Decision TreeBasic information

• Risk Analytics – A wide domain in financial and banking industry , basically analyzing
the risk of the customer.
• Ex – 1) Whether the customer will be a transactor or revolver , etc..
2) Whether a customer be delinquent or non-delinquent.
3) Whether a customer be defaulter or non-defaulter.
• ​Transactor​ – A person who pays his due amount balance full and on time.
• ​Revolver​ – A person who pays the minimum due amount but keeps revolving his
balance and does not pay the full amount.
• ​Delinquent​ - Delinquency means that you are behind on payments , a person who
fails to pay even the minimum due amount.
• ​Defaulter​ – Once you are delinquent for a certain period of time your lender will
declare you to be in default stage.Data Summary
• For further description of the variables please refer the attached data dictionary.Exploratory
Data Analysis
• The dataset consists of 7 variables and 11548 observations.
• We are trying to analyze attributes that leads to Delinquency.
• Dataset mainly contains Categorical variables.
• In the current dataset out of total 67% of the customers are delinquent, hence we need to
analyze the attributes which can reduce this delinquency rate.
• Hence in our analysis delinquent would be the target or the response variable i.e the
Dependent variable and other variables would be independent or predictor variables.
• There are no missing values in the dataset , hence no treatment is required for the
missing values.Identifying Dependent & Independent Variables

Dependent Variable
Independent Variables
term
delinquent
gender
age
FICOCART(Classification and Regression Tree)
• The analysis to understand the major segments of delinquent customers has been
implemented through CART.
• CART decision tree has been built in R and attached is the txt file that contains the code
used for it.
• Attached is the pdf that shows the output of the tree , detailed output is given in the
subsequent slides as well.CART(Classification and Regression Tree)
• The dataset has been split into training (70%) and test datasets (30%).
• Below are the libraries used in the code and the significance for using it.
Library Significance
caTools Building the training and test dataset
Rpart Built CART tree using rpart function and setting the control
parameters using rpart.control function.
Rattle Use the fancyRpartPlot function to display the tree
ROCR Calculate KS , AUC , etc. statistics.
Ineq Calculate GINI index.CART(Classification and Regression Tree)

• Below are the control parameters that are used while building the decision tree:
Parameter Value Significance
minsplit 1000 If the node will have at least 1000
observations then only it will split.
minbucket 100 The terminal nodes should have at least 100
observations.
cp (Complexity Parameter) 0 Allowing the full tree to be grown.
xval(Cross Validation) It will cross validate 10 times.
10Decision TreeDecision Tree output explanation & Insights
• First Node shows that there are 67% of delinquent customers and 33% non- delinquent
customers.
• FICO is the first variable that is split in the decision tree hence is the most important
variable as well while building any strategy.
• The highest risk segment is of 52% which means that 52% of the delinquent customers
are being taken in that segment which is of customers having FICO < 500 and loan term of
36 months.
• The second risk segment is of 14% which means that 14% of the delinquent customers
are being taken in that segment which is of customers having FICO > 500 and gender =
Male and age > 25 and loan term = 36 months. Rest of the segments are non-risk
segments.
• Next slide shows the detailed level decision tree as well.Decision Tree
• In the decision tree output 1 denotes delinquent and 0 represents non-delinquent.Insights and
implementation strategy
• Building and using this model the below insights are generated and the following
strategies can be implemented:
Risk Segments
Segment Meaning
1) FICO < 500 and loan
term = 36 months 1) Customers having low FICO score and having loan tenure
of 36 months.
2) FICO > 500 and gender
= Male and age > 25 and
loan term = 36 months. 2) Male customers more than 25 years old having higher
FICO score and having loan tenure of 36 months.
Strategy that can be implemented to decrease the delinquency rates
1) Reject the applications of those customers who have less FICO score and whose loan
tenure is 36 months.
2) Proper verification needs to be done for male customers before approving the loan
who have higher FICO score and have a loan tenure of 36 months.Insights and implementation
strategy
Non - Risk Segments
Segment Meaning
1) FICO < 500 and loan
term = 60 months 1) Customers having low FICO score and having loan tenure
of 60 months.
2) FICO > 500 and gender
= Female. 2) Female customers with higher FICO score.
3)FICO > 500 and gender = 3) Female customers more than 25 years old having higher
Female and age = 20-25.
FICO score and having loan tenure of 36 months.
Strategy that can be implemented to increase profits and having loyal customers
1) Increase the volume of accepting loans of female customers with higher FICO score
and whose loan tenure is of 36 months.
2) Accept the applications of those customers whose loan tenure would be 60 months
though having low FICO score.
3) Give some promotional offers as well to Female customers so that they can have a
better and long term relationship with the organization.Model performance and validation
• Classification error = (627+710)/7900 = 16.9%
• Classification error gives a clear picture of the % of the observations that have been
predicted wrongly and correctly by the model.
• The table of classification error gives the below insights:
Classification As per the above output
1) True Positive 1) 4575
2) True Negative 2) 1988
3) False Positive 3) 627
4) False Negative 4) 710
TPR = 86.5%
FPR = 23.9%Model performance and validation
• KS= 62.5%
KS is higher which suggests the model holds good.
• KS (Kolmogorov Statistic) is a non-parametric test that measures the degree of
separation between the positive and negative distributions.
• Higher the KS better the model is because higher KS means better the separation is by
the model , it’s value is between 0 and 1 and has to be interpreted in %.
• AUC = 85%
AUC is higher which also suggests that the model also
holds good.
• AUC (Area Under Curve) is the % of the box that is under the ROC (Receiver Operating
Characteristics) curve which is explained in the next slide.Model performance and
validationModel performance and validation – Test Dataset
The model was tested on test dataset as well and below are the output of the validation
measures:
In line with the training dataset hence
no issue with utilizing the model with
the test dataset.
Classification error
= (293+328)/ 3648
= 17% which is inline with the
training dataset.