Probability of Default Modelling—IFRS9 (Part 2)
The likelihood that a borrower will fail to repay their obligation within a given time horizon.
Dependent Variable Definition
The target variable will have values of 0 or 1. 0 defines a good customer, and 1 refers to a bad customer.
Bad customers will be defined using the below analysis.
Roll Rate analysis: Identify the DPD bucket where the point of no return (no roll backward or no cure rate)
Vintage analysis: Track the performance of a portfolio by grouping loans into cohorts based on their origination period.
Methodologies for estimating PD
- Judgemental method
- Statistical method
Judgmental method
It’s widely used when historical data is not available (like new loan products or NTB customers)
Dependents on 5 C’s
Character, Capital, Collateral, Capacity, Condition.
Statistical method
It’s a real-time credit decision tool bank use widely. The advantage of using a statistical method is that it produces a mathematical equation, which is an automated and faster solution for making credit decisions.
This method also comes with higher accuracy, as statistical and machine learning models consider hundreds of data points to identify defaulters.
Get Sagar H M’s stories in your inbox
Join Medium for free to get updates from this writer.
Data Source for PD Modelling:
- Demographic data: age, income, employment status, marital status, no. of years at current address, no. of years at job, postal code.
- Existing Relationship with Bank: Tenure, number of products, payment performance, previous claims.
- Credit Bureau variables: Credit score, External DPD, No. of enquiry, Loans history.
Steps of PD Modelling:
- Data Preparation
- Variable selection
- Model development
- Model validation
- Calibration
- Independent validation
- Supervisory approvals
- Model Implementation: Roll out to users
- Monitering set up
- Backtesting and benchmarking
Statistical techniques used for the model:
- Logistic Regression
- Survival analysis
- Random forest
- Gradient Boosting
- Markov Chain model
- Neural network
Model Performance in PD Model:
- Discrimination Power
- Risk ordering
- Calibration
