简介：程序调用了tensorflow.keras搭建的LSTM。通过输入前五天的"收盘价"(close)、"开盘价"(open)、"最高价"(high)、"最低价"(low)、来预测未来一天的收盘价。

本程序是通过tushare平台的API获取的数据。长短记忆网络是基于RNN网络修改过来的循环网络。可以解决RNN无法长距离的依赖问题。tushare ID：499921

直接看代码：

路径目录

baseDir = '/test' #当前目录
staticDir = os.path.join(baseDir,'Static') # 静态文件目录
resultDir = os.path.join(baseDir,'Result') # 结果文件目录

当前目录根据不同路径来设置

读取数据 通过tushare获取数据集

ts.set_token('这里是放你的token')
pro = ts.pro_api()
ts_code = '000001.SZ' # 平安股票
start_date = '2006-01-01'
end_date = '2020-01-01'
df = pro.daily(
    ts_code=ts_code,
    start_date=start_date,
    end_date=end_date
)

标准化数据集

df = df.set_index(['trade_date'])
ColName = df.columns.values
use_cols = ['close','open','high','low','vol','amount']
data = df[use_cols]
# 标准化数据集
outputCol = ['close'] # 输出列
inputCol = ['close','high','low','open'] # 输入列
X = data[inputCol]
Y = data[outputCol]
xScaler = StandardScaler()
yScaler = StandardScaler()
X = xScaler.fit_transform(X) 
Y = yScaler.fit_transform(Y)

确定输入天数和输出天数

#按时间
timeStep = 5 # 输入天数
outStep = 1 # 输出天数
xAll = list()
yAll = list()
# 按时间步骤整理数据 输入数据尺寸是(timeStep, 5) 输出尺寸是(outSize)
for row in range(data.shape[0]-timeStep-outStep+1):
    x = X[row:row+timeStep]
    y = Y[row+timeStep:row+timeStep+outStep]
    xAll.append(x)
    yAll.append(y)
xAll = np.array(xAll).reshape(-1,timeStep,len(inputCol))
yAll = np.array(yAll).reshape(-1,outStep)

按照时间序列将每五天的数据存入输入数据，5天后一天存入输出数据。

将所有数据分成测试集和训练集

testRate = 0.2 #测试比例
splitIndex = int(xAll.shape[0]*(1-testRate)) # 所有数据条数*训练集比例 = 训练集条数
xTrain = xAll[:splitIndex]
xTest = xAll[splitIndex:]
yTrain = yAll[:splitIndex]
yTest = yAll[splitIndex:]

搭建简单的LSTM网络

tanh：激活函数
timeStep:输入时间同步
inputColNum: 输入列数d
outStep: 输入时间步
learnRate: 学习率

def buildLSTM(timeStep, inputColNum,outStep,learnRate=1e-4):
    # 输入层
    inputLayer = Input(shape=(timeStep,inputColNum))
    # 中间层
    middle = LSTM(100,activation='tanh')(inputLayer)
    middle = Dense(100,activation='tanh')(middle)
    # 输出层 全连接
    outputLayer = Dense(outStep)(middle)
    # 建模
    model = Model(inputs=inputLayer,outputs=outputLayer)
    optimizer = Adam(lr=learnRate)
    model.compile(optimizer=optimizer,loss='mse') 
    model.summary()
    return model
#搭建LSTM
lstm = buildLSTM(timeStep=timeStep,inputColNum=len(inputCol),outStep=outStep,learnRate=1e-4)

训练网络

epochs = 1000 #迭代次数
batchSize = 500 #批处理量
lstm.fit(xTrain,yTrain,epochs=epochs,verbose=0,batch_size=batchSize) 

预测 测试集对比

yPredict = lstm.predict(xTest)
yPredict = yScaler.inverse_transform(yPredict)[:,0]
yTest = yScaler.inverse_transform(yTest)[:,0]
result = {'观测值':yTest,'预测值':yPredict}
result = pd.DataFrame(result)
result.index = data.index[timeStep+xTrain.shape[0]:result.shape[0]+timeStep+xTrain.shape[0]]
result.to_excel(resultDir+'\预测结果.xlsx')
result.head()
mae = MAE(result['观测值'],result['预测值'])
print('模型测试集MAE',mae)

可视化

fig,ax = plt.subplots(1,1)
ax.plot(result.index,result['预测值'],label='预测值')
ax.plot(result.index,result['观测值'],label='观测值')
ax.set_title('LSTM预测效果，MAE：%2f'%mae)
ax.legend()
ax.xaxis.set_major_locator(ticker.MultipleLocator(5))
fig.savefig(resultDir+'/预测折线图.png',dpi=500)