In [29]:
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import train_test_split
loan_data = pd.read_csv('./02-loan_data.csv')
df = loan_data.copy()
del df['customer_id']
columns_to_round = ['loan_amt_outstanding', 'total_debt_outstanding', 'income']
df[columns_to_round] = df[columns_to_round].apply(lambda x: x.round(2))
In [30]:
df.dtypes
Out[30]:
credit_lines_outstanding int64 loan_amt_outstanding float64 total_debt_outstanding float64 income float64 years_employed int64 fico_score int64 default int64 dtype: object
In [31]:
df.isnull().sum()
Out[31]:
credit_lines_outstanding 0 loan_amt_outstanding 0 total_debt_outstanding 0 income 0 years_employed 0 fico_score 0 default 0 dtype: int64
Our data appears to have already been cleaned.
In [32]:
df_train.head()
Out[32]:
| loan_amt_outstanding | total_debt_outstanding | income | years_employed | fico_score | default | |
|---|---|---|---|---|---|---|
| 4495 | 3322.04 | 1783.15 | 53092.33 | 4 | 588 | 0 |
| 6470 | 2153.90 | 6022.02 | 25992.88 | 4 | 687 | 1 |
| 2221 | 3666.26 | 5253.54 | 71090.00 | 5 | 622 | 0 |
| 7686 | 4485.62 | 6886.23 | 80505.30 | 5 | 681 | 0 |
| 9419 | 7624.47 | 14607.81 | 111898.20 | 3 | 756 | 0 |
In [33]:
# pred
y_train = df_train.default.values
y_val = df_val.default.values
y_test = df_test.default.values
del df_train['default']
del df_val['default']
del df_test['default']
Model (to predict PD)¶
In [34]:
df.corr()
Out[34]:
| credit_lines_outstanding | loan_amt_outstanding | total_debt_outstanding | income | years_employed | fico_score | default | |
|---|---|---|---|---|---|---|---|
| credit_lines_outstanding | 1.000000 | 0.080249 | 0.852210 | 0.022272 | -0.087900 | -0.258177 | 0.862815 |
| loan_amt_outstanding | 0.080249 | 1.000000 | 0.397403 | 0.835815 | -0.158416 | -0.031373 | 0.098978 |
| total_debt_outstanding | 0.852210 | 0.397403 | 1.000000 | 0.394397 | -0.174353 | -0.232246 | 0.758868 |
| income | 0.022272 | 0.835815 | 0.394397 | 1.000000 | 0.001814 | -0.010528 | 0.016309 |
| years_employed | -0.087900 | -0.158416 | -0.174353 | 0.001814 | 1.000000 | 0.255873 | -0.284506 |
| fico_score | -0.258177 | -0.031373 | -0.232246 | -0.010528 | 0.255873 | 1.000000 | -0.324515 |
| default | 0.862815 | 0.098978 | 0.758868 | 0.016309 | -0.284506 | -0.324515 | 1.000000 |
In [35]:
df.hist(figsize=(10,10))
Out[35]:
array([[<Axes: title={'center': 'credit_lines_outstanding'}>,
<Axes: title={'center': 'loan_amt_outstanding'}>,
<Axes: title={'center': 'total_debt_outstanding'}>],
[<Axes: title={'center': 'income'}>,
<Axes: title={'center': 'years_employed'}>,
<Axes: title={'center': 'fico_score'}>],
[<Axes: title={'center': 'default'}>, <Axes: >, <Axes: >]],
dtype=object)
It appears that the features follow a Gaussian (normal) distribution. Let's see if we our dataset is really suitable for a Gaussian Naive Bayes. So, we check if each feature, given a specific class of the target variable (default), follows a Gaussian distribution.
In [36]:
# Histogram of features when default = 1
df[df.default==1].hist(figsize=(10,10))
Out[36]:
array([[<Axes: title={'center': 'credit_lines_outstanding'}>,
<Axes: title={'center': 'loan_amt_outstanding'}>,
<Axes: title={'center': 'total_debt_outstanding'}>],
[<Axes: title={'center': 'income'}>,
<Axes: title={'center': 'years_employed'}>,
<Axes: title={'center': 'fico_score'}>],
[<Axes: title={'center': 'default'}>, <Axes: >, <Axes: >]],
dtype=object)
In [37]:
# Histogram of features when default = 0
df[df.default==0].hist(figsize=(10,10))
Out[37]:
array([[<Axes: title={'center': 'credit_lines_outstanding'}>,
<Axes: title={'center': 'loan_amt_outstanding'}>,
<Axes: title={'center': 'total_debt_outstanding'}>],
[<Axes: title={'center': 'income'}>,
<Axes: title={'center': 'years_employed'}>,
<Axes: title={'center': 'fico_score'}>],
[<Axes: title={'center': 'default'}>, <Axes: >, <Axes: >]],
dtype=object)
In [38]:
df_full_train, df_test = train_test_split(df, test_size=0.2, random_state=1)
df_train, df_val = train_test_split(df_full_train, test_size=0.25, random_state=1)
In [39]:
# Naive Bayes
from sklearn.naive_bayes import GaussianNB
model_g = GaussianNB()
model_g.fit(df_train, y_train)
Out[39]:
GaussianNB()In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.
GaussianNB()
In [40]:
# Accuracy on the training set
(y_train== model_g.predict(df_train)).mean()
Out[40]:
0.9813333333333333
In [41]:
# Accuracy on the validation set, to see if it needs tuning
(y_val== model_g.predict(df_val)).mean()
Out[41]:
0.9775
We get $97.8\%$ accuracy on the validation set. It looks like we don't need to tune the model.
In [42]:
# Accuracy on the test set:
(y_test== model_g.predict(df_test)).mean()
Out[42]:
0.977
In [43]:
df_full = loan_data.copy()
del df_full['customer_id']
y_fulltrain = df_full.default.values
del df_full['default']
In [44]:
df_full.columns
Out[44]:
Index(['credit_lines_outstanding', 'loan_amt_outstanding',
'total_debt_outstanding', 'income', 'years_employed', 'fico_score'],
dtype='object')
In [45]:
def prompt_user():
data = {}
data['credit_lines_outstanding'] = float(input("Enter the number of credit lines outstanding: "))
data['loan_amt_outstanding'] = float(input("Enter the amount of loan outstanding: "))
data['total_debt_outstanding'] = float(input("Enter the total debt outstanding: "))
data['income'] = float(input("Enter the income: "))
data['years_employed'] = float(input("Enter the number of years employed: "))
data['fico_score'] = float(input("Enter the FICO score: "))
return data
In [48]:
# takes user input about the loan data
user_data = prompt_user()
This is the function that takes in the properties of a loan and outputs the expected loss
In [49]:
model = GaussianNB().fit(df_full, y_fulltrain)
# loan here is the loan profile dict
def expectedLoss():
loan = pd.DataFrame(prompt_user(), index=[0])
#return model.predict_proba(loan)[:,1][0]*100 + '\%'
prob_default = model.predict_proba(loan)[:,1][0]
loan_amt = float(loan.loan_amt_outstanding)
return prob_default * loan_amt * (1 - 0.1) # recovery rate = 10% (given)
In [50]:
expectedLoss()
/var/folders/h0/722z94dd3fb0pfv4wg3qmdkh0000gn/T/ipykernel_74903/2507886040.py:7: FutureWarning: Calling float on a single element Series is deprecated and will raise a TypeError in the future. Use float(ser.iloc[0]) instead loan_amt = float(loan.loan_amt_outstanding)
Out[50]:
564.8283882093126
In [ ]: