In finance is a common practice to create risk scorecards to assess the credit worthiness for a given customer. Unfortunately, out of the box credit scoring tools are quite expensive and scatter, that's why we created this toolkit: to empower all credit scoring practicioners and spread the use of weight of evidence based scoring techniques for alternative uses cases (virtually any binary classification problem).
Explore the documentation»
Report Bug
Table of Contents
The general process for creating Weight of Evidence based scorecards is illustrated in the figure below :
For that matter, we implemented the following classes to address the necesary steps to perform credit scoring transformation:
Class for normalizing discrete data for a given relative frequency threshold
Class for discretizing continuous data into bins using several methods
Class for encoding discrete features into Weight of Evidence(WoE) transformation
Base class for selecting features based on their WoE transformation and Information Value statistic.
Class for selecting continuous features based on their WoE transformation and Information Value statistic.
Class for selecting discrete features based on their WoE transformation and Information Value statistic.
Implements credit risk scorecards following the methodology proposed in Siddiqi, N. (2012). Credit risk scorecards: developing and implementing intelligent credit scoring (Vol. 3). John Wiley & Sons.
You can simply install the module using pip
- pip
pip install woe-credit-scoring
import pandas as pd
from CreditScoringToolkit.frequency_table import frequency_table
from CreditScoringToolkit.DiscreteNormalizer import DiscreteNormalizer
from CreditScoringToolkit.WoeEncoder import WoeEncoder
from CreditScoringToolkit.WoeContinuousFeatureSelector import WoeContinuousFeatureSelector
from CreditScoringToolkit.WoeDiscreteFeatureSelector import WoeDiscreteFeatureSelector
from CreditScoringToolkit.CreditScoring import CreditScoring
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score
import matplotlib.pyplot as plt
import seaborn as sns
import warnings
warnings.filterwarnings('ignore')# Read example data for train and validation (loan applications)
train = pd.read_csv('train.csv')
valid = pd.read_csv('valid.csv') # Assign features lists by type, file contains "C_" prefix for continuous, and "D_" for discrete.
vard = [v for v in train.columns if v[:2]=='D_']
varc = [v for v in train.columns if v[:2]=='C_']# In this example, we aggregate categories with less than 10% of relative frequency
# into a new category called 'SMALL CATEGORIES', if new created category don't reach
# given relative frequency threshold (10%) then the most frequent category is imputed.
# All missing values are treatead as the separate category MISSING
dn = DiscreteNormalizer(normalization_threshold=0.1,default_category='SMALL CATEGORIES')
dn.fit(train[vard])
Xt = dn.transform(train[vard])
frequency_table(Xt,vard)****Frequency Table D_OCCUPATION_TYPE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
Laborers 166 0.166 166 0.166
MISSING 325 0.325 491 0.491
SMALL CATEGORIES 395 0.395 886 0.886
Sales staff 114 0.114 1000 1.000
****Frequency Table D_NAME_CONTRACT_TYPE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
Cash loans 897 0.897 897 0.897
Revolving loans 103 0.103 1000 1.000
****Frequency Table D_CODE_GENDER ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
F 659 0.659 659 0.659
M 341 0.341 1000 1.000
****Frequency Table D_FLAG_OWN_CAR ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
N 668 0.668 668 0.668
Y 332 0.332 1000 1.000
****Frequency Table D_FLAG_OWN_REALTY ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
N 287 0.287 287 0.287
Y 713 0.713 1000 1.000
****Frequency Table D_NAME_INCOME_TYPE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. \
Commercial associate 225 0.225 225
Pensioner 179 0.179 404
Working 596 0.596 1000
Cumm. Rel. Freq.
Commercial associate 0.225
Pensioner 0.404
Working 1.000
****Frequency Table D_NAME_EDUCATION_TYPE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. \
Higher education 243 0.243 243
Secondary / secondary special 757 0.757 1000
Cumm. Rel. Freq.
Higher education 0.243
Secondary / secondary special 1.000
****Frequency Table D_NAME_FAMILY_STATUS ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. \
Civil marriage 102 0.102 102
Married 620 0.620 722
SMALL CATEGORIES 117 0.117 839
Single / not married 161 0.161 1000
Cumm. Rel. Freq.
Civil marriage 0.102
Married 0.722
SMALL CATEGORIES 0.839
Single / not married 1.000
****Frequency Table D_NAME_HOUSING_TYPE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
House / apartment 878 0.878 878 0.878
SMALL CATEGORIES 122 0.122 1000 1.000
****Frequency Table D_FLAG_PHONE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
0 721 0.721 721 0.721
1 279 0.279 1000 1.000
****Frequency Table D_WEEKDAY_APPR_PROCESS_START ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
FRIDAY 145 0.145 145 0.145
MONDAY 170 0.170 315 0.315
SATURDAY 106 0.106 421 0.421
THURSDAY 235 0.235 656 0.656
TUESDAY 177 0.177 833 0.833
WEDNESDAY 167 0.167 1000 1.000
****Frequency Table D_NAME_TYPE_SUITE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
Family 126 0.126 126 0.126
Unaccompanied 874 0.874 1000 1.000
****Frequency Table D_HOUSETYPE_MODE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
MISSING 490 0.49 490 0.49
block of flats 510 0.51 1000 1.00
****Frequency Table D_WALLSMATERIAL_MODE ***
Abs. Freq. Rel. Freq. Cumm. Abs. Freq. Cumm. Rel. Freq.
MISSING 560 0.560 560 0.560
Panel 227 0.227 787 0.787
Stone, brick 213 0.213 1000 1.000
unary = [v for v in vard if Xt[v].nunique==1]
unary[]
# Now we proceed with feature selection, we have special classes for each type of feature (discrete,continuous)
# Discrete feature selector uses the given iv_threshold to select the best features only.
# For continuous feature selector, a variety of methods are available for selecting the best features , namely:
# -uniform: only uses equal-width discretized bins, selects number of bins with best IV value.
# -quantile: only uses equal-frequency discretized bins, selects number of bins with best IV value
# -kmeans: only uses discretized bins created by a K-Means clustering, selects number of bins with best IV value
# -gaussian: only uses discretized bins created by a Gaussian Mixture, selects number of bins with best IV value
# -dcc: stands for Discrete Competitive Combination, creates segments for all individual methods and then
# selects the best method and its corresponding best number of bins for each feature.
# -dec: stands for Discrete Exhaustive Combination, creates segments for all individual methods and then
# selects the best number of bins for each feature including every feasible method.
#
# One can configure IV threshold, minimun/maximum number of discretization bins, whether or not to keep only
# strictly monotonic segments and the number of pooling threads used in order to speed computations.
Xt = pd.concat([Xt,train[varc]],axis=1) #Merge continuous features matrix with the normalized discrete predictors Matrix
wcf = WoeContinuousFeatureSelector()
wdf = WoeDiscreteFeatureSelector()
# Perform feature selection
wcf.fit(Xt[varc],train['TARGET'],
max_bins=6,
strictly_monotonic=True,
iv_threshold=0.05,
method='dcc',
n_threads=20)
wdf.fit(Xt[vard],train['TARGET'],iv_threshold=0.1)
# Create new matrix with discrete and discretized best features
Xt = pd.concat([wdf.transform(Xt[vard]),wcf.transform(Xt[varc])],axis=1)
features = list(Xt.columns)
# Print selection results
print("Best continuous features: ", wcf.selected_features)
print("Best discrete features: ",wdf.selected_features)
print("Best Features selected: ",features)Best continuous features: [{'feature': 'disc_C_AMT_CREDIT_4_kmeans', 'iv': 0.08944178361036469, 'root_feature': 'C_AMT_CREDIT', 'nbins': '4', 'method': 'kmeans'}, {'feature': 'disc_C_AMT_GOODS_PRICE_3_quantile', 'iv': 0.09335492422758512, 'root_feature': 'C_AMT_GOODS_PRICE', 'nbins': '3', 'method': 'quantile'}, {'feature': 'disc_C_AMT_INCOME_TOTAL_3_gaussian', 'iv': 0.05045866117534799, 'root_feature': 'C_AMT_INCOME_TOTAL', 'nbins': '3', 'method': 'gaussian'}, {'feature': 'disc_C_OWN_CAR_AGE_3_kmeans', 'iv': 0.13493592841524896, 'root_feature': 'C_OWN_CAR_AGE', 'nbins': '3', 'method': 'kmeans'}, {'feature': 'disc_C_TOTALAREA_MODE_3_quantile', 'iv': 0.1259702243075047, 'root_feature': 'C_TOTALAREA_MODE', 'nbins': '3', 'method': 'quantile'}]
Best discrete features: {'D_CODE_GENDER': 0.10698023203218116}
Best Features selected: ['D_CODE_GENDER', 'disc_C_AMT_CREDIT_4_kmeans', 'disc_C_AMT_GOODS_PRICE_3_quantile', 'disc_C_AMT_INCOME_TOTAL_3_gaussian', 'disc_C_OWN_CAR_AGE_3_kmeans', 'disc_C_TOTALAREA_MODE_3_quantile']
# Weight of Evidence Transformation
we = WoeEncoder()
we.fit(Xt[features],train['TARGET'])
Xwt = we.transform(Xt[features])
Xwt.head()
<style scoped>
.dataframe tbody tr th:only-of-type {
vertical-align: middle;
}
</style>
.dataframe tbody tr th {
vertical-align: top;
}
.dataframe thead th {
text-align: right;
}
| D_CODE_GENDER | disc_C_AMT_CREDIT_4_kmeans | disc_C_AMT_GOODS_PRICE_3_quantile | disc_C_AMT_INCOME_TOTAL_3_gaussian | disc_C_OWN_CAR_AGE_3_kmeans | disc_C_TOTALAREA_MODE_3_quantile | |
|---|---|---|---|---|---|---|
| 0 | -0.397247 | 0.092313 | 0.185600 | -0.095193 | -0.099648 | -0.230676 |
| 1 | 0.271721 | 0.092313 | -0.309368 | -0.095193 | -0.099648 | -0.035932 |
| 2 | 0.271721 | -0.303484 | -0.309368 | -0.095193 | -0.099648 | -0.126945 |
| 3 | 0.271721 | -0.303484 | -0.309368 | -0.095193 | -0.099648 | -0.126945 |
| 4 | 0.271721 | 0.092313 | 0.185600 | -0.095193 | -0.099648 | -0.230676 |
lr = LogisticRegression()
lr.fit(Xwt,train['TARGET'])
print("AUC for training: ",roc_auc_score(y_score=lr.predict_proba(Xwt)[:,1],y_true=train['TARGET']))AUC for training: 0.6938732132419364
# In order to perform the scoring transformation, we need the WoE encoded data,
# the WoeEncoder fitted object and the logistic regression fitter object
# to produce a nice formatted scorecard
cs = CreditScoring()
cs.fit(Xwt,we,lr)
cs.scorecard
<style scoped>
.dataframe tbody tr th:only-of-type {
vertical-align: middle;
}
</style>
.dataframe tbody tr th {
vertical-align: top;
}
.dataframe thead th {
text-align: right;
}
| points | ||
|---|---|---|
| feature | attribute | |
| D_CODE_GENDER | F | 63 |
| M | 42 | |
| disc_C_AMT_CREDIT_4_kmeans | (1381796.72, 2370559.5] | 48 |
| (456025.115, 869398.644] | 49 | |
| (47969.999, 456025.115] | 57 | |
| (869398.644, 1381796.72] | 67 | |
| disc_C_AMT_GOODS_PRICE_3_quantile | (276000.0, 630000.0] | 49 |
| (44999.999, 276000.0] | 58 | |
| (630000.0, 2254500.0] | 59 | |
| MISSING | 12 | |
| disc_C_AMT_INCOME_TOTAL_3_gaussian | (173250.0, 337500.0] | 62 |
| (28403.999, 173250.0] | 53 | |
| (337500.0, 810000.0] | 43 | |
| disc_C_OWN_CAR_AGE_3_kmeans | (-0.001, 12.582] | 74 |
| (12.582, 41.679] | 43 | |
| (41.679, 65.0] | 41 | |
| MISSING | 52 | |
| disc_C_TOTALAREA_MODE_3_quantile | (-0.001, 0.053] | 49 |
| (0.053, 0.111] | 54 | |
| (0.111, 0.661] | 76 | |
| MISSING | 52 |
# Applying all transformations to the validation data is now easy and straightforward
# we can compute AUC to check model overfitting
Xv = pd.concat([wdf.transform(dn.transform(valid[vard])),wcf.transform(valid[varc])],axis=1)
Xwv = we.transform(Xv)
print("AUC for validation: ",roc_auc_score(y_score=lr.predict_proba(Xwv)[:,1],y_true=valid['TARGET']))AUC for validation: 0.6971411870981454
# We can check the score transformation distributions for training and validation
score = pd.concat([pd.concat([cs.transform(we.inverse_transform(Xwv))[['score']].assign(sample='validation'),valid['TARGET']],axis=1),
pd.concat([cs.transform(we.inverse_transform(Xwt))[['score']].assign(sample='train'),train['TARGET']],axis=1)
],ignore_index=True)
for s,d in score.groupby('sample'):
plt.figure()
plt.title(s)
sns.histplot(d['score'],legend=True,fill=True,bins=8)
# Finally, we can observe that, the greater the score, the lower the probability of being a
# bad customer (label=1) for both samples. Now all complexity is absorbed
score['score_range'] = pd.cut(score['score'],bins=6,include_lowest=True).astype(str)
for s,d in score.groupby('sample'):
aux = d.pivot_table(index='TARGET',
columns='score_range',
values='score',
aggfunc='count',
fill_value=0)
aux/=aux.sum()
aux = aux.T
plt.figure()
ax = aux.plot(kind='bar',stacked=True,color=['purple','black'])
plt.title(s)<Figure size 432x288 with 0 Axes>
<Figure size 432x288 with 0 Axes>
If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement".
Don't forget to give the project a star! Thanks again!
-
Fork the Project
-
Create your Feature Branch (
git checkout -b feature/AmazingFeature) -
Commit your Changes (
git commit -m 'Add some AmazingFeature') -
Push to the Branch (
git push origin feature/AmazingFeature) -
Open a Pull Request
Distributed under the GNU General Public License v3.0 License. See LICENSE for more information.
José G Fuentes - @jgusteacher - [email protected]
Project Link: https://github.com/JGFuentesC/woe_credit_scoring
If you use this software in scientific publications, we would appreciate citations to the following paper:
Combination of Unsupervised Discretization Methods for Credit Risk José G. Fuentes Cabrera, Hugo A. Pérez Vicente, Sebastián Maldonado,Jonás Velasco
-
Siddiqi, N. (2012). Credit risk scorecards: developing and implementing intelligent credit scoring (Vol. 3). John Wiley & Sons.. For his amazing textbook.
-
@othneildrew. For his amazing README template
-
Demo data. For providing example data.