【数据分析实战】金融评分卡建立

一、导入数据

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%config InlineBackend.figure_format = 'svg' 
import toad
from toad.plot import bin_plot, badrate_plot
import math
from imblearn.over_sampling import SMOTE, RandomOverSampler
import seaborn as sns
sns.set()

# 数据来源：kaggle项目"give me some credit"
credit_df0 = pd.read_csv('data/GiveMeSomeCredit/cs-training.csv')

# 查看数据集
credit_df0.head()

# 查看描述性统计信息
toad.detect(credit_df0)

列名含义：

• SeriousDlqin2yrs：超过90天或更糟的逾期拖欠
• RevolvingUtilizationOfUnsecuredLines：除了房贷车贷之外的信用卡账面金额（即贷款金额）/信用卡总额度
• age：贷款人年龄
• NumberOfTime30-59DaysPastDueNotWorse：35-59天逾期但不糟糕次数
• DebtRatio：负债比率
• MonthlyIncome：月收入
• NumberOfOpenCreditLinesAndLoans：开放式信贷和贷款数量，开放式贷款（分期付款如汽车贷款或抵押贷款）和信贷（如信用卡）的数量
• NumberOfTimes90DaysLate：借款者有90天或更高逾期的次数
• NumberRealEstateLoansOrLines：不动产贷款或额度数量
• NumberOfTime60-89DaysPastDueNotWorse：60-89天逾期但不糟糕次数
• NumberOfDependents：不包括本人在内的家属数量
  
  

二、EDA

# 丢弃编号列
credit_df1 = credit_df0.drop(['Unnamed: 0'], axis=1)
# 修改列名
colnames={
   'SeriousDlqin2yrs':'Isdlq',
          'age':'Age',
          'RevolvingUtilizationOfUnsecuredLines':'Revol',
          'NumberOfTime30-59DaysPastDueNotWorse':'Num30-59late',
          'NumberOfOpenCreditLinesAndLoans':'Numopen',
          'NumberOfTimes90DaysLate':'Num90late',
          'NumberRealEstateLoansOrLines':'Numestate',
          'NumberOfTime60-89DaysPastDueNotWorse':'Num60-89late',
          'NumberOfDependents':'Numdepend'}
credit_df1.rename(columns=colnames, inplace=True)
credit_df1.head()

# 查看好坏比
sns.countplot(credit_df1['Isdlq'])
print(f"好坏比：{
     np.round(100 * credit_df1['Isdlq'].mean(), 2)}%")  
# 好坏比：6.68%

2.1 查看Revol特征

# 查看可用额度比的描述性统计信息
credit_df1['Revol'].describe([0.99, 0.999])
"""
count    150000.000000
mean          6.048438
std         249.755371
min           0.000000
50%           0.154181
99%           1.092956
99.9%      1571.006000
max       50708.000000
Name: Revol, dtype: float64
"""

明显分布异常

# 画出Revol小于1的分布图
sns.distplot(credit_df1[credit_df1['Revol']<1]['Revol'], bins=10)

# 定义一个分箱并统计箱内坏客户率的函数
def show_rate_by_box(df, target_name, feature_name, bins):
    temp = pd.concat([df[target_name], pd.cut(df[feature_name], bins=bins, right=False)], axis=1)
    return pd.pivot_table(temp, index=[feature_name], values=[target_name], aggfunc=['mean', 'count'])

按理说Revol不应该大于1，所以我们重点查看大于1的数据违约率如何

# 初步分箱并查看各区间段的违约率分布，给后续分箱提供参考
revol_bins=[0,0.5,1,1.5,2,5,10,20,30,40,50,100,1000,5000,math.inf]
temp = show_rate_by_box(credit_df1, 'Isdlq', 'Revol', bins=revol_bins)
show_rate_by_box(credit_df1, 'Isdlq', 'Revol', bins=revol_bins)

# 画成图方便观看
plt.figure(figsize=(15, 5))
sns.barplot(x=temp.index, y=temp[( 'mean', 'Isdlq')])

结论：1到20坏客户比率明显上升，20以上又降下来，将异常值阈值确定为20。高于20后续统一删除。

2.2 Age

# 查看年龄的描述性统计信息
credit_df1['Age'].describe([0.01])
"""
count    150000.000000
mean         52.295207
std          14.771866
min           0.000000
1%           24.000000
50%          52.000000
max         109.000000
Name: Age, dtype: float64
"""

年龄小于18岁不符合业务逻辑，后续准备统一排除

# 查看要删除的有几人
len(credit_df1[credit_df1['Age']<18])
# 1

# 画出分布图
sns.distplot(credit_df1['Age'])

2.3 DebtRatio

# 查看负债率的描述性统计信息
credit_df1['DebtRatio'].describe([0.01, 0.99, 0.999])
"""
count    150000.000000
mean        353.005076
std        2037.818523
min           0.000000
1%            0.000000
50%           0.366508
99%        4979.040000
99.9%     10613.074000
max      329664.000000
Name: DebtRatio, dtype: float64
"""

# 画图查看分布
sns.distplot(credit_df1[</