基于频率分解的机器学习模型预测效果比较OA北大核心CHSSCDCSSCICSTPCD
Comparison of Prediction Effect of Machine Learning Models Based on Frequency Decomposition
文章引入一种小波变换与机器学习的组合预测方法,通过小波变换提取单变量时间序列的主要特征,并应用不同的机器学习模型进行预测分析.构建不同类型的机器学习模型对上证指数、恒生指数、纳斯达克指数和日经225指数进行预测,结果表明:在不增加任何被解释变量的情况下,经过小波变换的数据特征能较好地预测指数收益率;通过比较线性模型、机器学习模型和深度学习模型发现,线性模型在捕获小波变换特征方面表现最好;有效的数据降维方法是提高非线性模型样本外预测精度的重要手段,并且可以减少模型训练的时间;小波变换和贝叶斯混合模型的预测精度高于传统的ARMA模型.
This paper introduces a combined prediction method of wavelet transform and machine learning,extracts the main features of univariate time series by wavelet transform,and uses different machine learning models for prediction analysis.Differ-ent types of machine learning models are constructed to forecast the Shanghai Composite Index,Hang Seng Index,Nasdaq Index and Nikkei 225 Index.The results are shown as follows:Without adding any explained variables,the data features transformed by wavelet can predict the index return better.By comparing linear models,machine learning models and deep learning models,it is found that linear models perform best in capturing wavelet transform features.Effective dimensionality reduction method is an im-portant means to improve the out-of-sample prediction accuracy of nonlinear models and reduce the time of model training.The prediction accuracy of wavelet transform and Bayesian mixed model is better than that of traditional ARMA model.
陈煜之;李心悦;方毅
吉林大学 商学与管理学院,长春 130012东北师范大学 外国语学院,长春 130022吉林大学 商学与管理学院,长春 130012||吉林大学 数量经济研究中心,长春 130012
计算机与自动化
深度神经网络随机梯度下降长短期记忆神经网络小波变换随机森林
deep neural networkstochastic gradient descentlong short-term memory neural networkswavelet transformrandom forest
《统计与决策》 2024 (012)
29-34 / 6
国家自然科学基金资助项目(71871104)
评论