股票预测模型的改良
import numpy as np
import pandas as pd
from sklearn import svm
from sklearn.linear_model import LogisticRegression
import tushare as ts
from sklearn import cross_validationdata=ts.get_k_data('600000',start='2007-01-01',end='2018-04-13')print(data.head())
print(data.tail()) date open close high low volume code
0 2007-01-04 3.702 3.670 4.009 3.413 508894.23 600000
1 2007-01-05 3.670 3.562 3.670 3.338 357055.86 600000
2 2007-01-08 3.557 3.634 3.708 3.525 254888.76 600000
3 2007-01-09 3.596 3.897 3.916 3.591 329619.46 600000
4 2007-01-10 3.904 4.041 4.152 3.904 352768.36 600000
date open close high low volume code
2686 2018-04-09 11.53 11.50 11.59 11.49 167224.0 600000
2687 2018-04-10 11.52 11.77 11.79 11.51 287482.0 600000
2688 2018-04-11 11.79 11.91 12.02 11.75 312985.0 600000
2689 2018-04-12 11.91 11.78 11.96 11.76 188242.0 600000
2690 2018-04-13 11.83 11.69 11.89 11.69 140948.0 600000
data_SZ_index=ts.get_k_data('000001',index=True,start='2007-01-01',end='2018-04-13')print(data_SZ_index.head())
print(data_SZ_index.tail()) date open close high low volume code
0 2007-01-04 2728.19 2715.72 2847.61 2684.82 120156000.0 sh000001
1 2007-01-05 2668.58 2641.33 2685.80 2617.02 106156000.0 sh000001
2 2007-01-08 2621.07 2707.20 2708.44 2620.62 106813000.0 sh000001
3 2007-01-09 2711.05 2807.80 2809.39 2691.36 110751000.0 sh000001
4 2007-01-10 2838.11 2825.58 2841.74 2770.99 111769000.0 sh000001
date open close high low volume code
2738 2018-04-09 3125.44 3138.29 3146.09 3110.30 139608621.0 sh000001
2739 2018-04-10 3144.26 3190.32 3190.65 3139.08 168201359.0 sh000001
2740 2018-04-11 3197.37 3208.08 3220.85 3191.59 175867197.0 sh000001
2741 2018-04-12 3203.28 3180.16 3205.25 3177.05 148231313.0 sh000001
2742 2018-04-13 3192.04 3159.05 3197.90 3155.51 127552310.0 sh000001
num_data=len(data)
num_SZ_index=len(data_SZ_index)print(num_data,num_SZ_index)#股票会有停牌天数，但指数不会(2691, 2743)
from datetime import datetime
data['date'] = [datetime.strptime(x,'%Y-%m-%d') for x in data['date']]data['date'].head()0 2007-01-04
1 2007-01-05
2 2007-01-08
3 2007-01-09
4 2007-01-10
Name: date, dtype: datetime64[ns]
data_SZ_index['date'] = [datetime.strptime(x,'%Y-%m-%d') for x in data_SZ_index['date']]subdata_SZ_index=data_SZ_index[data_SZ_index['date'].isin(data['date'])]#数据对齐sub_index_open=subdata_SZ_index['open'].values #z做了对齐之后，丢失一部分大盘数据对index有影响，直接取数据部分
sub_index_close=subdata_SZ_index['close'].values
col_index=[]
y=[]
data_open=data['open']
data_close=data['close']
for i in xrange(2691):
if sub_index_close[i]>=sub_index_open[i]:
col_index.append(1)
else:
col_index.append(0)
if data_close[i]>=data_open[i]:
y.append(1)
else:
y.append(0)
x_data=data[['open','close','high','low','volume']].as_matrix()x=np.c_[x_data,col_index]#将大盘指数的涨跌合并到特征值中data_shape=x.shape
data_rows=data_shape[0]
data_cols=data_shape[1]
data_col_max=x.max(axis=0)
data_col_min=x.min(axis=0)
print(data_col_max,data_col_min)(array([ 1.36700000e+01, 1.37600000e+01, 1.40200000e+01,
1.35300000e+01, 1.19802410e+07, 1.00000000e+00]), array([ 2.42500000e+00, 2.47000000e+00, 2.65400000e+00,
2.41600000e+00, 2.89912100e+04, 0.00000000e+00]))
for i in xrange(0, data_rows, 1):#将输入数组归一化
for j in xrange(0, data_cols, 1):
x[i][j] = \
(x[i][j] - data_col_min[j]) / \
(data_col_max[j] - data_col_min[j])
print(x[0:2])[[ 0.11356158 0.10628875 0.1192152 0.08970668 0.04015505 0. ]
[ 0.11071587 0.09672276 0.08938941 0.08295843 0.02745024 0. ]]
y=y[1:2691]x=x[0:2690]clf1 = svm.SVC(kernel='rbf')
clf2 = LogisticRegression()
result1 = []
result2 = []
for i in range(5):
# x和y的验证集和测试集，切分80-20%的测试集
x_train, x_test, y_train, y_test = \
cross_validation.train_test_split(x, y, test_size=0.2)
# 训练数据进行训练
clf1.fit(x_train, y_train)
# 将预测数据和测试集的验证数据比对
result1.append(np.mean(y_test == clf1.predict(x_test)))
clf2.fit(x_train, y_train)
result2.append(np.mean(y_test == clf2.predict(x_test)))
print("svm classifier accuacy:")
print(result1)
print("LogisticRegression classifier accuacy:")
print(result2)svm classifier accuacy:
[0.53345724907063197, 0.53717472118959109, 0.55018587360594795, 0.51115241635687736, 0.54460966542750933]
LogisticRegression classifier accuacy:
[0.55390334572490707, 0.55576208178438657, 0.53717472118959109, 0.52416356877323422, 0.53903345724907059]
这篇代码主要改进是将大盘的涨跌加入到了模型的参数中，并且是通过股票昨天的情况来预测明天的涨跌，发现SVM和逻辑回归效果差不多