Credit_Risk_Modelling

Scorecard Model

Monotonic binning was done to calculate WOE(Weight of Evidence) values for the continuous variables using Pymob library. In order to understand the predictive power and each variable and later Information value(IV) was also calculated by summing up the WOE values of all the bins for feature selsction purpose.

How to calculate WOE and IV:

• WOE = ln(%Y / %Obs)
• IV = ∑((%Y- %Obs) * WOE)

Steps to calculate WOE and IV:

• Split Continuous Independent Variable (x) into 10 or 20 buckets (call variable 'rank'). If you have categorical independent variable, you don't need to split as they are already categorized.
• Calculate min and max of x by rank. Compute sum of target variable (y) by rank. Let’s name it as ‘SumY’.
• Calculate total count and % of observations falling in each bucket of rank variable.
• Calculate %Y which is calculated by SumY / ∑SumY.
• WOE = ln(%Y / %Obs). %Obs represents percentage of observations (calculated in step 3)
• IV = ∑((%Y- %Obs) * WOE)

Advantages of using WOE values

• It creates a separate bin for missing values ,i.e, imputation is not required.
• It can easily treat outliers as the values get binned.

Multicollinearity Detection

Variance Inflation Factor (VIF) was used to detect the multicollinearity among the independent variables. Conclusion: Multicollinearity did not exist among continuous variables.

Logistic regression model was used to calculate the probbability of the target variable which were later used to calculate the final Credit Score

Tree Models

In order to implement Tree Models like Decision Tree and Random Forest, separate data preprocessing was done, since WOE values can't be used with such models.

Missing value imputation

3000 records were having missing values of loan_int_rate, since loan_int_rate was dependent on loan_grade, so for each loan_grade median loan_int_rate was calculated and values were imputed accordingly. Other missing value records were just dropped.

Next Steps

Feature Encoding, Train Test Split and Hyperparameter Tuning using GridSearchCV was done on the dataset and model.

Performance Metrics

• KS statistic: It basically measures the degree of separation between the CDF’s of two classes.
• ROC AUC: It just gives the measurement of the area under the ROC curve made from predictions.
• Recall Score: It basically tells proportion of correctly classified positives out of total positives.
• F1 Score: It is just the weighted average of precision and recall.
• Capture rate: It is the proportion of actual positives in each bin. Predicted probabilities are sorted and divided into 10 bins. A Good model shows staircase pattern from top to bottom in case of capture rate

Results

Logistic Regression (Using WOE)

Metric	Train	Test
KS	60.6	60.8
ROC AUC	0.74	0.73
Capture Rate (10%)	37.07	36.8
Capture Rate (20%)	62.97	62.1
Recall	0.534	0.52
F1 Score	0.62	0.61

Decision Tree

Metric	Train	Test
KS	68.7	66.5
ROC AUC	0.85	0.84
Capture Rate (10%)	46.33	46.3
Capture Rate (20%)	73.9	72.2
Recall	0.76	0.75
F1 Score	0.75	0.74

Random Forest

Metric	Train	Test
KS	64.2	62.8
ROC AUC	0.788	0.79
Capture Rate (10%)	43.65	44
Capture Rate (20%)	68.78	68.1
Recall	0.615	0.62
F1 Score	0.7	0.7