LoanTap Case Study

Introduction
LoanTap is at the forefront of offering tailored financial solutions to milennials. Their innovative approach
seeks to harness data science for refining their credit underwriting process. The focus here is the Personal
Loan segment. A deep dive into the dataset can reveal patterns in borrower behaviour and creditworthiness.
Analyzing this dataset can provide crucial insights into the financial behaviours, spending habits and
potential risk associated with each borrower. The insights gained can optimize loan disbursal, balancing
customer outreach with risk management.

What is expected
Assuming you are a data scientist at LoanTap, you are tasked with analyzing the dataset to determine the
creditworthiness of potential borrowers. Your ultimate objective is to build a logistic regression model,
evaluate its performance, and provide actionable insights for the underwriting process.

1. Data
The analysis was done on the data located at -
https://drive.google.com/file/d/1ZPYj7CZCfxntE8p2Lze_4QO4MyEOy6_d/view?usp=sharing

2. Libraries
Below are the libraries required

In [1]: # libraries to analyze data

import numpy as np
import pandas as pd

# libraries to visualize data

import matplotlib.pyplot as plt
import seaborn as sns

import statsmodels.api as sm
from statsmodels.stats.outliers_influence import variance_inflation_factor

from sklearn.impute import KNNImputer

from sklearn.preprocessing import OneHotEncoder, MinMaxScaler
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report, confusion_matrix, ConfusionMatrixDisp
from sklearn.pipeline import make_pipeline

from imblearn.over_sampling import SMOTE

3. Data Loading
 Loading the data into Pandas dataframe for easily handling of data

In [2]: # read the file into a pandas dataframe

customer_df = pd.read_csv('LoanTapData.csv')
df = customer_df
# look at the datatypes of the columns
print('*************************************************')
print(df.info())
print('*************************************************\n')
print('*************************************************')
print(f'Shape of the dataset is {df.shape}')
print('*************************************************\n')
print('*************************************************')
print(f'Number of nan/null values in each column: \n{df.isna().sum()}')
print('*************************************************\n')
print('*************************************************')
print(f'Number of unique values in each column: \n{df.nunique()}')
print('*************************************************\n')
print('*************************************************')
print(f'Duplicate entries: \n{df.duplicated().value_counts()}')
print('*************************************************')

*************************************************
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 396030 entries, 0 to 396029
Data columns (total 27 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 loan_amnt 396030 non-null float64
1 term 396030 non-null object
2 int_rate 396030 non-null float64
3 installment 396030 non-null float64
4 grade 396030 non-null object
5 sub_grade 396030 non-null object
6 emp_title 373103 non-null object
7 emp_length 377729 non-null object
8 home_ownership 396030 non-null object
9 annual_inc 396030 non-null float64
10 verification_status 396030 non-null object
11 issue_d 396030 non-null object
12 loan_status 396030 non-null object
13 purpose 396030 non-null object
14 title 394274 non-null object
15 dti 396030 non-null float64
16 earliest_cr_line 396030 non-null object
17 open_acc 396030 non-null float64
18 pub_rec 396030 non-null float64
19 revol_bal 396030 non-null float64
20 revol_util 395754 non-null float64
21 total_acc 396030 non-null float64
22 initial_list_status 396030 non-null object
23 application_type 396030 non-null object
24 mort_acc 358235 non-null float64
25 pub_rec_bankruptcies 395495 non-null float64
26 address 396030 non-null object
dtypes: float64(12), object(15)
memory usage: 81.6+ MB
None
*************************************************

*************************************************
Shape of the dataset is (396030, 27)
*************************************************

*************************************************
Number of nan/null values in each column:
loan_amnt 0
term 0
int_rate 0
installment 0
grade 0
sub_grade 0
emp_title 22927
emp_length 18301
home_ownership 0
annual_inc 0
verification_status 0
issue_d 0
loan_status 0
purpose 0
title 1756
dti 0
earliest_cr_line 0
open_acc 0
pub_rec 0
revol_bal 0
revol_util 276
total_acc 0
initial_list_status 0
application_type 0
mort_acc 37795
pub_rec_bankruptcies 535
address 0
dtype: int64
*************************************************

*************************************************
Number of unique values in each column:
loan_amnt 1397
term 2
int_rate 566
installment 55706
grade 7
sub_grade 35
emp_title 173105
emp_length 11
home_ownership 6
annual_inc 27197
verification_status 3
issue_d 115
loan_status 2
purpose 14
title 48816
dti 4262
earliest_cr_line 684
open_acc 61
pub_rec 20
revol_bal 55622
revol_util 1226
total_acc 118
initial_list_status 2
application_type 3
mort_acc 33
pub_rec_bankruptcies 9
address 393700
dtype: int64
*************************************************

*************************************************
Duplicate entries:
False 396030
 Name: count, dtype: int64
*************************************************

In [3]: # look at the top 5 rows

df.head()

Out[3]: loan_amnt term int_rate installment grade sub_grade emp_title emp_length home_ownership annu

36
0 10000.0 11.44 329.48 B B4 Marketing 10+ years RENT 117
months

36 Credit
1 8000.0 11.99 265.68 B B5 4 years MORTGAGE 65
months analyst

36
2 15600.0 10.49 506.97 B B3 Statistician < 1 year RENT 43
months

36 Client
3 7200.0 6.49 220.65 A A2 6 years RENT 54
months Advocate

Destiny
60
4 24375.0 17.27 609.33 C C5 Management 9 years MORTGAGE 55
months
Inc.

5 rows × 27 columns

Not all columns are visible, so looking at groups of columns

In [4]: df[df.columns[10:20]]

Out[4]: verification_status issue_d loan_status purpose title dti earliest_cr_line open_acc

Jan-
0 Not Verified Fully Paid vacation Vacation 26.24 Jun-1990 16.0
2015

Jan- Debt
1 Not Verified Fully Paid debt_consolidation 22.05 Jul-2004 17.0
2015 consolidation

Jan- Credit card

2 Source Verified Fully Paid credit_card 12.79 Aug-2007 13.0
2015 refinancing

Nov- Credit card

3 Not Verified Fully Paid credit_card 2.60 Sep-2006 6.0
2014 refinancing

Apr- Charged Credit Card

4 Verified credit_card 33.95 Mar-1999 13.0
2013 Off Refinance

... ... ... ... ... ... ... ... ...

Oct- Debt
396025 Source Verified Fully Paid debt_consolidation 15.63 Nov-2004 6.0
2015 consolidation

Feb- Debt
396026 Source Verified Fully Paid debt_consolidation 21.45 Feb-2006 6.0
2015 consolidation

Oct- pay off credit

396027 Verified Fully Paid debt_consolidation 17.56 Mar-1997 15.0
2013 cards

Aug-
396028 Verified Fully Paid debt_consolidation Loanforpayoff 15.88 Nov-1990 9.0
2012

Jun- Toxic Debt

396029 Verified Fully Paid debt_consolidation 8.32 Sep-1998 3.0
2010 Payoff
 396030 rows × 10 columns

In [5]: df[df.columns[20:]]

Out[5]: revol_util total_acc initial_list_status application_type mort_acc pub_rec_bankruptcies

0174
0 41.8 25.0 w INDIVIDUAL 0.0 0.0 Gateway\r\nMendo
O

1076 Carney F
1 53.3 27.0 f INDIVIDUAL 3.0 0.0 347\r\nLoganmo

87025 Mark D
2 92.2 26.0 f INDIVIDUAL 0.0 0.0 269\r\nNew Sab

8
3 21.5 13.0 f INDIVIDUAL 0.0 0.0 Ford\r\nDelacruzs

6
4 69.8 43.0 f INDIVIDUAL 1.0 0.0 Roads\r\nGreggs

... ... ... ... ... ... ...

12951 W
396025 34.3 23.0 w INDIVIDUAL 0.0 0.0 Crossing\r\nJoh
D

0114 Fowler Fie

396026 95.7 8.0 f INDIVIDUAL 1.0 0.0
028\r\nRachelbor

953 Matthew Poin

396027 66.9 23.0 f INDIVIDUAL 0.0 0.0
414\r\nReedfor

7843 Blake Freew

396028 53.8 20.0 f INDIVIDUAL 5.0 0.0
229\r\nNew Mich

787
396029 91.3 19.0 f INDIVIDUAL NaN 0.0 Causeway\r\nBria
A

396030 rows × 7 columns

In [6]: df.describe()

Out[6]: loan_amnt int_rate installment annual_inc dti open_acc pub_rec

count 396030.000000 396030.000000 396030.000000 3.960300e+05 396030.000000 396030.000000 396030.000000

mean 14113.888089 13.639400 431.849698 7.420318e+04 17.379514 11.311153 0.178191

std 8357.441341 4.472157 250.727790 6.163762e+04 18.019092 5.137649 0.530671

min 500.000000 5.320000 16.080000 0.000000e+00 0.000000 0.000000 0.000000

25% 8000.000000 10.490000 250.330000 4.500000e+04 11.280000 8.000000 0.000000

50% 12000.000000 13.330000 375.430000 6.400000e+04 16.910000 10.000000 0.000000

75% 20000.000000 16.490000 567.300000 9.000000e+04 22.980000 14.000000 0.000000

max 40000.000000 30.990000 1533.810000 8.706582e+06 9999.000000 90.000000 86.000000

 Insight
There are 396030 entries with 27 columns
There are 22927 null/missing values in emp_title, 18301 in emp_length, 1756 in title, 276 in
revol_util, 37795 in mort_acc and 535 in pub_rec_bankruptcies
There are no duplicates
Based on the number of unique values in each column and their datatype, the columns term, grade,
emp_length, home_ownership, verification_status, loan_status, purpose, initial_list_status and
application_type can be converted to categorical datatype
issue_d and earliest_cr_line has date values so can be convered to datetime
The column sub_grade can be dropped as its info is already captured in column grade
The columns term and emp_length have space, special characters, alphanumeric mix.
Replace "36 months" with "short" and "60 months" with "long" in term column
Remove "year/years", replacing "< 1" with "0" and replacing "10+" with "10" in emp_length column.
Then convert emp_length column to categorical datatype
Extract zip code from address column and drop address column

In [7]: # Convert to category

categorical_columns = ['term', 'grade', 'emp_length', 'home_ownership', 'verification_st
df[categorical_columns] = df[categorical_columns].astype('category')

# Convert to datetime
df['issue_d'] = pd.to_datetime(df['issue_d'], format='%b-%Y')
df['earliest_cr_line'] = pd.to_datetime(df['earliest_cr_line'], format='%b-%Y')

# Drop "grade" column

df.drop(columns=['sub_grade'], inplace=True)

# Rename the values in 'term' column

df['term'].replace({' 36 months' : 'short', ' 60 months' : 'long'}, inplace=True)

# Rename employee_length column values

df['emp_length'] = df['emp_length'].replace({'< 1 year':'0 year'})
df['emp_length'] = df['emp_length'].str.replace(r'\D', '', regex=True)
df['emp_length'] = df['emp_length'].astype('category')

# Extract zip code from address

df['zip_code'] = df['address'].str[-5:].str.split().str[0].astype('category')
df.drop(columns='address', inplace=True)
print(f'Number of nan/null values in zip_code column: \n{df.zip_code.isna().sum()}')

df.info()

Number of nan/null values in zip_code column:

0
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 396030 entries, 0 to 396029
Data columns (total 26 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 loan_amnt 396030 non-null float64
1 term 396030 non-null category
2 int_rate 396030 non-null float64
3 installment 396030 non-null float64
4 grade 396030 non-null category
5 emp_title 373103 non-null object
6 emp_length 377729 non-null category
7 home_ownership 396030 non-null category
8 annual_inc 396030 non-null float64
9 verification_status 396030 non-null category
 10 issue_d 396030 non-null datetime64[ns]
11 loan_status 396030 non-null category
12 purpose 396030 non-null category
13 title 394274 non-null object
14 dti 396030 non-null float64
15 earliest_cr_line 396030 non-null datetime64[ns]
16 open_acc 396030 non-null float64
17 pub_rec 396030 non-null float64
18 revol_bal 396030 non-null float64
19 revol_util 395754 non-null float64
20 total_acc 396030 non-null float64
21 initial_list_status 396030 non-null category
22 application_type 396030 non-null category
23 mort_acc 358235 non-null float64
24 pub_rec_bankruptcies 395495 non-null float64
25 zip_code 396030 non-null category
dtypes: category(10), datetime64[ns](2), float64(12), object(2)
memory usage: 52.1+ MB

4. Exploratory Data Analysis

4.1. Handling null values

emp_title has 22927, emp_length has 18301, title has 1756, revol_util has 276, mort_acc has 37795 and
pub_rec_bankruptcies has 535 null values

Columns emp_title and title can be dropped as they would not have an effect on the loan approval
Null values in revol_util and pub_rec_bankruptcies are small in number and hence can be dropped

In [8]: df.drop(columns=['emp_title', 'title'], inplace=True)

df.dropna(subset = ['revol_util', 'pub_rec_bankruptcies'], inplace=True)

Let us check the distribution of remaining features before deciding on how to handle the null values

In [9]: fig, axs = plt.subplots(1,2, figsize=(10,4))

sns.histplot(ax = axs[0], data=df, x = 'emp_length')
sns.histplot(ax = axs[1], data=df, x = 'mort_acc')
fig.tight_layout()
plt.show()
 mort_acc is the number of mortgae accounts out of the total_acc. So I will replace the null values with
"mode" of mort_acc for different total_acc
For emp_length, I will use knn imputer to fill the missing data

In [10]: mode_mort_acc_df = df.groupby('total_acc')['mort_acc'].agg(lambda x: pd.Series.mode(x)[0

def fill_mort_acc(total_acc, mort_acc):

if np.isnan(mort_acc):
return mode_mort_acc_df[total_acc]
else:
return mort_acc

df['mort_acc'] = df.apply(lambda x: fill_mort_acc(x['total_acc'],x['mort_acc']), axis=1)

In [11]: imputer = KNNImputer(n_neighbors=5)

df['emp_length']=imputer.fit_transform(df[['emp_length']])
df['emp_length'] = df['emp_length'].astype('category')

In [12]: df.isna().sum()

loan_amnt 0
Out[12]:
term 0
int_rate 0
installment 0
grade 0
emp_length 0
home_ownership 0
annual_inc 0
verification_status 0
issue_d 0
loan_status 0
purpose 0
dti 0
earliest_cr_line 0
open_acc 0
pub_rec 0
revol_bal 0
revol_util 0
total_acc 0
initial_list_status 0
application_type 0
mort_acc 0
pub_rec_bankruptcies 0
zip_code 0
dtype: int64

There are no null values now

4.2. Detecting outliers

4.2.1. Outliers for every continuous variable
In [13]: # helper function to detect outliers using IQR method
def detectOutliers_iqr(df):
q1 = df.quantile(0.25)
q3 = df.quantile(0.75)
iqr = q3-q1
lower_outliers = df[df<(q1-1.5*iqr)]
higher_outliers = df[df>(q3+1.5*iqr)]
return lower_outliers, higher_outliers
In [14]: # helper function to detect outliers using standard deviation method
def detectOutliers_std(df):
mean = df.mean()
std = df.std()
upper_limit = mean+(3*std)
lower_limit = mean-(3*std)
lower_outliers = df[df<lower_limit]
higher_outliers = df[df>upper_limit]
return lower_outliers, higher_outliers

In [15]: numerical_columns = df.select_dtypes(include=np.number).columns

column_outlier_dictionary = {}
for column in numerical_columns:
lower_outliers, higher_outliers = detectOutliers_iqr(df[column])
column_outlier_dictionary[column] = [lower_outliers, higher_outliers]
#print('*'*50)
#print(f'Outliers of \'{column}\' column are:')
#print("Lower outliers:\n", lower_outliers)
#print("Higher outliers:\n", higher_outliers)
#print('*'*50, end="\n")

In [16]: for key, value in column_outlier_dictionary.items():

print(f'The column \'{key}\' has {len(value[0]) + len(value[1])} outliers')

The column 'loan_amnt' has 190 outliers

The column 'int_rate' has 3144 outliers
The column 'installment' has 11114 outliers
The column 'annual_inc' has 16649 outliers
The column 'dti' has 275 outliers
The column 'open_acc' has 10297 outliers
The column 'pub_rec' has 57730 outliers
The column 'revol_bal' has 21205 outliers
The column 'revol_util' has 12 outliers
The column 'total_acc' has 8491 outliers
The column 'mort_acc' has 6837 outliers
The column 'pub_rec_bankruptcies' has 45111 outliers

In [17]: num_cols = 6
num_rows = int(np.ceil(len(numerical_columns)/num_cols))
fig, axs = plt.subplots(num_rows,num_cols,figsize=(10,15))
for idx in range(len(numerical_columns)):
ax = plt.subplot(num_rows, num_cols, idx+1)
sns.boxplot(ax = ax, data=df, y = numerical_columns[idx])
plt.tight_layout()
plt.show()
In [18]: numerical_columns = df.select_dtypes(include=np.number).columns
numerical_columns = list(numerical_columns)
numerical_columns.remove('pub_rec')
numerical_columns.remove('pub_rec_bankruptcies')
numerical_columns = pd.core.indexes.base.Index(numerical_columns)
column_outlier_dictionary = {}
for column in numerical_columns:
lower_outliers, higher_outliers = detectOutliers_std(df[column])
column_outlier_dictionary[column] = [lower_outliers, higher_outliers]
#print('*'*50)
#print(f'Outliers of \'{column}\' column are:')
#print("Lower outliers:\n", lower_outliers)
#print("Higher outliers:\n", higher_outliers)
#print('*'*50, end="\n")

In [19]: for key, value in column_outlier_dictionary.items():

print(f'The column \'{key}\' has {len(value[0]) + len(value[1])} outliers')

The column 'loan_amnt' has 184 outliers

The column 'int_rate' has 754 outliers
The column 'installment' has 5042 outliers
The column 'annual_inc' has 3190 outliers
The column 'dti' has 12 outliers
The column 'open_acc' has 4873 outliers
The column 'revol_bal' has 4771 outliers
The column 'revol_util' has 16 outliers
The column 'total_acc' has 3396 outliers
The column 'mort_acc' has 6837 outliers

4.2.2. Remove the outliers

Based on the boxplot, the number of outliers using IQR method and standard deviation method, I will
remove the outliers using the standard deviation method except for columns pub_rec and
pub_rec_bankruptcies which will be removed based on manual check.

In [20]: remove_outliers = True

if True == remove_outliers:
master_index = pd.core.indexes.base.Index([])
for key, value in column_outlier_dictionary.items():
lower_outliers = value[0]
higher_outliers = value[1]
master_index = master_index.union(lower_outliers.index).union(higher_outliers.in
df.drop(master_index, inplace=True)
else:
print('Not removing any outliers')

In [21]: df['pub_rec'].value_counts()

pub_rec
Out[21]:
0.0 315552
1.0 47129
2.0 5107
3.0 1424
4.0 481
5.0 218
6.0 108
7.0 47
8.0 31
10.0 11
9.0 10
11.0 6
13.0 4
 12.0 4
19.0 2
40.0 1
17.0 1
86.0 1
24.0 1
15.0 1
Name: count, dtype: int64

In [22]: df['pub_rec_bankruptcies'].value_counts()

pub_rec_bankruptcies
Out[22]:
0.0 327200
1.0 40774
2.0 1716
3.0 332
4.0 75
5.0 30
6.0 6
7.0 4
8.0 2
Name: count, dtype: int64

In [23]: df = df[df['pub_rec'] < 11]

df = df[df['pub_rec_bankruptcies'] < 6]

In [24]: numerical_columns = df.select_dtypes(include=np.number).columns

num_cols = 6
num_rows = int(np.ceil(len(numerical_columns)/num_cols))
fig, axs = plt.subplots(num_rows,num_cols,figsize=(10,15))
for idx in range(len(numerical_columns)):
ax = plt.subplot(num_rows, num_cols, idx+1)
sns.boxplot(ax = ax, data=df, y = numerical_columns[idx])
plt.tight_layout()
plt.show()
In [25]: df.info()

<class 'pandas.core.frame.DataFrame'>
Index: 370106 entries, 0 to 396029
Data columns (total 24 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 loan_amnt 370106 non-null float64
1 term 370106 non-null category
2 int_rate 370106 non-null float64
3 installment 370106 non-null float64
4 grade 370106 non-null category
5 emp_length 370106 non-null category
6 home_ownership 370106 non-null category
7 annual_inc 370106 non-null float64
8 verification_status 370106 non-null category
9 issue_d 370106 non-null datetime64[ns]
10 loan_status 370106 non-null category
11 purpose 370106 non-null category
12 dti 370106 non-null float64
13 earliest_cr_line 370106 non-null datetime64[ns]
14 open_acc 370106 non-null float64
15 pub_rec 370106 non-null float64
16 revol_bal 370106 non-null float64
17 revol_util 370106 non-null float64
18 total_acc 370106 non-null float64
19 initial_list_status 370106 non-null category
20 application_type 370106 non-null category
21 mort_acc 370106 non-null float64
22 pub_rec_bankruptcies 370106 non-null float64
23 zip_code 370106 non-null category
dtypes: category(10), datetime64[ns](2), float64(12)
memory usage: 45.9 MB

Insight
The number of columns reduced to 370106 from the original 396030 rows

4.3. Univariate analysis

4.3.1. Numerical Variables

In [26]: num_cols = 2
num_rows = int(np.ceil(len(numerical_columns)/num_cols))
fig, axs = plt.subplots(num_rows,num_cols,figsize=(10,15))
for idx in range(len(numerical_columns)):
ax = plt.subplot(num_rows, num_cols, idx+1)
sns.histplot(ax = ax, data=df, x = numerical_columns[idx], kde=True)
plt.tight_layout()
plt.show()
 Insight
Most of the features are right skewed except for revol_util which is slightly left skewed

4.3.2. Categorical Variables

In [27]: categorical_columns = df.select_dtypes(include='category').columns
num_cols = 2
num_rows = int(np.ceil(len(categorical_columns)/num_cols))
fig, axs = plt.subplots(num_rows,num_cols,figsize=(12,30))
for idx in range(len(categorical_columns)):
ax = plt.subplot(num_rows, num_cols, idx+1)
sns.countplot(ax = ax, data=df, y = categorical_columns[idx], stat='percent')
plt.tight_layout()
plt.show()
Insight
 Most of the loan is taken for a short term which is for 3 years
Maximum loan are assigned the grade B followed by grade C
Maximum loan are taken by borrowers whose employement tenure is more than 10 years
Maximum loan are taken by borrowers whose house is mortgaged followed by those whose who are in
rented house
Most of the loan is fully_paid
Most of the loan is taken for debt consolidation followed by credit card
Majority of the loan is taken by individuals

4.4. Bivariate analysis

In [28]: sns.boxplot(data=df, x='loan_status', y='loan_amnt')
plt.show()

Insight
The median of the loan amount slighly higher for loans which were charged off

In [29]: fig, axs = plt.subplots(2,2,figsize=(10,6))

sns.scatterplot(ax=axs[0,0], data=df, x='loan_amnt', y='installment')
sns.scatterplot(ax=axs[0,1], data=df, x='loan_amnt', y='annual_inc')
sns.scatterplot(ax=axs[1,0], data=df, x='annual_inc', y='dti')
sns.scatterplot(ax=axs[1,1], data=df, x='total_acc', y='open_acc')
fig.tight_layout()
plt.show()
 Insight
It is very clear from the plot that, in general, the installment increases as the loan amount increases
It is very obvious that people with higher income can afford to take higher loan
As income increases debt to income ratio reduces
Most of the borrower's accounts are active accounts

In [30]: fig, axs = plt.subplots(2,2,figsize=(10,6))

sns.histplot(ax=axs[0,0], data=df, x='pub_rec', hue='loan_status')
sns.histplot(ax=axs[0,1], data=df, x='mort_acc', hue='loan_status')
sns.histplot(ax=axs[1,0], data=df, x='pub_rec_bankruptcies', hue='loan_status')
fig.tight_layout()
plt.show()
 Insight
I will group pub_rec into 2 groups: 0 for having 0 negative records and 1 for having more than 0
negative records
Same for mort_acc and pub_rec_bankruptcies

In [31]: group_0_list = [0.0]

pub_rec_list = list(df['pub_rec'].explode().unique())
group_1_list = list(set(pub_rec_list) - set(group_0_list))
df['any_neg_rec'] = df['pub_rec'].replace(group_0_list, 0)
df['any_neg_rec'] = df['any_neg_rec'].replace(group_1_list, 1)
df['any_neg_rec'] = df['any_neg_rec'].astype('category')

group_0_list = [0.0]
pub_rec_list = list(df['mort_acc'].explode().unique())
group_1_list = list(set(pub_rec_list) - set(group_0_list))
df['any_mort'] = df['mort_acc'].replace(group_0_list, 0)
df['any_mort'] = df['any_mort'].replace(group_1_list, 1)
df['any_mort'] = df['any_mort'].astype('category')

group_0_list = [0.0]
pub_rec_list = list(df['pub_rec_bankruptcies'].explode().unique())
group_1_list = list(set(pub_rec_list) - set(group_0_list))
df['any_bankruptcies'] = df['pub_rec_bankruptcies'].replace(group_0_list, 0)
df['any_bankruptcies'] = df['any_bankruptcies'].replace(group_1_list, 1)
df['any_bankruptcies'] = df['any_bankruptcies'].astype('category')

df.drop(columns = ['pub_rec', 'mort_acc', 'pub_rec_bankruptcies'], inplace=True)

In [32]: fig, axs = plt.subplots(2,2,figsize=(10,6))

sns.histplot(ax=axs[0,0], data=df, x='any_neg_rec', hue='loan_status', stat="proportion"
sns.histplot(ax=axs[0,1], data=df, x='any_mort', hue='loan_status', stat="proportion", m
sns.histplot(ax=axs[1,0], data=df, x='any_bankruptcies', hue='loan_status', stat="propor
fig.tight_layout()
plt.show()
In [33]: fig, axs = plt.subplots(3,2,figsize=(12,10))
sns.histplot(ax=axs[0,0], data=df, x='term', hue='loan_status', stat="proportion", multip
sns.histplot(ax=axs[0,1], data=df, x='home_ownership', hue='loan_status', stat="proporti
sns.histplot(ax=axs[1,0], data=df, x='verification_status', hue='loan_status', stat="prop
sns.histplot(ax=axs[1,1], data=df, x='initial_list_status', hue='loan_status', stat="prop
sns.histplot(ax=axs[2,0], data=df, x='application_type', hue='loan_status', stat="propor
fig.tight_layout()
plt.show()

fig, axs = plt.subplots(4,1,figsize=(15,15))

sns.histplot(ax=axs[0], data=df, x='grade', hue='loan_status', stat="proportion", multip
sns.histplot(ax=axs[1], data=df, x='emp_length', hue='loan_status', stat="proportion", m
sns.histplot(ax=axs[2], data=df, y='purpose', hue='loan_status', stat="proportion", mult
sns.histplot(ax=axs[3], data=df, x='zip_code', hue='loan_status', stat="proportion", mul
fig.tight_layout()
plt.show()
 Insight
Having a negative or bankruptcy record doesnt seem to impact the loan getting paid back or not
Loan taken for short term, i.e. 3 years are most likely to be fully paid back
Loan taken by people whose house ownership is of type any is on are paid back 100%.
Suprisingly, loans which are not verified are more likely to be paid back
Initial list status also do not seem to impact loan status
Loan taken as joint application type are more likely to be paid back
Loans with grade A and B are more likely to get paid back
Employment duration does not seem to impact loan status
Loan taken for wedding are more likely to be paid back
Loan taken by people with zip code 00813 and 05113 are fully paid back whereas loan taken by
people with zip code 11650, 86630 and 93700 are all charged off

In [34]: df.drop(columns=['any_neg_rec', 'any_bankruptcies', 'initial_list_status', 'emp_length']

 4.5. Multivariate analysis
In [35]: fig, ax = plt.subplots(figsize=(8,6))
sns.heatmap(df.select_dtypes(include=np.number).corr(), annot=True, linewidth=0.5, cmap =
ax.xaxis.tick_top()
plt.xticks(rotation=30, ha='left')
plt.show()

Insight
loan amount is highly correlated with installment
There is good correlation between loan amount - annual income, loan amount - revol balance,
installment - annual income, installment - revol balance, open account - total account

5. Data Preprocessing
In [36]: df.info()
<class 'pandas.core.frame.DataFrame'>
Index: 370106 entries, 0 to 396029
Data columns (total 20 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 loan_amnt 370106 non-null float64
 1 term 370106 non-null category
2 int_rate 370106 non-null float64
3 installment 370106 non-null float64
4 grade 370106 non-null category
5 home_ownership 370106 non-null category
6 annual_inc 370106 non-null float64
7 verification_status 370106 non-null category
8 issue_d 370106 non-null datetime64[ns]
9 loan_status 370106 non-null category
10 purpose 370106 non-null category
11 dti 370106 non-null float64
12 earliest_cr_line 370106 non-null datetime64[ns]
13 open_acc 370106 non-null float64
14 revol_bal 370106 non-null float64
15 revol_util 370106 non-null float64
16 total_acc 370106 non-null float64
17 application_type 370106 non-null category
18 zip_code 370106 non-null category
19 any_mort 370106 non-null category
dtypes: category(9), datetime64[ns](2), float64(9)
memory usage: 37.1 MB

The date features will not have an impact on the loan status, so i will drop issue_d and earliest_cr_line
columns

In [37]: df.drop(columns=['issue_d', 'earliest_cr_line'], inplace=True)

5.1. Multicollinearity Check

In [38]: features_df = df.drop(columns=['loan_status']) # Drop target column
features_df.select_dtypes(include='category').columns
features_df = features_df.drop(columns=features_df.select_dtypes(include='category').col
features_df = sm.add_constant(features_df) # Adding a constant column for the intercept
vif_df = pd.DataFrame()
vif_df['Features'] = features_df.columns
vif_df['VIF'] = [variance_inflation_factor(features_df.values, idx) for idx in range(len
vif_df['VIF'] = round(vif_df['VIF'], 2)
vif_df = vif_df.sort_values(by='VIF', ascending=False)
vif_df

Out[38]: Features VIF

0 const 24.91

1 loan_amnt 11.58

3 installment 11.04

6 open_acc 2.00

9 total_acc 1.87

7 revol_bal 1.76

4 annual_inc 1.62

8 revol_util 1.47

5 dti 1.40

2 int_rate 1.23

Insight
 loan amount is highly correlated with installment which is also shown here by high VIF values. I will
drop installment

In [39]: features_df = features_df.drop(columns=['installment'])

features_df = sm.add_constant(features_df) # Adding a constant column for the intercept
vif_df = pd.DataFrame()
vif_df['Features'] = features_df.columns
vif_df['VIF'] = [variance_inflation_factor(features_df.values, idx) for idx in range(len
vif_df['VIF'] = round(vif_df['VIF'], 2)
vif_df = vif_df.sort_values(by='VIF', ascending=False)
vif_df

Out[39]: Features VIF

0 const 24.76

5 open_acc 2.00

8 total_acc 1.86

6 revol_bal 1.75

3 annual_inc 1.62

1 loan_amnt 1.49

7 revol_util 1.46

4 dti 1.40

2 int_rate 1.22

Insight
Based on the above VIF scores, I can conclude that there are no more multicolinear numerical features
I will drop installment from the dataframe

In [40]: df.drop(columns=['installment'], inplace=True)

5.2. Encode categorical variables

In [41]: final_df = df.copy()
final_df.reset_index(inplace=True, drop=True)

Sepearte out target and feature columns

In [42]: X = final_df.drop(columns=['loan_status'])
y = final_df['loan_status']

Encode target variable

In [43]: y = y.replace({'Fully Paid': 0, 'Charged Off': 1}).astype(int)

Encode features with just 2 classes as 0 or 1

In [44]: X['term'] = X['term'].replace({'short': 0, 'long': 1}).astype('int8')

X['any_mort'] = X['any_mort'].astype('int8')
 One-Hot-Encoding for remaining categorical features

In [45]: X.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 370106 entries, 0 to 370105
Data columns (total 16 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 loan_amnt 370106 non-null float64
1 term 370106 non-null int8
2 int_rate 370106 non-null float64
3 grade 370106 non-null category
4 home_ownership 370106 non-null category
5 annual_inc 370106 non-null float64
6 verification_status 370106 non-null category
7 purpose 370106 non-null category
8 dti 370106 non-null float64
9 open_acc 370106 non-null float64
10 revol_bal 370106 non-null float64
11 revol_util 370106 non-null float64
12 total_acc 370106 non-null float64
13 application_type 370106 non-null category
14 zip_code 370106 non-null category
15 any_mort 370106 non-null int8
dtypes: category(6), float64(8), int8(2)
memory usage: 25.4 MB

In [46]: categorical_columns = X.select_dtypes(include='category').columns

categorical_columns

Index(['grade', 'home_ownership', 'verification_status', 'purpose',

Out[46]:
'application_type', 'zip_code'],
dtype='object')

In [47]: encoder = OneHotEncoder(sparse_output=False)

encoded_data = encoder.fit_transform(X[categorical_columns])
encoded_df = pd.DataFrame(encoded_data, columns = encoder.get_feature_names_out(categori
X = pd.concat([X, encoded_df], axis=1)
X.drop(columns = categorical_columns, inplace=True)
X.head()

Out[47]: loan_amnt term int_rate annual_inc dti open_acc revol_bal revol_util total_acc any_mort ... zip_code

0 10000.0 0 11.44 117000.0 26.24 16.0 36369.0 41.8 25.0 0 ...

1 8000.0 0 11.99 65000.0 22.05 17.0 20131.0 53.3 27.0 1 ...

2 15600.0 0 10.49 43057.0 12.79 13.0 11987.0 92.2 26.0 0 ...

3 7200.0 0 6.49 54000.0 2.60 6.0 5472.0 21.5 13.0 0 ...

4 24375.0 1 17.27 55000.0 33.95 13.0 24584.0 69.8 43.0 1 ...

5 rows × 53 columns

5.3. Train-test split

In [48]: X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)
X_train.shape, X_test.shape, y_train.shape, y_test.shape

((296084, 53), (74022, 53), (296084,), (74022,))

Out[48]:
 5.4. Perform data normalization/standardization
Data normalization/standardization is required so that features with higher scales do not dominate the
model's performance. Hence all features should have same scale\

Data before scaling

In [49]: X_train.head()

Out[49]: loan_amnt term int_rate annual_inc dti open_acc revol_bal revol_util total_acc any_mort ... zip

133405 27000.0 1 16.29 82302.0 25.52 13.0 12014.0 48.6 29.0 1 ...

365868 6000.0 0 18.55 45000.0 19.37 8.0 3219.0 73.2 11.0 0 ...

71124 8975.0 0 9.71 65000.0 7.98 10.0 3932.0 34.5 58.0 0 ...

33923 9600.0 0 6.62 58000.0 25.01 10.0 57236.0 36.3 19.0 1 ...

30512 18000.0 0 11.53 75000.0 8.50 9.0 9916.0 35.8 17.0 0 ...

5 rows × 53 columns

In [50]: min_max_scaler = MinMaxScaler()

# Fit min_max_scaler to training data
min_max_scaler.fit(X_train)
# Scale the training and testing data
X_train = pd.DataFrame(min_max_scaler.transform(X_train), columns=X_train.columns)
X_test = pd.DataFrame(min_max_scaler.transform(X_test), columns=X_test.columns)

Data after scaling

In [51]: X_train.head()

Out[51]: loan_amnt term int_rate annual_inc dti open_acc revol_bal revol_util total_acc any_mort ... zip_c

0 0.697828 1.0 0.506230 0.308276 0.372229 0.48 0.154818 0.394481 0.457627 1.0 ...

1 0.144832 0.0 0.610521 0.161417 0.282526 0.28 0.041481 0.594156 0.152542 0.0 ...

2 0.223173 0.0 0.202584 0.240157 0.116394 0.36 0.050669 0.280032 0.949153 0.0 ...

3 0.239631 0.0 0.059991 0.212598 0.364790 0.36 0.737568 0.294643 0.288136 1.0 ...

4 0.460829 0.0 0.286571 0.279528 0.123979 0.32 0.127782 0.290584 0.254237 0.0 ...

5 rows × 53 columns

Check for imbalance in target class

In [52]: y_train.value_counts(normalize=True)*100

loan_status
Out[52]:
0 80.249186
1 19.750814
Name: proportion, dtype: float64

We can see a clear imbalance in the target class with 1 being ~20% and 0 being ~80%. Hence, I will use
SMOTE to fix this imbalance

In [53]: sm = SMOTE(random_state=0)
 X_train, y_train = sm.fit_resample(X_train, y_train)
y_train.value_counts(normalize=True)*100

loan_status
Out[53]:
0 50.0
1 50.0
Name: proportion, dtype: float64

6. Build Logistic Regression model

Train the model

In [54]: model = LogisticRegression(solver='lbfgs', max_iter=300)

model.fit(X_train, y_train)
y_train_pred = model.predict(X_train)

Classification metrics and confusion matrix for Training data

In [55]: print(classification_report(y_train, y_train_pred))

precision recall f1-score support

0 0.81 0.80 0.80 237605

1 0.80 0.82 0.81 237605

accuracy 0.81 475210

macro avg 0.81 0.81 0.81 475210
weighted avg 0.81 0.81 0.81 475210

In [56]: # Confusion Matrix

cm = confusion_matrix(y_train, y_train_pred)
disp = ConfusionMatrixDisplay(cm)
disp.plot()
plt.title('Confusion Matrix')
plt.show()
 Model prediction

In [57]: y_test_pred = model.predict(X_test)

Classification metrics and confusion matrix for Testing data

In [58]: print(classification_report(y_test, y_test_pred))

precision recall f1-score support

0 0.95 0.80 0.86 59619

1 0.49 0.81 0.61 14403

accuracy 0.80 74022

macro avg 0.72 0.80 0.74 74022
weighted avg 0.86 0.80 0.82 74022

In [59]: # Confusion Matrix

cm = confusion_matrix(y_test, y_test_pred)
disp = ConfusionMatrixDisplay(cm)
disp.plot()
plt.title('Confusion Matrix')
plt.show()
 Insight
Recall is high indicating that the model is able to identify 80% of the actual defaulters and 80% of non-
defaulting customers
Precision for class 1 (defaulters) is low. Of all the predicted defaulters, only 50% are actual defaulters
With this model there is a risk of denying loans to deserving customers due to low precision score for
defaulters

In [60]: feature_imp = pd.DataFrame({'Columns':X_train.columns, 'Coefficients':model.coef_[0]}).r

plt.figure(figsize=(8,8))
sns.barplot(data=feature_imp, y = 'Columns', x = 'Coefficients')
plt.title("Feature Importance for Model")
plt.yticks(fontsize=8)
plt.ylabel("Feature")
plt.tight_layout()
plt.show()
 Insight
The features zip_code_29597, zip_code_05113, zip_code_00813, annual_inc and
application_type_joint have got high positive weightage and features zip_code_86630,
zip_code_11650, zip_code_93700, dti and open_acc have got high negative weightage indicating
their major contribution towards target variable

ROC and AUC

In [61]: # Predict probabilities for the test set
probs = model.predict_proba(X_test)[:,1]

# Compute the false positive rate, true positive rate, and thresholds
fpr, tpr, thresholds = roc_curve(y_test, probs)

# Compute the area under the ROC curve

roc_auc = auc(fpr, tpr)
 # Plot the ROC curve
plt.figure()
plt.plot(fpr, tpr, color='darkorange', lw=2, label='ROC curve (area = %0.2f)' % roc_auc)
plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic')
plt.legend(loc="lower right")
plt.show()

Insight
ROC curve illustrates the trade off between TPR(True Positive Rate) and FPR(False Positive Rate) for
various thresholds
The AU-ROC value of 0.91 signifies that the model is able to differenciate well between the two classes
Let us also look at PR Curve(Precision Recall Curve)

Precision Recall Curve

In [62]: precision, recall, thr = precision_recall_curve(y_test, probs)

# Area under Precision Recall Curve

apc = average_precision_score(y_test, probs)

# Plot the precision-recall curve

plt.plot(recall, precision, marker='.', label='PR curve (area = %0.2f)' % apc)
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.title('Precision-Recall Curve')
 plt.legend(loc="lower left")
plt.show()

Insight
PR curve illustrates the trade off between Precision and Recall for various thresholds
The model has a AU-PRC value of 0.78 which is not that high. It is better than the random model which
has a AU-PRC value of 0.5.
This clearly indicates that we simply cannot conclude on the model's performance from just the ROC
curve.

Find the threshold where precision and recall meet

In [63]: plt.figure()
plt.plot(thr,precision[0:len(thr)],label='precision',color='blue')
plt.plot(thr,recall[0:len(thr)],label='recall',color='orange')
intersection_thr = thr[np.where(precision == recall)[0][0]].round(4)
plt.axvline(intersection_thr, linestyle='--', color='red')
plt.text(intersection_thr, 0.01, str(intersection_thr), ha='left', color='red')
plt.title("Precision-recall curve at different thresholds")
plt.xlabel("Threshold values")
plt.ylabel("Precision and Recall values")
plt.legend(loc="upper right")
plt.grid()
plt.show()
In [64]: y_pred = model.predict_proba(X_test)[:,1]
threshold_considered = intersection_thr
y_pred_custom = (y_pred>threshold_considered).astype('int')
print(classification_report(y_test,y_pred_custom))

precision recall f1-score support

0 0.92 0.92 0.92 59619

1 0.65 0.65 0.65 14403

accuracy 0.87 74022

macro avg 0.78 0.78 0.78 74022
weighted avg 0.87 0.87 0.87 74022

In [65]: # Confusion Matrix

cm = confusion_matrix(y_test, y_pred_custom)
disp = ConfusionMatrixDisplay(cm)
disp.plot()
plt.title('Confusion Matrix')
plt.show()
Insight
With the new threshold, precision for class 1 (defaulters) has increased at a cost of decrease in Recall
The overall F1-score and accuracy has increased

7. Insights
80% of the customers have fully paid their loan and 20% are defaulters
Loan amount and installment are highly correlated as it is obvious that high loan amount will have high
installment amount
Loan taken for short term, i.e. 3 years are most likely to be fully paid back
Most of the people have home ownership as mortgage
Suprisingly, loans which are not verified are more likely to be paid back
Loan taken as joint application type are more likely to be paid back
People with grade A are more likely to fully pay their loan
Loan taken for wedding are more likely to be paid back
People from zip code 00813, 05113 fully pay back their loans whereas people from zip code 11650,
86630, 93700 are all defaulters
The features zip_code_29597, zip_code_05113, zip_code_00813, annual_inc, application_type_joint,
zip_code_86630, zip_code_11650, zip_code_93700, dti, open_acc affected the model outcome
heavily
As per the ROC curve and AU-ROC value of 0.91, the model is able to differenciate well between the
defaulters and non-defaulters
 As per the PRC and AU-PRC value of 0.97, the model is able to return accurate results as well as return
majority of all positive results(high recall)

8. Recommendation
The bank can provide more short term loans, i.e. for 3 years, without much risk
Provide more joint loans and scrutinize more individual and direct pay application types
Analyze carefully the loan applications of customers with grades D, E, F and G. Do not provide them
loans or provide smaller loans to these customers
Reduce the loan given for small bussiness or analyze their application in detail before giving out loan to
small bussiness
Do not provide loans to customers with zip code 11650, 86630, 93700