基于CEEMDAN与LSTM的人民币汇率分析与预测

RMB Exchange Rate Analysis and Its Forecasting Based on CEEMDAN and LSTM

由于汇率波动的复杂性和非线性特点,使得汇率预测一直是金融领域中最具挑战的课题之一。本文提出了一种基于分解-优化预测-集成的综合研究方法来分析预测人民币汇率。首先,利用自适应噪声完备经验模态分解(CCEMDAN)方法,将原始汇率序列分解为若干不同频率的分量序列;接着通过Hilbert谱分析和相关统计方法确定高频、低频和趋势等三种结构分量,并将高频分量序列合并优化为一个新的分量序列;然后运用长短时记忆神经网络(LSTM)模型分别对各分量进行预测;最后将这些预测结果集成得到汇率的最终预测结果。本文以美元、欧元、英镑和日元兑人民币4种汇率为研究对象,研究发现:1)4种汇率的价格及波动受趋势分量和低频分量的影响较大,受高频分量的影响较小;2)欧元、英镑和日元兑人民币汇率受随机波动的影响要远大于美元兑人民币汇率所受到的影响,对短线投资者和机构来说,相比美元兑人民币汇率,关注另外三种汇率的高频分量可能具有更重要的意义;3)对比了其它10种模型(包括5种综合模型和5种单一模型)的预测结果,本文所提出的模型无论在预测精度还是在预测方向准确率上,表现都是最佳的,也充分说明该模型预测的有效性。此外,本文所提出的研究方法框架对其它金融时间序列的研究也具有一定的借鉴和参考价值。

Due to the complexity and non-linear characteristics of exchange rate fluctuations,exchange rate forecasting has been one of the most challenging topics in the financial field.This paper proposes a idea of decomposition-optimization and prediction-integration to predict the RMB exchange rate.First,complete ensemble empirical mode decomposition with adaptive noise(CCEMDAN)method is used to decompose the original exchange rate sequence into several component sequences of different frequencies.Next,we use Hilbert spectrum and statistical method to determine the three structural components of high frequency,low frequency and trend,and the high frequency component sequence is combined into a new component sequence.Then,each component is predicted by short-term memory neural network(LSTM)model.Finally,these predictions are integrated to obtain the final exchange rate prediction result.This paper takes the four exchange rates of the USD,EUR,GBP and JPY against RMB as research objects.The experimental results show that:1)The prices and fluctuations of the exchange rates are mainly affected by the trend component and the low-frequency component,and are less affected by the high-frequency component;2)EUR,GBP and JPY against RMB are much more affected by random fluctuations than USD against RMB.It may be more important to focus on the high-frequency components of the other three exchange rates than USD against RMB to short-term investors and institutions;3)Compared with the prediction results of the other 10 models(including 5 comprehensive models and 5 single models),the model proposed in this paper performs the best in both prediction accuracy and prediction direction accuracy.In addition,the method framework proposed in this paper also has certain reference value for the study of other financial time series.