基于贝叶斯优化混合CNN-RNN进行股市预测（Matlab代码实现）

最新推荐文章于 2024-07-01 21:22:25 发布

荔枝科研社

最新推荐文章于 2024-07-01 21:22:25 发布

阅读量826

点赞数

分类专栏：神经网络预测预测与分类文章标签： cnn rnn matlab

本文链接：https://blog.csdn.net/weixin_46039719/article/details/129249547

版权

神经网络预测预测与分类专栏收录该内容

136 篇文章 58 订阅

订阅专栏

👨‍🎓个人主页：研学社的博客

💥💥💞💞欢迎来到本博客❤️❤️💥💥

🏆博主优势：🌞🌞🌞博客内容尽量做到思维缜密，逻辑清晰，为了方便读者。

⛳️座右铭：行百里者，半于九十。

📋📋📋本文目录如下：🎁🎁🎁

目录

💥1 概述

📚2 运行结果

🎉3 参考文献

🌈4 Matlab代码实现

💥1 概述

通过文献调研，现阶段大部分股票择时的研究还是集中在传统的统计学模型上，将深度学习的理念应用于股票择时领域的研究寥寥无几。近年来，深度学习模型取得了长足的发展，正因为如此，CNN模型和RNN模型越来越多的进入到了人们的视野之中，而由CNN和 RNN组合形成的混合CNN-RNN模型势必有着良好的发展前;景。本文模型构建的大体思路就是将先进的CNN-RNN混合模型理念应用到股票择时令顶域,建立出能解决股票择时问题的模型，填补该领域的空白。

要点

贝叶斯优化分别用于调整混合CNN-RNN和浅层网络。

贝叶斯优化算法使用一个简单的过程来包含离散值。

此代码中使用了生成类似股票市场的过程（股票序列）。

📚2 运行结果

部分代码：

Lag = 1:optVars.Lag;
MiniBatchSize = str2double(char(optVars.MiniBatchSize));

% validation data

cv = cvpartition(size(TrainStandardizeddata,2),'HoldOut',0.2);
dataTrain = TrainStandardizeddata(cv.training);
dataValida = TrainStandardizeddata(cv.test);

% Data Preparation Train
XTrain = lagmatrix(dataTrain,Lag);
XTrain = XTrain(max(Lag)+1:end,:)';

YTrain = dataTrain(max(Lag)+1:end);
XrTrain = cell(size(XTrain,2),1);% independent variable lagged Lag months
YrTrain = zeros(size(YTrain,2),1);
for i=1:size(XTrain,2)
XrTrain{i,1} = XTrain(:,i);
YrTrain(i,1) = YTrain(:,i);
end

% Data Preparation Valida
XVal = lagmatrix(dataValida,Lag);
XVal = XVal(max(Lag)+1:end,:)';

YVal = dataValida(max(Lag)+1:end);
XrVal = cell(size(XVal,2),1);% independent variable lagged Lag months
YrVal = zeros(size(YVal,2),1);
for i=1:size(XVal,2)
XrVal{i,1} = XVal(:,i);
YrVal(i,1) = YVal(:,i);
end

% Testing data is calculated once the optimixation is finished
if exist('mviz')
XTest = lagmatrix(TestStandardizeddata,Lag);
XTest = XTest(max(Lag)+1:end,:)';
YTest = TestStandardizeddata(max(Lag)+1:end);
XrTest = cell(size(XTest,2),1);% independent variable lagged Lag months
YrTest = zeros(size(YTest,2),1);
for i=1:size(XTest,2)
XrTest{i,1} = XTest(:,i);
YrTest(i,1) = YTest(:,i);
end
end

numFeatures = size(XTrain,1);% it depends on the roll back window (No of features, one output)
% CNN-RNN
% you can follow this template and create your own Architecture
layers = [...
% Here input the sequence. No need to be modified
sequenceInputLayer([numFeatures 1 1],'Name','input')
sequenceFoldingLayer('Name','fold')

% from here do your engeneering design of your CNN feature
% extraction
convolution2dLayer(optVars.FilterSize1,optVars.NoFilter1,'Padding','same',...
'WeightsInitializer','he','Name','conv');
eluLayer('Name','elu')
convolution2dLayer(optVars.FilterSize2,optVars.NoFilter2,'Padding','same',...
'WeightsInitializer','he','Name','conv1');
eluLayer('Name','elu1')
maxPooling2dLayer(5,'Padding','same','Name','MPL')

% here you finish your CNN design and next step is to unfold and
% flatten. Keep this part like this
sequenceUnfoldingLayer('Name','unfold')
flattenLayer('Name','flatten')

% from here the RNN design. Feel free to add or remove layers
bilstmLayer(128,'Name','bilstm','RecurrentWeightsInitializer','he')
dropoutLayer(0.25,'Name','drop1')
% this last part you must change the outputmode to last
lstmLayer(64,'OutputMode',"last",'Name','lstm',...
'RecurrentWeightsInitializer','he')

% here finish the RNN design

% use a fully connected layer with one neuron because you will predict one step ahead
fullyConnectedLayer(1,'Name','fc')
regressionLayer('Name','output') ];