one country's currency on other countries' commodities. Exchange rate fluctuations refer to changes in the value of exchange rates that fluctuate up and down, including currency depreciation and appreciation, and also refer to changes in the value of a currency relative to another country's currency. Volatility is an indicator that measures the quantity and frequency of exchange rate changes. Exchange rate volatility is an indicator to describe the degree of change in the income of foreign exchange financial assets, and it is also one of the methods to measure foreign exchange risk. Exchange rate fluctuations have an important impact on the economic and financial systems. Due to the non-stationary and nonlinear characteristics, accurate forecasting of exchange rate volatility has always been the focus and difficulty of financial research. As a measure of risk, volatility modeling is important not only to researchers trying to understand the dynamics of volatility, but also to policymakers and regulators. Because it is closely related to the operation and stability of the financial market, and the financial market is directly related to the operation and fluctuation of the real economy. Volatility is also an important input parameter for portfolio decision-making models or option pricing, so the measurement of exchange rate volatility is related to forecasting, which is of great significance to investment decision-making, derivatives pricing, and risk management.
With the development of information technology, machines are doing more and more “smart” things such as recognizing faces in photos, recognizing voices, and making predictions. The current research work shows that the forecasting effect based on machine learning is better than other traditional forecasting models. This paper mainly uses machine learning algorithms to build a forecasting model of RMB exchange rate volatility. In order to improve the accuracy of forecasting exchange rate volatility, this paper adopts the realized volatility calculated based on high-frequency data of RMB exchange rate and machine learning method to decompose, integrate and model the data (main data sources: Monopoly data center and Wind database) , an efficient multiscale EEMD-PSR-SVR-ARIMA forecasting model is proposed. The specific process is as follows: First, the complex time series is decomposed into eigenmode functions and trend items of different scales by using the method of Ensemble Empirical Mode Decomposition (EEMD). The core of the SVR algorithm process is the kernel function. What it does is transform the original input space into a high-dimensional space where linear decision boundaries can be easily identified. In the prediction process, the phase space reconstruction method is used to calculate the optimal input dimension of support vector regression, the particle swarm optimization algorithm is used to select its optimal parameters, and then the support vector with the optimal input dimension and optimal parameters is adopted. The regression model makes predictions for different eigenmode functions. On the basis of the EEMD-PSR-SVR model, we build a multi-scale model to improve performance, and use a divide-and-conquer method. The intrinsic mode function is partially predicted by SVR. The trend item belongs to the low-frequency subsequence, and the ARIMA model is used for prediction. For linear low-frequency, the time series has a good prediction effect, which reflects its good at capturing linear trends. Finally, the prediction results of different intrinsic mode functions and trend items are integrated as the final result. The principles of “decomposition and integration” and “divide and conquer” also effectively improve the prediction accuracy and direction prediction accuracy of the model.
At the same time, in order to comprehensively evaluate the prediction effect of the model proposed in this paper, different reference group models will be introduced for comparative analysis, and the prediction performance of the model can be evaluated from three aspects: The relative level of prediction, the absolute level and the direction of prediction. The empirical results show that the EEMD-PSR-SVR-ARIMA model can effectively improve the accuracy of exchange rate volatility prediction. Facing the complex and changeable foreign exchange market environment, there are still a lot of problems in the study of exchange rate volatility, which need to be explored and studied. With the vigorous development of machine learning methods, especially deep learning methods, we will further explore the modeling and prediction of volatility.

Key words: exchange rate volatility forecast, ensemble empirical mode decomposition, phase space reconstruction, support vector regression

摘要： 汇率波动率是刻画外汇金融资产收益变化程度的指标,也是度量外汇风险的方法之一,汇率波动对经济与金融系统都有重要的影响。由于非平稳和非线性的特征,准确预测汇率波动率一直是金融研究的重点和难点。为了提高预测汇率波动率的准确性,本文采用基于人民币汇率高频数据计算的已实现波动率和机器学习方法,对数据进行分解集成和建模,提出了一种有效的多尺度EEMD-PSR-SVR-ARIMA预测模型。具体过程如下:首先,采用集合经验模态分解(EEMD)的方法将复杂的时间序列分解成不同尺度的本征模态函数和趋势项;然后采用支持向量回归(SVR)的方法对本征模态函数进行预测,并利用相空间重构和粒子群优化的方法来确定SVR模型的输入维数与参数。同时,使用差分自回归移动平均模型(ARIMA)预测趋势项;最后集成得到模型预测的结果。实证结果表明 EEMD-PSR-SVR-ARIMA 模型可以有效地提高汇率波动率预测的精度。

关键词: 汇率波动率预测, 集成经验模态分解, 相空间重构, 支持向量回归