import numpy as np
import pandas as pd
from pylab import plt
plt.style.use('seaborn')
%matplotlib inline
import pandas_datareader as pdr



start = '2015-01-01'
end = '2019-12-31'
data = pdr.get_data_yahoo('600001',start,end)



data



data_cp = data



data = data['Close']



data.columns = ['prices']



data = pd.DataFrame(data)



data



data.info



data.plot(figsize=(10, 6));



data['returns'] = np.log(data / data.shift(1))



data.head()



lags = 5
cols = []
for lag in range(1, lags+1):
    col = 'ret_%d' % lag
    data[col] = data['returns'].shift(lag)
    cols.append(col)



data



data.dropna(inplace=True)



data.head()







# OLS regression
reg = np.linalg.lstsq(data[cols].values, np.sign(data['returns'].values))[0]



reg



pred = np.sign(np.dot(data[cols].values, reg))



pred



np.sign(data['returns'].values)



data['ols_pred'] = pred



c = np.sign(data['returns'] * data['ols_pred'])



c.value_counts()



c.value_counts()[1] / (c.value_counts().sum())



data['ols_returns'] = data['returns'] * data['ols_pred']



data[['returns', 'ols_returns']].cumsum().apply(np.exp).plot(figsize=(10, 6));



from sklearn import linear_model
lm = linear_model.LogisticRegression(C = 1e6)
lm.fit(data[cols], np.sign(data['returns']))



data['log_pred'] = lm.predict(data[cols])
data.head()



data['log_returns'] = data['returns'] * data['log_pred']



data[['returns', 'ols_returns', 'log_returns']].cumsum(
        ).apply(np.exp).plot(figsize=(10, 6));



import tensorflow as tf



tf.logging.set_verbosity(tf.logging.ERROR)



mean = data['returns'].mean()
std = data['returns'].std()



print(mean, std)



fc = tf.contrib.layers.real_valued_column('returns', dimension=lags)



fcb = [tf.contrib.layers.bucketized_column(fc,
                    boundaries=[-0.0005, 0.0001, 0.0005])]



model = tf.contrib.learn.DNNClassifier(hidden_units=[50, 50],
                                       feature_columns=fcb)



def get_data():
    fc = {'returns': tf.constant(data[cols].values)}
    la = tf.constant((data['returns'] > 0).astype(int).values,
                     shape=[len(data), 1])
    return fc, la



model.fit(input_fn=get_data, steps=100)



model.evaluate(input_fn=get_data, steps=1)



data['dnn_pred'] = list(model.predict(input_fn=get_data))
data['dnn_pred'] = np.where(data['dnn_pred'] > 0, 1.0, -1.0)



data['dnn_pred']



data['dnn_returns'] = data['returns'] * data['dnn_pred']



data[['returns', 'ols_returns', 'log_returns', 'dnn_returns']].cumsum(
        ).apply(np.exp).plot(figsize=(10, 6));

问题：

读取AMZN股票的历史收盘价数据，训练神经网络或者其他机器学习模型，
以延迟特征（日收益率）为输入进行涨跌预测。
并将模型预测结果转化为策略，应用与历史交易（可以是训练数据也可以是其他时间段）
计算算法交易策略下的回报是如何