Task - Case: Acceptance of Loan Offers Universal Bank is a relatively young bank growing rapidly in terms of overall customer acquisition. The majority of these customers are liability customers (depositors) with varying sizes of relationship with the bank. The customer base of asset customers (borrowers) is quite small, and the bank is interested in expanding this base rapidly to bring in more loan business. In particular, it wants to explore ways of converting its liability customers to personal loan customers (while retaining them as depositors).
A campaign that the bank ran last year for liability customers showed a healthy conversion rate of over 9%. This has encouraged the retail marketing department to devise smarter campaigns with better target marketing. Your goal is to model the previous campaign’s customer behavior to analyze what combination of factors make a customer more likely to accept a personal loan. This will serve as the basis for the design of a new campaign.
The file UniversalBank.xls contains data on 5000 liability customers of Universal Bank who were targeted in the previous personal loan campaign. The data include customer demographic information (age, income etc.), the customer’s relationship with the bank (mortgage, securities account etc.), and the customer response to the last campaign (Personal Loan). A 1 in the Personal Loan column indicates the loan offer was accepted. The descriptions of the variables are in the Description worksheet in the file. (Read the descriptions to get a better idea about the variables.) Use the dataset to answer the following questions. (Hint: In the dataset, since the row containing attribute names is not the first row of the worksheet, inspect your dataset, set your dataframes correctly, or adjust your data set.)
1. Create a scatterplot of Age vs. Income, using color to differentiate customers who accepted the loan and those who did not. Which variable (i.e., age or income) appears to be potentially more useful in classifying customers? Explain.
import pandas as pd
from matplotlib import pyplot as plt
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report, confusion_matrix, accuracy_score
#Standard import functions for pythond coding
xls = pd.ExcelFile("/Users/Charles/Desktop/GMU/GBUS738/UniversalBank.xlsx")
df = pd.read_excel(xls, 'Data')
#Imports and reads our data files, also denotes file as df / xls for remainder of code
len(df[df['Personal Loan']==1])
df1=df[df['Personal Loan']==0]
df2=df[df['Personal Loan']==1]
#Denotes those who did NOT accept a personal loan as a 0, and those who did as a 1. Utilized for sorting later
plt.scatter('Age','Income',label='Loan not accepted',data=df1)
plt.scatter('Age','Income',label='Loan accepted',data=df2)
#Plots 2 independent charts on 1 graph. 1 chart for those who did not accept (0 as denoted in line 44), and those who did (1 in line 45)
plt.xlabel('Age in years')
plt.ylabel('Income in $000')
#adds X and Y axis labels
plt.title('Loan Acceptance classification based on age and income')
#adds plot title
plt.legend()
#adds legend
plt.show()
#returns plot Based on the information from UniversalBank, and our scatterplot, it does appear there is a stronger correlation between Income and loan acceptance, as compared to age and loan acceptance. Based on visually looking at our scatter plot, we can see that plot points listed in orange represent an individual taking a given loan, where blue represents those not taking the loan. While we can see there is a spread across all age groups of those taking or not taking loans, we can see a stronger correlation between income and those who accepted loans. It does appear that those individuals who have an income of >$100,000, are more strongly correlated with loan acceptance, as compared to those who have an income of <$100,000. Based on the graph, those who have an income sub $100,000, had very few loans, where as the majority of loans were above the income range of $100,000.
2. Build a logistic regression model to classify customers into those who are likely to accept personal loan offer and those who are not. Use all the available variables as predictors except ID and ZIP Code. (Hint: Since the Logistic Regression operator expects binominal or polynominal target variables, if the target variable is numeric, you will have to convert it to binominal by using the Numerical to Binominal operator.)
- a. Evaluate the overall predictive accuracy of the model as well as the accuracy of each class using appropriate metrics.
x=df[['Age','Experience','Income', 'Family','CCAvg','Education','Mortgage','Securities Account','CD Account','Online','CreditCard']].values
y=df['Personal Loan'].values
y=df['Personal Loan']
x_train, x_test, y_train, y_test = train_test_split(x,y, test_size=0.7)
model = LogisticRegression()
mfit=model.fit(x_train, y_train)
y_pred=mfit.predict(x_test)
x=0
for i in y_test:
if i ==0:
x=x+1
else:
pass
print(x)
#tn, fp, fn, tp = confusion_matrix(np.array(y_test), np.array(y_pred), labels=[0,1]).ravel()
#print(tn)
#print(fp)
#print(fn)
#print(tp)
#score=accuracy_score(y_test, y_pred)
#print(score)
#we can remove 90-98 and just use line 99 for simpler code? thoughts?
print(confusion_matrix(y_true=y_test,y_pred=y_pred))Confusion Matrix
[[1354 18]
[ 53 75]]
Utilizing the results from our Confusion Matrix, we can generate an overall accuracy report, as well as accuracy of each class:
accuracy=(1353+75)/(1354+18+53+75)
print(accuracy)
accuracy_class0=1354/(1354+19)
print(accuracy_class0)
accuracy_class1=75/(75+53)
print(accuracy_class1)Overall Accuracy = 95%
Accuracy of predicted class 0 = 99%
Accuracy of predicted class 1 = 60%
- b. What was the default cutoff probability used to generate the classifications?
.5
- c. Assuming that the dataset contains a representative sample of the liability customers of the bank, if you target 100 customers randomly (i.e., without the aid of any predictive model), how many of them would potentially accept a personal loan offer?
random_subset = df.sample(n=100)
#calls for a random sample of 100
random_subset=pd.DataFrame(random_subset)
#puts name to random subset
x1=random_subset.drop(['ID'],axis =1)
x1=x1.drop(['ZIP Code'],axis =1)
x1=x1.drop(['Personal Loan'], axis =1)
#calls to drop ID and Zip code from the column
y1=random_subset['Personal Loan']
#calls for random sample of our indepent variable
y1_pred=mfit.predict(x1)
print(confusion_matrix(y1, y1_pred))7 Total people would accept the loan offer
- d. Now if you use your model in Part (2) to select 100 customers with the highest probability of loan acceptance, how many of them would potentially accept a personal loan offer? (Hint: Revise the process from Part (2) to generate a lift chart.)
clf = LogisticRegression(
penalty='l2', dual=False,
tol=0.0001, C=1.0, fit_intercept=True, intercept_scaling=1,
class_weight=None, random_state=None, solver='lbfgs',
max_iter=10000, multi_class='auto', verbose=0,
warm_start=False, n_jobs=None, l1_ratio=None)
x=df[['Age','Experience','Income', 'Family','CCAvg','Education','Mortgage','Securities Account','CD Account','Online','CreditCard']].values
y=df['Personal Loan'].values
x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.3,train_size=0.7,random_state=0)
clf.fit(x_train,y_train)
y_proba=clf.predict_proba(x_test)
print(y_proba)
probability=y_proba[:,1]
probability=probability.reshape(-1,1)
y_test=y_test.reshape(-1,1)
probability_test=np.append(probability,y_test,axis=1)
probability_test=probability_test[probability_test[:, 0].argsort()[::-1]]
probability_test=probability_test[0:100,:]
print(probability_test)
Accepted=probability_test[:,1].sum()
print(Accepted)79%
- e. What percentage of customers who accepted the loan were incorrectly classified by the model in Part (2)?
Incorrect_classified=53/(53+75)
print(Incorrect_classified)41%
3. Suppose the bank is interested in improving the accuracy of identifying the potential positive responders, i.e., those who would accept the loan offer. Create a new process to develop a logistic regression model to classify customers into those who are likely to accept personal loan and those who are not using all the available variables—except ID and ZIP Code — as predictors. However, this time modify the cutoff probability in such a way that the accuracy of identifying the positive responders is at least 70%. Compare the predictive accuracy of this revised model with that of the model developed in Part (2). (Again, try to be analytical instead of just noting the numbers)
y_proba1=clf.predict_proba(x_test)
probability1=y_proba1[:,1]
probability1=probability1.reshape(-1,1)
predicted=[1 if i>0.3 else 0 for i in probability1]
predicted=np.asarray(predicted)
predicted=predicted.reshape(-1,1)
y_test=y_test.reshape(-1,1)
probability_test1=np.append(probability,y_test, axis=1)
probability_test1=np.append(probability_test1,predicted,axis=1)
probability_test1=probability_test1[probability_test1[:, 0].argsort()[::-1]]
positive = probability_test1[probability_test1[:,1]==1, :]
true=(positive[:,1]==positive[:,2]).sum()
print(true/positive.shape[0])75%
When we adjusted the test cutoff in line 214, to .3 or 70% we have increased our models accuracy by 15%, up from 60% to 75%.
4. Aside from the problem of predicting the likelihood of accepting loan offers, think of two other business problems where logistic regressions can be utilized for predictive modeling. For each problem, identify a target variable and four possible predictor variables.
Logistic regression could be utilized in many busniess applications. One example of where Logistic regression could be utilized is in the auto insurance industry. The target variable would be the monthly premium for a given consumer. Predictor variables could include: Prior tickets, Age, Credit Score, Gender, Vehical Price, Time with company, or Average distance driven per year. Another example of where logistic regression could be utilized is in predicting the probability of a presidential election. The predictor variables in this case could include: Incombant party, Amount of money spent on election, Average hours per week spent campainging, or Amount of money spent on negative advertisement. While both of these models are very different, both examples could be utilized for logistic regression modeling.