Credit Risk Modelling using Machine Learning: A Gentle Introduction

Category

• Programming

Tags

• Decision Trees
• Logistic Regression
• Machine Learning
• R Programming
• Random Forests

Assume you are given a dataset for a large bank and you are tasked to come up with a credit risk score for each customer.You have just been briefed that you are going to work on this project and you have to come up with a prototype demonstrating how this problem could be solved.

Approach

The credit risk scoring is a very complicated process with a lot of due diligence on data, model reviews internal controls and sign offs. As a first step you could follow the steps outlined below with the accompanying code to create a straw man version of your approach.

The first step in building your prototype will be obtaining a sample dataset and performing high level analysis on it.

#setting up the data and performing high level analysis#
########################################################
#downloading the data
#https://raw.githubusercontent.com/obaidpervaizgill/CreditRiskModelling/master/credit.csv

#loading data
credit <- read.csv("credit.csv")

#identifying the structure of variables
str(credit)

#getting summary of the variables
summary(credit) 

#getting the column names
colnames(credit)
#[1] "checking_balance"     "months_loan_duration" "credit_history"       "purpose"              "amount"               "savings_balance"     
#[7] "employment_duration"  "percent_of_income"    "years_at_residence"   "age"                  "other_credit"         "housing"             
#[13] "existing_loans_count" "job"  

#tabulating dependent variables
table(credit$default)

#No missing values in the data
#Note : I would have used "mice" package in R to impute missing values if there were any

#Normalizing or standardizing data
#Note : I would have scaled the variables using standardization or minmax normalization, but I havent done this here! 

#Removing correlated features
#Note : I would have removed correlated feature based on an 80 percent correlation rule in the correlation matrix

#spliting data into test and train
library(caTools)
split <- sample.split(credit$default, SplitRatio = 0.70)
train <- subset(cbind(credit,split), cbind(credit,split)$split == TRUE)
test <- subset(cbind(credit,split), cbind(credit,split)$split == FALSE)

#checking proportions across train and test
prop.table(table(train$default))
prop.table(table(test$default))

The second step in your prototype will be to train an explainable model, such as a logistic regression model so that you can identify and explain the driving variables.

#training a model using logistic regression#
############################################

#training a model
creditLogReg <- glm(train$default ~ ., data = train[,c(-17,-18)], family = "binomial" ) #removing split feature and dependent variable
summary(creditLogReg) #summary of the model output
#Note: In theory I should rerun the model removing the non-significant features but since I want to demonstrate multiple model usage I would let it slide

#predicing on test data
predCreditLogReg  0.5) 
#Note: we want our model to be optimally sensitive hence we use 0.5 as the threshold, redudcing the threshold will make the model more sensitive

#computing the accuracy of the model
accuracyCreditLogReg  0.5))[1,1]) + (as.matrix(table(test$default, predCreditLogReg > 0.5))[2,2]))/nrow(test)

#computing the baseline model for comparison
baseLineAccuracy <- max(table(test$default))/nrow(test)

print(accuracyCreditLogReg) 
print(baseLineAccuracy)
#Note : Our simple logistic regression model beats the baseline model

#assesing the robustness of model
library(ROCR)
rocrPredCreditLogReg <- prediction(predCreditLogReg,test$default) 
areaUnderCurve <- as.numeric(performance(rocrPredCreditLogReg, "auc")@y.values) #out of sample auc
print(areaUnderCurve) 
#Note : Closer to 1 is better, 0.78 here is not bad for a first model

The third step in your prototype will be to train an more complicated model to assess if you can improve over your explainable model through additional tuning as well.

#training a model using decision trees#
#######################################
library("rpart")
library("rpart.plot")

#training a model
creditDecTree <- rpart(train$default ~ ., data = train[,c(-17,-18)], method = "class", minbucket = 1) #min bucket is minimum number of observations in a terminal nore
summary(creditDecTree) #summary of the model output

#plotting a decision tree to see splits
prp(creditDecTree)

#predicting on test data
predictCreditDecTree <- predict(creditDecTree, newdata = test[,c(-17,-18)], type = "class") #getting classes rather than probability

#computing the accuracy of the model
table(test$default,predictCreditDecTree) #since we dont have a probability here so we dont set a threshold

accuracyCreditDecTree <- ((as.matrix(table(test$default, predictCreditDecTree))[1,1]) + (as.matrix(table(test$default, predictCreditDecTree))[2,2]))/nrow(test)

#computing the baseline model for comparison
baseLineAccuracy <- max(table(test$default))/nrow(test)

print(accuracyCreditDecTree)
print(baseLineAccuracy)
#Note: Our decision tree model beats the basline model in terms of accuracy

#assesing the robustness of model
library(ROCR)
rocrPredictCreditDecTree <- prediction((predict(creditDecTree, newdata = test[,c(-17,-18)])[,2]), test$default) #getting probability and then picking predicted class
areaUnderCurve <- as.numeric(performance(rocrPredictCreditDecTree, "auc")@y.values) #out of sample auc
print(areaUnderCurve) 

#tuning a model using decision trees#
#####################################
library(caret)

#tuning for complexity parameter, this penalizes model complexity and avoids overfitting
tuneGridDecTree <- expand.grid(.cp=seq(0.01,0.5,0.01))

#creating a list of parameters to be passed onto the model
fitControlDecTree <- trainControl(method = "cv", number = 10)


tunedCreditDecTree <- train(train$default ~., data = train[,c(-17,-18)],
                            method = "rpart",
                            trControl = fitControlDecTree,
                            tuneGrid = tuneGridDecTree)

tunedPredictCreditDecTree <- predict(tunedCreditDecTree, newdata=test[,c(-17,-18)], type="raw")

#copmuting the accuracy of the model
table(test$default,tunedPredictCreditDecTree) #since we dont have a probability here so we dont set a threshold

accuracyTunedCreditDecTree <- ((as.matrix(table(test$default, tunedPredictCreditDecTree))[1,1]) + (as.matrix(table(test$default, tunedPredictCreditDecTree))[2,2]))/nrow(test)

The final step in your prototype will be to train using a highly robust and more black box model to assess if you can improve over your existing approaches, to see if it is worthwhile to pursue this path.

#training a model using random forest#
#######################################
library(randomForest)

#training a model
creditRandFor <- randomForest(as.factor(train$default) ~., data = train[,c(-17,-18)],nodesize =25, ntree = 200)
summary(creditRandFor) #summary of the model output

#identifying the most important variables based on mean gini decrease
varImpPlot(creditRandFor)
#Note : Show how each split result in low impurities or increased homogeneity

#predicting on test data
predictCreditRandFor <- predict(creditRandFor, newdata = test[,c(-17,-18)])

#computing the accuracy of the model
table(test$default,predictCreditRandFor) #since we dont have a probability here so we dont set a threshold

accuracyCreditRandFor <- ((as.matrix(table(test$default, predictCreditRandFor))[1,1]) + (as.matrix(table(test$default, predictCreditRandFor))[2,2]))/nrow(test)

#computing the baseline model for comparison
baseLineAccuracy <- max(table(test$default))/nrow(test)

print(accuracyCreditRandFor)
print(baseLineAccuracy)
#Note: Our random forest model beats the basline model in terms of accuracy

#assesing the robustness of model
library(ROCR)
rocrPredictCreditRandFor <- prediction((predict(creditRandFor, newdata = test[,c(-17,-18)], type = "prob")[,2]), test$default) #getting probability and then picking predicted class
areaUnderCurve <- as.numeric(performance(rocrPredictCreditRandFor, "auc")@y.values) #out of sample auc
print(areaUnderCurve)
#Note : Very high area under the curve but slighltly less than logistic regression
#Note : Very high accuracy as good as logistic regression

#tuning a model using random forest#
#######################################
#Note : We can tune it using tuneRF package but repeated cross validation using caret produces much better results
library(caret)

#tuning for mtry, this the number of variables randomly sampled for splits
tuneGridRandFor <- expand.grid(.mtry=c(1:sqrt(ncol(train[,c(-17,-18)]))))

#creating a list of parameters to be passed onto the model
fitControlRandFor <- trainControl(method = "repeatedcv", 
                             number = 5, repeats = 3,
                             #fivefold cross validation repeated 10 times
                             classProbs = TRUE,
                             summaryFunction = twoClassSummary) 

tunedCreditRandFor <- train(as.factor(train$default) ~., data = train[,c(-17,-18)],
                            method = "rf",
                            trControl = fitControlRandFor,
                            verbose = TRUE,
                            metric = "ROC",
                            tuneGrid = data.frame(tuneGridRandFor),
                            importance = TRUE)

tunedPredictCreditRandFor <- predict(tunedCreditRandFor, newdata = test[,c(-17,-18)])

#computing the accuracy of the model
table(test$default,tunedPredictCreditRandFor) #since we dont have a probability here so we dont set a threshold

accuracyTunedCreditRandFor <- ((as.matrix(table(test$default, tunedPredictCreditRandFor))[1,1]) + (as.matrix(table(test$default, tunedPredictCreditRandFor))[2,2]))/nrow(test)

Conclusion

Depending on the problem you are trying to solve, you could pick a model that serves your case, simplest is always the better unless the complicated one is significantly better. Also note that while there may be a temptation to jump into models, most improvement in model performance come from data wrangling and creating new features for your models.

• How to manage credentials and secrets safely in R
• Keras implementation and pushing it to dockerhub
• Multiclass classification using Tensorflow
• Understanding Customer Attrition Using Categorical Features in Python
• KNNImputer for Missing Value Imputation in Python using scikit-learn