基于AFD-LSTM模型的金融信号去噪与预测

Denoising and Prediction of Financial Signals Based on AFD-LSTM Model

旨在提高金融时间序列数据预测的准确性,提出了将自适应Fourier分解(AFD)方法和长短时记忆神经网络模型(LSTM)相结合的预测模型。AFD是一种信号处理方法,拥有Fourier变换所不具备的自适应性,可以快速提取金融信号的特征,达到去除信号噪声污染的目的。而LSTM模型可以挖掘时间序列数据的依赖关系,对于存在长记忆性的金融时间序列数据的预测十分有效。基于AFD-LSTM方法以美元兑人民币汇率,深证700股票指数,伦敦黄金现价和广东省碳排放配额(GDEA)价格四种金融信号数据为研究对象进行了实证分析,并将其和自适应噪声完备经验模态分解方法(CEEMDAN)进行对比。结果表明,基于AFD方法去噪后的金融数据所训练的LSTM网络模型具有较高的预测精度,不依赖于LSTM模型的层数参数,稳定性较强。

Aiming to improve the accuracy of data prediction of financial time series,the paper proposes a prediction model that combines the Adaptive Fourier Decomposition(AFD)method with the Long Short-Term Memory(LSTM)neural network model.AFD is a signal processing method,which has the adaptability that Fourier transform does not have.It can quickly extract the characteristics of financial signals and achieve the purpose of removing signal noise pollution.The LSTM model can explore the dependency relationships of time series data,which is very effective for predicting financial time series data with long memory.Based on AFD-LSTM model,the research conducts the empirical analysis on four kinds of financial signal data,namely,USD/RMB exchange rate,SZSE 700 stock index,current price of gold in London and Guangdong Province carbon emission quota(GDEA)price and compares them with the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN).The results show that the LSTM network model trained based on the financial data denoising by the AFD method has high predictive accuracy and does not depend on the layer number parameters of the LSTM model and has strong stability.