Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • 357 views

    Retail Credit Scoring

    Uploaded by

    Rohan Mishra
    1. The document describes using logistic regression and classification trees to build models for predicting credit default using customer data and compares their performance. 2. Logistic regression identified monthly income as not significant, and classification trees had a higher correct prediction rate of 73% compared to 57% for logistic regression. 3. Based on the results, classification trees provide a better approach for this problem compared to logistic regression.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Retail Credit Scoring

    Uploaded by

    Rohan Mishra
    357 views9 pages
    1. The document describes using logistic regression and classification trees to build models for predicting credit default using customer data and compares their performance. 2. Logistic regression identified monthly income as not significant, and classification trees had a higher correct prediction rate of 73% compared to 57% for logistic regression. 3. Based on the results, classification trees provide a better approach for this problem compared to logistic regression.

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    1. The document describes using logistic regression and classification trees to build models for predicting credit default using customer data and compares their performance. 2. Logistic regression identified monthly income as not significant, and classification trees had a higher correct prediction rate of 73% compared to 57% for logistic regression. 3. Based on the results, classification trees provide a better approach for this problem compared to logistic regression.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    357 views9 pages

    Retail Credit Scoring

    Uploaded by

    Rohan Mishra
    1. The document describes using logistic regression and classification trees to build models for predicting credit default using customer data and compares their performance. 2. Logistic regression identified monthly income as not significant, and classification trees had a higher correct prediction rate of 73% compared to 57% for logistic regression. 3. Based on the results, classification trees provide a better approach for this problem compared to logistic regression.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 9

    Retail Credit Scoring for

    Auto Finance Ltd.

    Group-7
    Section A

    DAUR Assignment 3
    Method:
    - For the ease calculation and clean the data to remove factors which might not contribute
    to the analysis, we have removed several fields from the data frame
    - We have used logistics regression to build the model for the classification of defaulters
    (Only 2 classes: Defaulters and Non-defaulters)
    - Using the cross validation approach post building the model we have tested the training
    data on test data (since it is mentioned to split the data into two half, we have used equal
    number (14453) of samples for test and training set)
    - We have used classification tree approach to classify the data post this
    - For above methods, we have built confusion matrix and determined true and false
    prediction as a percentage

    Analysis
    Since, Monthly income in thousands is coming out be not significant in the model, we have
    removed this field and run the logistics regression one more time, without “MTHINCTH”

    Odds Ratio, ROC, Confusion matrix, Hosmer-Lemeshow Test


    Output
    Test and training set

    Output

    ROC plot

    fraction of days for which the prediction was correct: 0.573


    misclassification rate: 0.43
    Classification Tree

    Output

    Tree

    fraction of days for which the prediction was correct: 0.73


    misclassification rate: 0.27
    Conclusion

    1. “MTHINCTH” doesn’t have a significant impact on the credit defaulting


    2. On the basis of results from the confusion matrix (the classification & misclassification
    rate) we can say that the true classification value has increased using Tree model, so
    classification tree is be a better approach than logistic regression for this case
    3. The analysis can be further extended by using 3 classes of defaulters
    Code
    # Load the R library "ISLR"

    library(ISLR)

    library(tree)

    cr<-read.csv("Downloads/DAUR/Retail credit.csv",header=T)

    detach(cr)

    attach(cr)

    names(cr)

    dim(cr)

    cr <- subset( cr, select = -c(1:3,21) )

    names(cr)

    dim(cr)

    mod_1=glm(DefaulterFlag~.-DefaulterType, data=cr, family=binomial)

    summary(mod_1)

    mod_2=glm(DefaulterFlag~. -MTHINCTH-DefaulterType, data=cr,family=binomial)

    summary(mod_2)

    require(MASS)

    exp(cbind(Odds_Ratio=coef(mod_2), confint(mod_2)))

    # Using the "predict()" function to obtain the probabilities of the form "P(Y=1|X)"

    # The "type=response" ensures the output of the form "P(Y=1|X)", rather than other information
    such as the logit

    mod_2.probs=predict(mod_2,type="response")

    # Optimal Cut-off Point

    library(Epi)

    # ROC Plot

    library(pROC)

    R=roc(DefaulterFlag,mod_2.probs)

    plot(roc(DefaulterFlag,mod_2.probs),col="blue",legacy.axes = TRUE)

    coords(R, "best", ret = "threshold")


    # Conversion of probabilities into class labels

    mod_2.pred=rep("No",28906)

    mod_2.pred[mod_2.probs>.6845515]="Yes"

    mod_train.predict=ifelse(mod_2.pred=="Yes",1,0)

    # Creating Confusion Matrix to check how many observations are correctly or incorrectly classified

    table(mod_train.predict,DefaulterFlag)

    # Calculating the fraction of days for which the prediction was correct

    mean(mod_train.predict==DefaulterFlag)

    # Calculating the misclassification rate

    mean(mod_train.predict!=DefaulterFlag)

    # Hosmer-Lemeshow Test for checking the model

    default_new<- ifelse(DefaulterFlag=="Yes", 1, 0)

    library(ResourceSelection)

    hoslem.test(default_new, fitted(mod_2))

    set.seed(1)

    # Creating a hold-out data set

    # Creating the Training Data Set

    train=sample(1:28906,14453)

    # Training Data

    cr_train=cr[train,]

    dim(cr_train)

    # Test Data

    cr_test=cr[-train,]
    dim(cr_test)

    # Creating an array of "Status" variable for Training Data

    df_train=DefaulterFlag[train]

    # Creating an array of "Status" variable for Test Data

    df_test=DefaulterFlag[-train]

    # Fitting a new logistic regression model based on the training data set

    mod_train=glm(DefaulterFlag~.-DefaulterType

    ,data=cr, subset=train,family=binomial)

    # Predicting "P(Y=1|X)" for the training data set based on the fitted logistic regression model

    mod_probs_train=predict(mod_train,cr_test,type="response")

    names(cr_train)

    dim(cr_test)

    names(cr_test)

    # ROC Plot

    R=roc(df_train,mod_probs_train)

    plot(roc(df_train,mod_probs_train),col="blue",legacy.axes = TRUE)

    coords(R, "best", ret = "threshold")

    # Conversion of probabilities into class labels

    mod_pred_train=rep("No",14453)

    mod_pred_train[mod_probs_train>.644594]="Yes"

    mod_train.predict=ifelse(mod_pred_train=="Yes",1,0)

    # Creating Confusion Matrix to check how many observations are correctly or incorrectly classified
    table(mod_train.predict,df_train)

    # Calculating the fraction of days for which the prediction was correct

    mean(mod_train.predict==df_train)

    # Calculating the misclassification rate

    mean(mod_train.predict!=df_train)

    dim(cr_train)

    require(rpart.plot)

    require(rpart)

    r <- rpart(DefaulterFlag~.-DefaulterType,data=cr, subset=train, method = "class")

    rpart.plot(r, type=3, extra=101, fallen.leaves = T)

    p <-predict(r,cr_test,type = "class")

    # Confusion Matrix

    table(df_test,p)

    # Calculating the fraction of days for which the prediction was correct

    mean(df_test==p)

    # Calculating the misclassification rate

    mean(df_test!=p)

    You might also like