WOE and IV are important concepts in credit risk analysis used to find out features which are relevant to predict whether or not a person is a possible loan defaulter. Lets say I have a person with some variables (could be age, income, past credit score etc., and I need to predict based on past data which ppl are likely to default. Now I have 1000 variables. All the variables will not be of equal importance while predicting loan default. So WOE and IV help me in finding out which variable out of all the variables should I use in any model. Below is a sample df and a generic function to find WOE and IV for any df. My doubt is stated below.
X_train (independent variables in credit risk analysis)
Var1 Var2 ............. Var1000
30 Unknown ............. 80000
33 Success ............. 90000
45 Failure ............. 899900
y_train (dependent variable in credit risk analysis to denote customer default or not in a loan)
0
1
0
0...and so on
Code to find the most relevant and important variables from a dataframe (WOE and IV)
import pandas as pd
import numpy as np
import pandas.core.algorithms as algos
from pandas import Series
import scipy.stats.stats as stats
import re
import traceback
import string
max_bin = 20
force_bin = 3
# define a binning function
def mono_bin(Y, X, n = max_bin):
df1 = pd.DataFrame({"X": X, "Y": Y})
justmiss = df1[['X','Y']][df1.X.isnull()]
notmiss = df1[['X','Y']][df1.X.notnull()]
r = 0
while np.abs(r) < 1:
try:
d1 = pd.DataFrame({"X": notmiss.X, "Y": notmiss.Y, "Bucket": pd.qcut(notmiss.X, n)})
d2 = d1.groupby('Bucket', as_index=True)
r, p = stats.spearmanr(d2.mean().X, d2.mean().Y)
n = n - 1
except Exception as e:
n = n - 1
if len(d2) == 1:
n = force_bin
bins = algos.quantile(notmiss.X, np.linspace(0, 1, n))
if len(np.unique(bins)) == 2:
bins = np.insert(bins, 0, 1)
bins[1] = bins[1]-(bins[1]/2)
d1 = pd.DataFrame({"X": notmiss.X, "Y": notmiss.Y, "Bucket": pd.cut(notmiss.X, np.unique(bins),include_lowest=True)})
d2 = d1.groupby('Bucket', as_index=True)
d3 = pd.DataFrame({},index=[])
d3["MIN_VALUE"] = d2.min().X
d3["MAX_VALUE"] = d2.max().X
d3["COUNT"] = d2.count().Y
d3["EVENT"] = d2.sum().Y
d3["NONEVENT"] = d2.count().Y - d2.sum().Y
d3=d3.reset_index(drop=True)
if len(justmiss.index) > 0:
d4 = pd.DataFrame({'MIN_VALUE':np.nan},index=[0])
d4["MAX_VALUE"] = np.nan
d4["COUNT"] = justmiss.count().Y
d4["EVENT"] = justmiss.sum().Y
d4["NONEVENT"] = justmiss.count().Y - justmiss.sum().Y
d3 = d3.append(d4,ignore_index=True)
d3["EVENT_RATE"] = d3.EVENT/d3.COUNT
d3["NON_EVENT_RATE"] = d3.NONEVENT/d3.COUNT
d3["DIST_EVENT"] = d3.EVENT/d3.sum().EVENT
d3["DIST_NON_EVENT"] = d3.NONEVENT/d3.sum().NONEVENT
d3["WOE"] = np.log(d3.DIST_EVENT/d3.DIST_NON_EVENT)
d3["IV"] = (d3.DIST_EVENT-d3.DIST_NON_EVENT)*np.log(d3.DIST_EVENT/d3.DIST_NON_EVENT)
d3["VAR_NAME"] = "VAR"
d3 = d3[['VAR_NAME','MIN_VALUE', 'MAX_VALUE', 'COUNT', 'EVENT', 'EVENT_RATE', 'NONEVENT', 'NON_EVENT_RATE', 'DIST_EVENT','DIST_NON_EVENT','WOE', 'IV']]
d3 = d3.replace([np.inf, -np.inf], 0)
d3.IV = d3.IV.sum()
return(d3)
def char_bin(Y, X):
df1 = pd.DataFrame({"X": X, "Y": Y})
justmiss = df1[['X','Y']][df1.X.isnull()]
notmiss = df1[['X','Y']][df1.X.notnull()]
df2 = notmiss.groupby('X',as_index=True)
d3 = pd.DataFrame({},index=[])
d3["COUNT"] = df2.count().Y
d3["MIN_VALUE"] = df2.sum().Y.index
d3["MAX_VALUE"] = d3["MIN_VALUE"]
d3["EVENT"] = df2.sum().Y
d3["NONEVENT"] = df2.count().Y - df2.sum().Y
if len(justmiss.index) > 0:
d4 = pd.DataFrame({'MIN_VALUE':np.nan},index=[0])
d4["MAX_VALUE"] = np.nan
d4["COUNT"] = justmiss.count().Y
d4["EVENT"] = justmiss.sum().Y
d4["NONEVENT"] = justmiss.count().Y - justmiss.sum().Y
d3 = d3.append(d4,ignore_index=True)
d3["EVENT_RATE"] = d3.EVENT/d3.COUNT
d3["NON_EVENT_RATE"] = d3.NONEVENT/d3.COUNT
d3["DIST_EVENT"] = d3.EVENT/d3.sum().EVENT
d3["DIST_NON_EVENT"] = d3.NONEVENT/d3.sum().NONEVENT
d3["WOE"] = np.log(d3.DIST_EVENT/d3.DIST_NON_EVENT)
d3["IV"] = (d3.DIST_EVENT-d3.DIST_NON_EVENT)*np.log(d3.DIST_EVENT/d3.DIST_NON_EVENT)
d3["VAR_NAME"] = "VAR"
d3 = d3[['VAR_NAME','MIN_VALUE', 'MAX_VALUE', 'COUNT', 'EVENT', 'EVENT_RATE', 'NONEVENT', 'NON_EVENT_RATE', 'DIST_EVENT','DIST_NON_EVENT','WOE', 'IV']]
d3 = d3.replace([np.inf, -np.inf], 0)
d3.IV = d3.IV.sum()
d3 = d3.reset_index(drop=True)
return(d3)
def data_vars(df1, target):
stack = traceback.extract_stack()
filename, lineno, function_name, code = stack[-2]
vars_name = re.compile(r'\((.*?)\).*$').search(code).groups()[0]
final = (re.findall(r"[\w']+", vars_name))[-1]
x = df1.dtypes.index
count = -1
for i in x:
if i.upper() not in (final.upper()):
if np.issubdtype(df1[i], np.number) and len(Series.unique(df1[i])) > 2:
conv = mono_bin(target, df1[i])
conv["VAR_NAME"] = i
count = count + 1
else:
conv = char_bin(target, df1[i])
conv["VAR_NAME"] = i
count = count + 1
if count == 0:
iv_df = conv
else:
iv_df = iv_df.append(conv,ignore_index=True)
iv = pd.DataFrame({'IV':iv_df.groupby('VAR_NAME').IV.max()})
iv = iv.reset_index()
return(iv_df,iv)
Calling function
final_iv, IV = data_vars(X_train, y_train)
Output
VAR_NAME MIN_VALUE MAX_VALUE COUNT EVENT EVENT_RATE NONEVENT NON_EVENT_RATE DIST_EVENT DIST_NON_EVENT WOE IV
0 Var1 19 39 2290 259 0.113100 2031 0.886900 0.497121 0.50775 -0.021156 0.000452
1 Var1 40 87 2231 262 0.117436 1969 0.882564 0.502879 0.49225 0.021363 0.000452
64 Var2 failure failure 490 63 0.128571 427 0.871429 0.120921 0.10675 0.124650 0.461890
65 Var2 other other 197 38 0.192893 159 0.807107 0.072937 0.03975 0.606982 0.461890
66 Var2 success success 129 83 0.643411 46 0.356589 0.159309 0.01150 2.628499 0.461890
67 Var2 unknown unknown 3705 337 0.090958 3368 0.909042 0.646833 0.84200 -0.263692 0.461890
and so on....
If you see there are 2 values for the same variable Var1
and 4 values for Var2
variable. I am unable to understand what is the meaning of multiple rows for the same variable.
What is this function data_vars trying to do?
The results are binned, probably under the assumption that some features can have different WOE/IV scores across their distribution. For example, Var1 has different WOE scores for [19,39] than [40,87]. Whether this binning makes sense for categorical variables, like Var2, is a different story.