前言

本文主要是采用lstm对汇率涨跌进行预测，是一个二分类的预测问题。

步骤解析

数据构造

原始数据是单变量数据

import  pandas as pd

file_path  = r"./huilv.csv"
data = pd.read_csv(file_path, usecols=[1],encoding='gbk')
data['level'] = -1   

 	美元	level
0	7.2996	-1
1	7.2775	-1
2	7.2779	-1
3	7.2695	-1
4	7.2791	-1
...	...	...
3655	6.9475	-1
3656	6.9131	-1
3657	6.8926	-1
3658	6.8912	-1
3659	6.8265	-1
3660 rows × 2 columns

构造成两列数据，后一列为涨跌情况

for i in range(len(data)-1):
    if data.iloc[i+1,0] -  data.iloc[i,0] > 0:
         data.iloc[i+1,1] = 1
    else :
         data.iloc[i+1,1] = 0

	美元	level
0	7.2996	-1
1	7.2775	0
2	7.2779	1
3	7.2695	0
4	7.2791	1
...	...	...
3655	6.9475	0
3656	6.9131	0
3657	6.8926	0
3658	6.8912	0
3659	6.8265	0
3660 rows × 2 columns

构造训练和测试集

def scaler(X_train,X_test):
    """
    数据归一化
    """
    mm = MinMaxScaler()
    mm.fit(X_train)
    mmX_train = mm.transform(X_train)
    mmX_test = mm.transform(X_test)
    return mmX_train,mmX_test

def create_dataset(dataset, dataset_label, look_back=1):
    dataX, dataY = [], []
    for i in range(len(dataset_label)-look_back-1):
        a = dataset[i:(i + look_back), :]
        dataX.append(a)
        b = dataset_label[i+look_back, 0]
        dataY.append(b)
        print('a: %s, b: %s' % (a, b))
    return numpy.array(dataX), numpy.array(dataY)

构造模型

# create and fit the LSTM network
model = Sequential()
model.add(LSTM(64, input_shape=(look_back,1)))  # 与上面的重构格式对应，要改都改，才能跑通代码
model.add(Dense(1))
model.compile(loss='binary_crossentropy', optimizer='adam',metrics=['acc'])
history =  model.fit(trainX, trainY, epochs=100, batch_size=1000, validation_split=0.2)

损失函数图像

预测

# 模型预测并计算结果
from sklearn.metrics import confusion_matrix
y_pred = model.predict(testX)
# 计算混淆矩阵
cm = confusion_matrix(testY, y_pred.argmax(axis=1))
print('Confusion Matrix')
print(cm)
# 评价指标
print("Precision: ",  (cm[0][0] + cm[1][1])/ (cm[0][0]+cm[1][0]+cm[0][1]+cm[1][1]))
print("Accuracy: ",  (cm[0][0] + cm[0][1])/ (cm[0][0]+cm[1][0]+cm[0][1]+cm[1][1]))
print("Recall",  (cm[0][0] )/ (cm[0][0]+cm[1][0]))

23/23 [==============================] - 0s 1ms/step
Confusion Matrix
[[384   0]
 [344   0]]
 Precision:  0.5274725274725275
Accuracy:  0.5274725274725275
Recall 0.5274725274725275

总结

• lstm 二分类预测，效果看起来一般，算是第一次接触分类问题
• 结果有更进一步的优化，本案例只做跑通整个流程

备注

需要完整源代码和数据集,或者想要沟通交流,请私聊,谢谢.