简介
本文通过使用LendingClub的数据,采用卡方分箱(ChiMerge)、WOE编码、计算IV值、单变量和多变量(VIF)分析,然后使用逻辑回归模型进行训练,在变量筛选时也可尝试添加L1约束或通过随机森林筛选变量,最后进行模型评估。
关键词:卡方分箱,WOE,IV值,变量分析,逻辑回归
一、数据预处理
数据清洗:数据选择、格式转换、缺失值填补
由于贷款期限(term)有多个种类,申请评分卡模型评估的违约概率必须在统一的期限中,并且不宜太长,因此选择36months的数据作为本次建模数据,60%训练,40%测试。
folderOfData = os.path.join(os.getcwd(), 'data')
allData = pd.read_csv(os.path.join(folderOfData,'application.csv'),header = 0, encoding = 'latin1')
allData['term'] = allData['term'].apply(lambda x: int(x.replace(' months','')))
# 处理标签:Fully Paid是正常用户;Charged Off是违约用户
allData['y'] = allData['loan_status'].map(lambda x: int(x == 'Charged Off'))
allData1 = allData.loc[allData.term == 36]
trainData, testData = train_test_split(allData1,test_size=0.4)
进一步清洗:
- 将int_rate利息转换为小数形式
- 将emp_length处理为:10+为11,<1为0,空为-1
- desc为有记录和无记录两种情况
- 日期处理
- 两个日期之间月数计算
# 将带%的百分比变为浮点数
trainData['int_rate_clean'] = trainData['int_rate'].map(lambda x: float(x.replace('%',''))/100)
# 将工作年限进行转化,否则影响排序
trainData['emp_length_clean'] = trainData['emp_length'].map(CareerYear)
# 将desc的缺失作为一种状态,非缺失作为另一种状态
trainData['desc_clean'] = trainData['desc'].map(DescExisting)
# 处理日期。earliest_cr_line的格式不统一,需要统一格式且转换成python的日期
trainData['app_date_clean'] = trainData['issue_d'].map(lambda x: ConvertDateStr(x))
trainData['earliest_cr_line_clean'] = trainData['earliest_cr_line'].map(lambda x: ConvertDateStr(x))
# 处理mths_since_last_delinq。注意原始值中有0,所以用-1代替缺失
trainData['mths_since_last_delinq_clean'] = trainData['mths_since_last_delinq'].map(lambda x:MakeupMissing(x))
trainData['mths_since_last_record_clean'] = trainData['mths_since_last_record'].map(lambda x:MakeupMissing(x))
trainData['pub_rec_bankruptcies_clean'] = trainData['pub_rec_bankruptcies'].map(lambda x:MakeupMissing(x))
二、变量衍生和挑选
- 衍生:
- 考虑申请额度与收入的占比
- 考虑earliest_cr_line到申请日期的跨度,计算月份数
# 考虑申请额度与收入的占比
trainData['limit_income'] = trainData.apply(lambda x: x.loan_amnt / x.annual_inc, axis = 1)
# 考虑earliest_cr_line到申请日期的跨度,计算月份数
trainData['earliest_cr_to_app'] = trainData.apply(lambda x: MonthGap(x.earliest_cr_line_clean,x.app_date_clean), axis = 1)
- 挑选:
我们初步挑选变量如下,分为两类:数值型(连续型)的和类别型的变量。
num_features = ['int_rate_clean','emp_length_clean','annual_inc', 'dti', 'delinq_2yrs', 'earliest_cr_to_app','inq_last_6mths', \
'mths_since_last_record_clean', 'mths_since_last_delinq_clean','open_acc','pub_rec','total_acc','limit_income','earliest_cr_to_app']
cat_features = ['home_ownership', 'verification_status','desc_clean', 'purpose', 'zip_code','addr_state','pub_rec_bankruptcies_clean']
三、卡方分箱法
采用卡方(ChiMerge)分箱,要求分箱完成之后:
- 不超过5箱(本模型默认不超过5箱)
- 坏样本率(Bad Rate)单调
- 每箱同时包含好坏样本
- 如有特殊值如-1单独成一箱,此箱不参与Bad Rate单调性检验
连续型的变量可以直接进行分箱,对于类别型的变量分为以下几种情况:
- 当类别型变量取值比较多时(本例中大于5),先用bad rate 进行编码,然后放入连续型变量列表中,使用连续型变量分箱的方法进行分箱。
- 当取值较少时(本例中小于等于5),分两种情况:
(1)如果每种类别同时包含好坏样本,则无需分箱;
(2)如果有类别只包含好坏样本的一种,则需要合并;
具体操作如下:
第一步,检查类别型变量中,哪些变量取值超过5。
more_value_features = []
less_value_features = []
# 第一步,检查类别型变量中,哪些变量取值超过5
for var in cat_features:
valueCounts = len(set(trainData[var]))
print valueCounts
if valueCounts > 5:
more_value_features.append(var) #取值超过5的变量,需要bad rate编码,再用卡方分箱法进行分箱
else:
less_value_features.append(var)
第二步,当取值<5时:如果每种类别同时包含好坏样本,无需分箱;如果有类别只包含好坏样本的一种,需要合并。
merge_bin_dict = {} #存放需要合并的变量,以及合并方法
var_bin_list = [] #由于某个取值没有好或者坏样本而需要合并的变量
for col in less_value_features:
binBadRate = BinBadRate(trainData, col, 'y')[0]
if min(binBadRate.values()) == 0 : #由于某个取值没有坏样本而进行合并
print '{} need to be combined due to 0 bad rate'.format(col)
combine_bin = MergeBad0(trainData, col, 'y')
merge_bin_dict[col] = combine_bin
newVar = col + '_Bin'
trainData[newVar] = trainData[col].map(combine_bin)
var_bin_list.append(newVar)
if max(binBadRate.values()) == 1: #由于某个取值没有好样本而进行合并
print '{} need to be combined due to 0 good rate'.format(col)
combine_bin = MergeBad0(trainData, col, 'y',direction = 'good')
merge_bin_dict[col] = combine_bin
newVar = col + '_Bin'
trainData[newVar] = trainData[col].map(combine_bin)
var_bin_list.append(newVar)
第三步,当取值>5时:用bad rate进行编码,放入连续型变量里。
br_encoding_dict = {} #记录按照bad rate进行编码的变量,及编码方式
for col in more_value_features:
br_encoding = BadRateEncoding(trainData, col, 'y')
trainData[col+'_br_encoding'] = br_encoding['encoding']
br_encoding_dict[col] = br_encoding['bad_rate']
num_features.append(col+'_br_encoding')
第四步,分箱,对连续型变量列表num_features进行卡方分箱。本文分箱后的最多的箱数为5箱。
continous_merged_dict = {}
for col in num_features:
max_interval = 5 # 分箱后的最多的箱数
print "{} is in processing".format(col)
if -1 not in set(trainData[col]): #-1会当成特殊值处理。如果没有-1,则所有取值都参与分箱
cutOff = ChiMerge(trainData, col, 'y', max_interval=max_interval,special_attribute=[],minBinPcnt=0)
trainData[col+'_Bin'] = trainData[col].map(lambda x: AssignBin(x, cutOff,special_attribute=[]))
monotone = BadRateMonotone(trainData, col+'_Bin', 'y') # 检验分箱后的单调性是否满足
while(not monotone):
# 检验分箱后的单调性是否满足。如果不满足,则缩减分箱的个数。
max_interval -= 1
cutOff = ChiMerge(trainData, col, 'y', max_interval=max_interval, special_attribute=[],
minBinPcnt=0)
trainData[col + '_Bin'] = trainData[col].map(lambda x: AssignBin(x, cutOff, special_attribute=[]))
if max_interval == 2:
# 当分箱数为2时,必然单调
break
monotone = BadRateMonotone(trainData, col + '_Bin', 'y')
newVar = col + '_Bin'
trainData[newVar] = trainData[col].map(lambda x: AssignBin(x, cutOff, special_attribute=[]))
var_bin_list.append(newVar)
else:
# 如果有-1,则除去-1后,其他取值参与分箱
cutOff = ChiMerge(trainData, col, 'y', max_interval=max_interval, special_attribute=[-1],
minBinPcnt=0)
trainData[col + '_Bin'] = trainData[col].map(lambda x: AssignBin(x, cutOff, special_attribute=[-1]))
monotone = BadRateMonotone(trainData, col + '_Bin', 'y',['Bin -1'])
while (not monotone):
max_interval -= 1
# 如果有-1,-1的bad rate不参与单调性检验
cutOff = ChiMerge(trainData, col, 'y', max_interval=max_interval, special_attribute=[-1],
minBinPcnt=0)
trainData[col + '_Bin'] = trainData[col].map(lambda x: AssignBin(x, cutOff, special_attribute=[-1]))
if max_interval == 3:
# 考虑特殊值,当分箱数为3-1=2时,必然单调
break
monotone = BadRateMonotone(trainData, col + '_Bin', 'y',['Bin -1'])
newVar = col + '_Bin'
trainData[newVar] = trainData[col].map(lambda x: AssignBin(x, cutOff, special_attribute=[-1]))
var_bin_list.append(newVar)
continous_merged_dict[col] = cutOff
四、WOE编码和IV值
经常上一步的分箱后,分箱后的变量有如下几种情况:
- 初始取值个数小于5,且不需要合并的类别型变量。
- 初始取值个数小于5,需要合并的类别型变量,并且合并后的新变量不再需要合并。
- 初始取值个数超过5,需要合并的类别型变量,并且合并后的新变量不再需要合并。
- 连续型变量进行卡方分箱。
如下取到每个变量分箱后的WOE和该变量的IV值:
WOE_dict = {}
IV_dict = {}
for var in all_var:
woe_iv = CalcWOE(trainData, var, 'y')
WOE_dict[var] = woe_iv['WOE']
IV_dict[var] = woe_iv['IV']
将变量IV值进行降序排列,得到结果如下:
IV_dict_sorted = sorted(IV_dict.items(), key=lambda x: x[1], reverse=True)
IV_values = [i[1] for i in IV_dict_sorted]
IV_name = [i[0] for i in IV_dict_sorted]
plt.title('feature IV')
plt.bar(range(len(IV_values)),IV_values)
得到的IV值如下图所示:
image.png
五、变量分析
单变量分析和多变量分析,均基于WOE编码后的值。
- 选择IV值大于等于0.01的变量
- 比较两两线性相关性。如果相关系数的绝对值高于阈值,剔除IV较低的一个。
#选取IV>=0.01的变量
high_IV = {k:v for k, v in IV_dict.items() if v >= 0.01}
high_IV_sorted = sorted(high_IV.items(),key=lambda x:x[1],reverse=True)
short_list = high_IV.keys()
short_list_2 = []
for var in short_list:
newVar = var + '_WOE'
trainData[newVar] = trainData[var].map(WOE_dict[var])
short_list_2.append(newVar)
#对于上一步的结果,计算相关系数矩阵,并画出热力图进行数据可视化
trainDataWOE = trainData[short_list_2]
f, ax = plt.subplots(figsize=(10, 8))
corr = trainDataWOE.corr()
sns.heatmap(corr, mask=np.zeros_like(corr, dtype=np.bool), cmap=sns.diverging_palette(220, 10, as_cmap=True),square=True, ax=ax)
f.savefig('sns_heatmap_high_IV.png')
根据IV值挑选的变量的相关系数矩阵热力图:
image.png
单变量两两间的线性相关性检验:
(1)将候选变量按照IV进行降序排列
(2)计算第i和第i+1的变量的线性相关系数
(3)对于系数超过阈值的两个变量,剔除IV较低的一个
此处阈值为0.7,大于0.7则表示有相关性。见如下代码:
deleted_index = []
cnt_vars = len(high_IV_sorted)
for i in range(cnt_vars):
if i in deleted_index:
continue
x1 = high_IV_sorted[i][0]+"_WOE"
for j in range(cnt_vars):
if i == j or j in deleted_index:
continue
y1 = high_IV_sorted[j][0]+"_WOE"
roh = np.corrcoef(trainData[x1],trainData[y1])[0,1]
if abs(roh)>0.7:
x1_IV = high_IV_sorted[i][1]
y1_IV = high_IV_sorted[j][1]
if x1_IV > y1_IV:
deleted_index.append(j)
else:
deleted_index.append(i)
multi_analysis_vars_1 = [high_IV_sorted[i][0]+"_WOE" for i in range(cnt_vars) if i not in deleted_index]
多变量分析:VIF
一般要小于10,本次结果max_VIF为:1.5093709849027372,则多变量之间排除共线性。
X = np.matrix(trainData[multi_analysis_vars_1])
VIF_list = [variance_inflation_factor(X, i) for i in range(X.shape[1])]
max_VIF = max(VIF_list)
print max_VIF
六、逻辑回归模型
要求:
1,变量显著
2,符号为负
将多变量分析后的变量带入LR模型中,
y = trainData['y']
X = trainData[multi_analysis]
X['intercept'] = [1]*X.shape[0]
LR = sm.Logit(y, X).fit()
summary = LR.summary()
pvals = LR.pvalues
pvals = pvals.to_dict()
逐步剔除p值不显著的变量
varLargeP = {k: v for k,v in pvals.items() if v >= 0.1}
varLargeP = sorted(varLargeP.items(), key=lambda d:d[1], reverse = True)
while(len(varLargeP) > 0 and len(multi_analysis) > 0):
# 每次迭代中,剔除最不显著的变量,直到
# (1) 剩余所有变量均显著
# (2) 没有特征可选
varMaxP = varLargeP[0][0]
print varMaxP
if varMaxP == 'intercept':
print 'the intercept is not significant!'
break
multi_analysis.remove(varMaxP)
y = trainData['y']
X = trainData[multi_analysis]
X['intercept'] = [1] * X.shape[0]
LR = sm.Logit(y, X).fit()
pvals = LR.pvalues
pvals = pvals.to_dict()
varLargeP = {k: v for k, v in pvals.items() if v >= 0.1}
varLargeP = sorted(varLargeP.iteritems(), key=lambda d: d[1], reverse=True)
summary = LR.summary()
逻辑回归结果如下:
LLR p-value: 2.460e-280
========================================================================================================
coef std err z P>|z| [0.025 0.975]
--------------------------------------------------------------------------------------------------------
zip_code_br_encoding_Bin_WOE -0.9467 0.045 -21.258 0.000 -1.034 -0.859
int_rate_clean_Bin_WOE -0.8742 0.055 -15.779 0.000 -0.983 -0.766
annual_inc_Bin_WOE -0.7039 0.095 -7.383 0.000 -0.891 -0.517
purpose_br_encoding_Bin_WOE -0.8559 0.087 -9.785 0.000 -1.027 -0.684
inq_last_6mths_Bin_WOE -0.7831 0.104 -7.537 0.000 -0.987 -0.579
addr_state_br_encoding_Bin_WOE -0.2423 0.121 -1.997 0.046 -0.480 -0.005
limit_income_Bin_WOE -0.4409 0.134 -3.299 0.001 -0.703 -0.179
mths_since_last_record_clean_Bin_WOE -0.7616 0.141 -5.416 0.000 -1.037 -0.486
total_acc_Bin_WOE -0.2963 0.173 -1.710 0.087 -0.636 0.043
dti_Bin_WOE -0.7897 0.196 -4.021 0.000 -1.175 -0.405
emp_length_clean_Bin_WOE -0.7229 0.200 -3.611 0.000 -1.115 -0.331
intercept -2.1014 0.027 -78.645 0.000 -2.154 -2.049
========================================================================================================
可以看到p值均显著,且系数为负。
计算auc值,结果为:0.74
trainData['prob'] = LR.predict(X)
auc = roc_auc_score(trainData['y'],trainData['prob']) #AUC = 0.73
七、验证模型
用同样的方法,对验证集数据进行处理后,放入模型,如下得到
auc=0.65
ks = 0.22
表明模型有一定的预测能力和区分度
testData['intercept'] = [1]*testData.shape[0]
#预测数据集中,变量顺序需要和LR模型的变量顺序一致
#例如在训练集里,变量在数据中的顺序是“负债比”在“借款目的”之前,对应地,在测试集里,“负债比”也要在“借款目的”之前
testData2 = testData[list(LR.params.index)]
testData['prob'] = LR.predict(testData2)
#计算KS和AUC
auc = roc_auc_score(testData['y'],testData['prob'])
ks = KS(testData, 'prob', 'y')
计算评分:
basePoint = 250
PDO = 200
testData['score'] = testData['prob'].map(lambda x:Prob2Score(x, basePoint, PDO))
testData = testData.sort_values(by = 'score')
结果如下,分值与频数的分布近似为正态分布。根据业务需要以及相应的风险比例,划分评分区间,合理应用评分卡模型。
image.png
网友评论