山东-日照分赛场-公积金贷款逾期预测

数据链接：http://data.sd.gov.cn/cmpt/cmptDetail.html?id=26

评分标准：给定一个阀值，可根据混淆矩阵计算TPR（覆盖率）和FPR（打扰率） TPR = TP /（TP + FN） FPR = FP /（FP + TN） 其中，TP、FN、FP、TN分别为真正例、假反例、假正例、真反例。 这里的评分指标，首先计算了3个覆盖率TPR： TPR1：FPR=0.001时的TPR TPR2：FPR=0.005时的TPR TPR3：FPR=0.01时的TPR 最终成绩= 0.4 * TPR1 + 0.3 * TPR2 + 0.3 * TPR3

代码如下：

def tpr_weight_funtion(y_true,y_predict):
    d = pd.DataFrame()
    d['prob'] = list(y_predict)
    d['y'] = list(y_true)
    d = d.sort_values(['prob'], ascending=[0])
    y = d.y
    PosAll = pd.Series(y).value_counts()[1]
    NegAll = pd.Series(y).value_counts()[0]
    pCumsum = d['y'].cumsum()
    nCumsum = np.arange(len(y)) - pCumsum + 1
    pCumsumPer = pCumsum / PosAll
    nCumsumPer = nCumsum / NegAll
    TR1 = pCumsumPer[abs(nCumsumPer-0.001).idxmin()]
    TR2 = pCumsumPer[abs(nCumsumPer-0.005).idxmin()]
    TR3 = pCumsumPer[abs(nCumsumPer-0.01).idxmin()]
    return 0.4 * TR1 + 0.3 * TR2 + 0.3 * TR3

其他的就不多说的，直接上第一版的代码

# -*- coding: utf-8 -*-
"""
Created on Wed Jan 27 09:44:50 2021

@author: Administrator
"""

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import datetime
import warnings
warnings.filterwarnings('ignore')

#%%导入数据
train = pd.read_csv('E:\山东竞赛\公积金逾期预测\公积金逾期预测-数据\公积金逾期预测-数据\train.csv')
test = pd.read_csv('E:\山东竞赛\公积金逾期预测\公积金逾期预测-数据\公积金逾期预测-数据\test.csv')

'''
XINGBIE 3
CSNY 393
HYZK 4  
ZHIYE 7  
ZHICHEN 4
ZHIWU 2
XUELI 2
DWJJLX 28
DWSSHY 21
'''
#%%性别
train.XINGBIE.value_counts(normalize=True)  #男女比列63%：36%
test.XINGBIE.value_counts(normalize=True)  #65%:34%


#%%先复制一个表
train_copy = train.copy()


#%%出生年月
train_copy.CSNY.value_counts(normalize=True).sort_index()

train_copy = train_copy[~train_copy.CSNY.isin([190364572800])]
train_copy.shape  #(39999, 21)
train_copy['birthday'] = train_copy.apply(lambda x:datetime.datetime.fromtimestamp(x.CSNY),axis=1)

train_copy.birthday.value_counts(normalize=True).sort_index().to_excel('tmp25.xlsx')
import pycard as pc
train_copy['to_date'] = datetime.date(2020, 9, 30)
train_copy['age'] = pc.month_diff(train_copy['to_date'],train_copy['birthday'])
train_copy['age'] = train_copy['age']/12
train_copy['age'] = train_copy['age'].astype('int')

train_copy.age.value_counts(normalize=True).sort_index().to_excel('tmp25.xlsx')


#%%测试集
test.CSNY.value_counts(normalize=True).sort_index()

test.CSNY[test.CSNY==150445008000] = 1330531200
test.shape  #(39999, 21)
test['birthday'] = test.apply(lambda x:datetime.datetime.fromtimestamp(x.CSNY),axis=1)


test['to_date'] = datetime.date(2020, 9, 30)
test['age'] = pc.month_diff(test['to_date'],test['birthday'])
test['age'] = test['age']/12
test['age'] = test['age'].astype('int')

test.age.value_counts(normalize=True).sort_index().to_excel('tmp25.xlsx')



#%%婚姻状况
train_copy.HYZK.value_counts(normalize=True)  #99.99%都是一类
test.HYZK.value_counts(normalize=True)  #100%都是一类

train_copy.pop('HYZK')
test.pop('HYZK')

#%%职业
train_copy.ZHIYE.value_counts(normalize=True)  #99.99%都是一类
test.ZHIYE.value_counts(normalize=True)  #100%都是一类

train_copy.pop('ZHIYE')
test.pop('ZHIYE')

#%%职称
train_copy.ZHICHEN.value_counts(normalize=True)  #99.99%都是一类
test.ZHICHEN.value_counts(normalize=True)  #100%都是一类

train_copy.pop('ZHICHEN')
test.pop('ZHICHEN')

#%%职务
train_copy.ZHIWU.value_counts(normalize=True)  #99.99%都是一类
test.ZHIWU.value_counts(normalize=True)  #3-7分，有干扰数据

train_copy.pop('ZHIWU')
test.pop('ZHIWU')

#%%学历
train_copy.XUELI.value_counts(normalize=True)  #99.99%都是一类
test.XUELI.value_counts(normalize=True)  #100%都是一类

train_copy.pop('XUELI')
test.pop('XUELI')

#%%单位
train_copy.DWJJLX.value_counts(normalize=True)  #99.99%都是一类
test.DWJJLX.value_counts(normalize=True)  #100%都是一类

#%%重复，删除一个
train_copy.pop('GRYJCE')
test.pop('GRYJCE')

#%%个人缴存基

#年利率；济南现行政策中5年以下是2.75,5年以上是3.25。职工若第二次贷款，利率上浮10%。

#%%类别型变量的iv值计算
cate_col = ['DWJJLX', 'DWSSHY','GRZHZT','DKLL']

import pycard as pc
cate_iv_woedf = pc.WoeDf()
clf = pc.NumBin()
for i in cate_col:
    cate_iv_woedf.append(pc.cross_woe(train_copy[i] ,train_copy.label))
cate_iv_woedf.to_excel('tmp25')

#%%类别变量处理
map_dict={140:0,
142:0,
143:0,
149:0,
171:0,
172:0,
173:0,
174:0,
175:0,
210:0,
230:0,
310:0,
330:0,
390:0,
130:0,
159:0,
300:0,
120:1,
160:1,
200:1,
150:1,
141:1,
170:1,
110:2,
900:3,
100:4,
179:4,
190:4}

train_copy['DWJJLX_bin'] = train_copy.DWJJLX.map(map_dict)

map_dict1 = {19:0,
11:0,
10:0,
7:0,
3:0,
0:0,
6:0,
2:1,
4:1,
13:1,
9:2,
1:2,
12:3,
14:3,
8:3,
17:4,
18:4,
5:4,
16:5,
15:5,
20:5}
train_copy['DWSSHY_bin'] = train_copy.DWSSHY.map(map_dict1)
from numpy import *
train_copy['GRZHZT_bin'] = pd.cut(train_copy.GRZHZT,bins=[-inf, 1, inf])
train_copy['DKLL_bin'] = pd.cut(train_copy.DKLL,bins=[-inf, 2.75, inf])

#%%测试集
test['DWJJLX_bin'] = test.DWJJLX.map(map_dict)
test['DWSSHY_bin'] = test.DWSSHY.map(map_dict1)
test['GRZHZT_bin'] = pd.cut(test.GRZHZT,bins=[-inf, 1, inf])
test['DKLL_bin'] = pd.cut(test.DKLL,bins=[-inf, 2.75, inf])

#%%数值型
num_col=['GRJCJS', 
       'GRZHYE', 'GRZHSNJZYE', 'GRZHDNGJYE', 'DWYJCE', 'DKFFE', 'DKYE','age','tiqu', 'DKFFE_GRJCJS', 'DKFFE_GRZHYE',
       'DKFFE_DWYJCE', 'DKFFE_DKYE', 'DKFFE_age', 'GRJCJS_age', 'GRZHYE_age']
train_copy[num_col].describe()

#%%构造变量
#提取金额 GRZHDNGJYE = -提取金额+DWYJCE*12
train_copy['tiqu'] = train_copy.DWYJCE*12*2-train_copy.GRZHDNGJYE
train_copy['tiqu'].max()

train_copy['DKFFE_GRJCJS'] = train_copy.DKFFE/train_copy.GRJCJS
train_copy['DKFFE_GRZHYE'] = train_copy.DKFFE/train_copy.GRZHYE
train_copy['DKFFE_DWYJCE'] = train_copy.DKFFE/train_copy.DWYJCE
train_copy['DKFFE_DKYE'] = train_copy.DKYE/train_copy.DKFFE
train_copy['DKFFE_age'] = train_copy.DKFFE/train_copy.age
train_copy['GRJCJS_age'] = train_copy.GRJCJS/train_copy.age
train_copy['GRZHYE_age'] = train_copy.GRZHYE/train_copy.age

#%%数值型变量的iv
num_col=['GRJCJS', 
       'GRZHYE', 'GRZHSNJZYE', 'GRZHDNGJYE', 'DWYJCE', 'DKFFE', 'DKYE','age','tiqu', 'DKFFE_DKYE']

num_iv_woedf = pc.WoeDf()
clf = pc.NumBin()
for i in num_col:
    clf.fit(train_copy[i] ,train_copy.label)
    clf.generate_transform_fun()
    num_iv_woedf.append(clf.woe_df_)
num_iv_woedf.to_excel('tmp25')


train_copy['GRJCJS_bin'] = pd.cut(train_copy.GRJCJS,bins=[-inf, 1197.25, 3745.75,  4489.25, 9801.75, 12446.0,  inf])
train_copy['GRZHYE_bin'] = pd.cut(train_copy.GRZHYE,bins=[-inf, 1324.0175, 5607.4651, 16688.3525, 25816.3701, 39361.7637, 65252.7266,  inf])
train_copy['GRZHSNJZYE_bin'] = pd.cut(train_copy.GRZHSNJZYE,bins=[-inf, 240.7625, 397.5475, 21479.625, 30042.8545,  inf])
train_copy['GRZHDNGJYE_bin'] = pd.cut(train_copy.GRZHDNGJYE,bins=[-inf, 204.94, 237.5275, 3905.64, 4846.5601, 6378.79,  inf])
train_copy['DWYJCE_bin'] = pd.cut(train_copy.DWYJCE,bins=[-inf, 351.595, 657.11,  739.3075, 1384.8, 1702.08,  inf])
train_copy['DKFFE_bin'] = pd.cut(train_copy.DKFFE,bins=[-inf, 129487.0, 168487.0, 250487.0, 339987.0,  inf])
train_copy['DKYE_bin'] = pd.cut(train_copy.DKYE,bins=[-inf, 237.0875, 40012.8418, 63379.4609, 113304.7031, 141377.0547, 150237.0547, 272994.5781, inf])
train_copy['age_bin'] = pd.cut(train_copy.age,bins=[-inf, 10.0, 29.5, 31.5, 32.5, 34.5, 35.5, 48.5, inf])


train_copy['tiqu_bin'] = pd.cut(train_copy.tiqu,bins=[-inf, 6500.6951, 9509.8101,  14111.0801,  15543.5, 22686.7305, inf])
train_copy['DKFFE_DKYE_bin'] = pd.cut(train_copy.DKFFE_DKYE,bins=[-inf, 0.5167, 0.7419, 0.8772, 0.9046, 0.929, 0.9385, 0.9727, inf])


train_copy = train_copy[['label','DWJJLX_bin', 'DWSSHY_bin',
       'GRZHZT_bin', 'DKLL_bin','GRJCJS_bin', 'GRZHYE_bin', 'GRZHSNJZYE_bin', 'GRZHDNGJYE_bin',
       'DWYJCE_bin','DKFFE_bin', 'DKYE_bin', 'age_bin', 'tiqu_bin',
       'DKFFE_DKYE_bin']]

#%%调箱后的iv 
iv_col = ['DWJJLX_bin', 'DWSSHY_bin',
       'GRZHZT_bin', 'DKLL_bin','GRJCJS_bin', 'GRZHYE_bin', 'GRZHSNJZYE_bin', 'GRZHDNGJYE_bin',
       'DWYJCE_bin', 'DKFFE_bin', 'DKYE_bin', 'age_bin', 'tiqu_bin',
       'DKFFE_DKYE_bin']

cate_iv_woedf = pc.WoeDf()
for i in iv_col:
    cate_iv_woedf.append(pc.cross_woe(train_copy[i] ,train_copy.label))
cate_iv_woedf.to_excel('tmp25')

len(iv_col)

#%%woe转换
cate_iv_woedf.bin2woe(train_copy,iv_col)
model_col = ['label','DWJJLX_woe',
       'DWSSHY_woe', 'GRZHZT_woe', 'DKLL_woe', 'GRJCJS_woe', 'GRZHYE_woe',
       'GRZHSNJZYE_woe', 'GRZHDNGJYE_woe', 'DWYJCE_woe', 'DKFFE_woe', 'DKYE_woe','age_woe', 'tiqu_woe',
       'DKFFE_DKYE_woe']


model_data = train_copy[model_col]
model_data.info()
model_data = model_data.astype(float64)


#%%建模
import pandas as pd
import matplotlib.pyplot as plt #导入图像库
import matplotlib
import seaborn as sns
import statsmodels.api as sm
from sklearn.metrics import roc_curve, auc,confusion_matrix
from sklearn.model_selection import train_test_split

X = model_data[['DWJJLX_woe',
       'DWSSHY_woe', 'GRZHZT_woe', 'DKLL_woe', 'GRJCJS_woe', 'GRZHYE_woe',
       'GRZHSNJZYE_woe', 'GRZHDNGJYE_woe', 'DWYJCE_woe', 'DKFFE_woe', 'DKYE_woe','age_woe', 'tiqu_woe',
       'DKFFE_DKYE_woe']]
Y = model_data['label']


x_train,x_test,y_train,y_test=train_test_split(X,Y,test_size=0.3,random_state=0)

#(10127, 44)

X1=sm.add_constant(x_train)   #在X前加上一列常数1，方便做带截距项的回归
logit=sm.Logit(y_train.astype(float),X1.astype(float))
result=logit.fit()
result.summary()
result.params


resu_1 = result.predict(X1.astype(float))
fpr, tpr, threshold = roc_curve(y_train, resu_1)
rocauc = auc(fpr, tpr)
plt.plot(fpr, tpr, 'b', label='AUC = %0.2f' % rocauc)
plt.legend(loc='lower right')
plt.plot([0, 1], [0, 1], 'r--')
plt.xlim([0, 1])
plt.ylim([0, 1])
plt.ylabel('真正率')
plt.xlabel('假正率')
plt.show()



X3 = sm.add_constant(x_test)
resu = result.predict(X3.astype(float))
fpr, tpr, threshold = roc_curve(y_test, resu)
rocauc = auc(fpr, tpr)
plt.plot(fpr, tpr, 'b', label='AUC = %0.2f' % rocauc)
plt.legend(loc='lower right')
plt.plot([0, 1], [0, 1], 'r--')
plt.xlim([0, 1])
plt.ylim([0, 1])
plt.ylabel('真正率')
plt.xlabel('假正率')
plt.show()

def tpr_weight_funtion(y_true,y_predict):
    d = pd.DataFrame()
    d['prob'] = list(y_predict)
    d['y'] = list(y_true)
    d = d.sort_values(['prob'], ascending=[0])
    y = d.y
    PosAll = pd.Series(y).value_counts()[1]
    NegAll = pd.Series(y).value_counts()[0]
    pCumsum = d['y'].cumsum()
    nCumsum = np.arange(len(y)) - pCumsum + 1
    pCumsumPer = pCumsum / PosAll
    nCumsumPer = nCumsum / NegAll
    TR1 = pCumsumPer[abs(nCumsumPer-0.001).idxmin()]
    TR2 = pCumsumPer[abs(nCumsumPer-0.005).idxmin()]
    TR3 = pCumsumPer[abs(nCumsumPer-0.01).idxmin()]
    return 0.4 * TR1 + 0.3 * TR2 + 0.3 * TR3
tpr_weight_funtion(y_train,resu_1)

然后是测试集

#%%测试集
test['tiqu'] = test.DWYJCE*12*2-test.GRZHDNGJYE
test['DKFFE_DKYE'] = test.DKYE/test.DKFFE
test['GRJCJS_bin'] = pd.cut(test.GRJCJS,bins=[-inf, 1197.25, 3745.75,  4489.25, 9801.75, 12446.0,  inf])
test['GRZHYE_bin'] = pd.cut(test.GRZHYE,bins=[-inf, 1324.0175, 5607.4651, 16688.3525, 25816.3701, 39361.7637, 65252.7266,  inf])
test['GRZHSNJZYE_bin'] = pd.cut(test.GRZHSNJZYE,bins=[-inf, 240.7625, 397.5475, 21479.625, 30042.8545,  inf])
test['GRZHDNGJYE_bin'] = pd.cut(test.GRZHDNGJYE,bins=[-inf, 204.94, 237.5275, 3905.64, 4846.5601, 6378.79,  inf])
test['DWYJCE_bin'] = pd.cut(test.DWYJCE,bins=[-inf, 351.595, 657.11,  739.3075, 1384.8, 1702.08,  inf])
test['DKFFE_bin'] = pd.cut(test.DKFFE,bins=[-inf, 129487.0, 168487.0, 250487.0, 339987.0,  inf])
test['DKYE_bin'] = pd.cut(test.DKYE,bins=[-inf, 237.0875, 40012.8418, 63379.4609, 113304.7031, 141377.0547, 150237.0547, 272994.5781, inf])
test['age_bin'] = pd.cut(test.age,bins=[-inf, 10.0, 29.5, 31.5, 32.5, 34.5, 35.5, 48.5, inf])


test['tiqu_bin'] = pd.cut(test.tiqu,bins=[-inf, 6500.6951, 9509.8101,  14111.0801,  15543.5, 22686.7305, inf])
test['DKFFE_DKYE_bin'] = pd.cut(test.DKFFE_DKYE,bins=[-inf, 0.5167, 0.7419, 0.8772, 0.9046, 0.929, 0.9385, 0.9727, inf])

#%%建模
cate_iv_woedf.bin2woe(test,iv_col)


model_col.remove('label')
test.isnull().sum()
model_data_1 = test[model_col]
model_data_1 = model_data_1.astype(float)


X4 = sm.add_constant(model_data_1)
resu = result.predict(X4.astype(float))

最终结果才0.305，我看auc还是不错的，但是结果就是，这样，最高的那位好像达到了0.615

 目前没有时间去做了，等过年再慢慢研究吧