- The model we built here will use all possible factors to predict data on customers to find who are defaulters and non‐defaulters next month.
- The goal is to find the whether the clients are able to pay their next month credit amount.
- Identify some potential customers for the bank who can settle their credit balance.
- To determine if their customers could make the credit card payments on‐time.
- Default is the failure to pay interest or principal on a loan or credit card payment.
3. Overview
● Banking / Financial Institutes plays a significant role in providing financial service.
● To maintain the integrity, bank/institute must be careful when investing in customers to avoid
financial loss.
● Before giving credit to borrowers, the bank must come to about the potential of customers.
● The term credit scoring, determines the relation between defaulters and loan characteristics.
4. Goal
● The model we built here will use all possible factors to predict data on customers to find who are
defaulters and non-defaulters next month.
● The goal is to find the whether the clients are able to pay their next month credit amount.
● Identify some potential customers for the bank who can settle their credit balance.
● To determine if their customers could make the credit card payments on-time.
● Default is the failure to pay interest or principal on a loan or credit card payment.
5. Process Design
● The process is done by Supervised learning Algorithm.
● The idea behind using this is we have a prior knowledge on our output values.
● It acts as a guide to teach the algorithms what conclusion it should come up with.
● Common algorithms in supervised learning includes logistic regression, naive bayes, support vector
machines and decision tree classifier.
8. Dataset Overview
● Oriented: UCI Machine Learning Repository. (Link)
● Attributes: 24
● Tuples: 30,000
Customers
data
Customers
data
9. Continue…
Independent
variables:
• Customer ID
• Credit limit
• Gender
• Age
• Marital status
• Level of education
• History of their past payments made (April to
September) (g1 to g6)
• Amount of bill statement (h1 to h6)
• Amount of previous payment (i1 to i6)
Dependent
variables:
• default – A customer who will be default next month
payment (0: no, 1: yes)
10. Dataset overview
Graph shows total number
of records for defaulters
and non-defaulters.
If they would do payment
or not (yes=1 no=0) for
next month
22% - default
78% - non-default
(Screenshot: Taken during evaluation process)
11. Continue…
1. It shows count for ‘sex’ attribute
1 - Male and 2 - Female
2. It shows default count for ‘marriage’ attribute
0,3 - Others, 1 - Married, 2 - Single
(Screenshot 1,2: Taken during evaluation process)
12. Continue…
2. It shows Limit of balance with respect to age.
1. It shows count for Total ‘Credit limit’ attribute
values with respect to Number of Account
(Screenshot 1,2: Taken during evaluation process)
14. Data Preprocessing
● Data set is divided in 60:40 ratio for train and test respectively.
● ID column was dropped as its unnecessary for our modeling.
● The attribute name ‘PAY_0’ was converted to ‘PAY_1’ for naming convenience.
● Pay_0: No consumption of credit card=-2, Pay duly(paid on time)=-1, payment delay for one month=1,
payment delay for two months=2, payment delay for nine months and above=9.
● Marital Status 0 is converted to value 3 (Others).
● No Null values in dataset.
● Robust Scaler – It is used to convert variable in the same scale. Ex. in Limit Balance column there are
different range of values which are converted in proper scale.
15. ● Education ● Gender ● Marital status
Feature Engineering
Education Value
Graduate 1
University 2
High School 3
Others 4
Gender Value
Male 1
Female 2
Marital Status Value
- 0
Married 1
Single 2
Others 3
16. One-Hot Encoding
● The process in which Categorical variables are converted into a form, where we can apply algorithms to do
prediction easily.
● (n-1) features
● Ex. There are 4 features in Education Converted to 3 features (n-1)
18. Proposed Models
Logistic Regression
•It is used for Binary classification.
•Outputs have a nice probabilistic interpretation, and the algorithm can be regularized to avoid over fitting.
•In logistic regression the hypothesis is that the conditional probability p of class belongs to ”1”
•if probability is greater than threshold probability, generally 0.5, else it belongs to the class ”0”.
•Ex. Y 𝑖 = ቊ
1 , 𝑝 ≥ 0.5
0, 𝑝 < 0.5
Support Vector Machine
•Kernel: linear and enabled probability: ‘true’
•Soft-SVM search used in the process.
•Hard-SVM searches for the decision boundary that separates the training data separately with the largest
margin.
•Soft-SVM is based on the assumption that learning data is not perfectly separable.
19. Continue…
Feed Forward Neural Network
•Feed forward NN is used with Backpropagation algorithm. Below parameters decided tuned parameters
•Activation function : Rectified Linear Unit(ReLU) and Sigmoid
•SGD: adam
•Epochs: 100
•Loss function: Binary cross entropy
•Input layer : 26, Hidden layer : 2, Output: 1
•Early Stopping
Voting Classifier
•Estimators: list of classifiers
•Combined average of Logistic Regression + Decision tree Classifier + Support Vector Machine (SVM)
Naïve Bayes, KNN and Decision Tree classifier
20. Evaluation Process
Evaluation Metrics:
● Accuracy: Accuracy determine how often the model predicts default and non-default correctly.
● Precision: Precision calculates whenever our models predicts it is default how often it is correct.
● Recall: Recall regulate the actual default that the model is actually predict.
● Precision Recall Curve: PRC will display the tradeoff between precision and recall threshold.
Cross Validation:
● K Fold cross validation; k = 5
21. Confusion Metrics
True Positive – A person who is defaulter and predicted as defaulter.
True Negative – A person who is non-defaulter and predicted as non-defaulter.
False Positive – A person who is predicted defaulter is non-defaulter.
False Negative – A person who is predicted non-defaulter is defaulter.
# Non-defaulter (predicted) - 0 Defaulter (predicted) - 1
Non-defaulter (actual) - 0 TN FP
Defaulter (actual) - 1 FN TP
26. Precision Recall curve with threshold
● Precision-Recall comparison with respect to the
threshold value.
● Adjust precision and recall value by adjusting
threshold value
● We set threshold value 0.2
● False Negative value is improved by 40%.
28. Conclusion
● We investigated the data, checking for data unbalancing, visualizing the features and understanding the
relationship between different features.
● We used both train-validation split and cross-validation to evaluate the model effectiveness to predict the target
value, i.e. detecting if a credit card client will default next month.
● We then investigated five predictive models:
○ We started with Logistic Regression, Naïve bayes, SVM, KNN, Classification Tree and Feed-forward NN and
Voting classifier accuracy is almost same.
○ We choose based model based on minimum value of False Negative value.
○ This would also inform the issuer’s decisions on who to give a credit card to and what credit limit to
provide.
31. QR Code
S C A N Q R C O D E A N D A C C ES S P P T A N D C O D E ( A C C E S S T H E C O D E : H T T P : / / V A T S A L S H A H . I N / P R O J E C T S . H T M L )
32. References
● Li, Xiao-Lin, and Yu Zhong. An overview of personal credit scoring: techniques and future work. Journal:
International Journal of Intelligence Science ISSN 2163-0283. 2012.
● Federal Reserve. (2017) “Report to the Congress on the Profitability of Credit Card Operations of Depository
Institutions”.
● Taiwo Oladipupo Ayodele. (2010) “Types of Machine Learning Algorithms”, New Advances In Machine
Learning, Yagang Zhang (Ed.), Intech
● Default Credit Card Clients Dataset, https://www.kaggle.com/uciml/default-of-credit-card-clients-dataset/
● Davis J., Goadrich M. The Relationship Between Precision-Recall and ROC Curves.
ACM New York, NY, USA 2006. ISBN:1-59593-383-2.
● Arlot, Sylvain, and Alain Celisse. A survey of cross-validation procedures for model
selection. eprint arXiv:0907.4728. DOI:10.1214/09-SS054.
● https://www.analyticsvidhya.com/blog/2016/03/introduction-deep-learning-fundamentals-neural-
networks/
34. Contact Us
You can share your feedback on below Email-IDs.
● Hetarth Bhatt – hbhatt7@uwo.ca
● Khushali Patel – kpate372@uwo.ca
● Rajaraman Ganesan – rganesa@uwo.ca
● Vatsal Shah – vshh56@uwo.ca (http://vatsalshah.in)