IRB Credit Risk Modelling
Understanding the Internal Ratings-Based Approach under Basel
Framework
Presenter - Debangshu
Date - 20/09/2025
Key topics to be covered
● Introduction to Basel Guidelines ● Pre-Implementation Model Testing
● Key Components of Credit Risk Modelling ● Post-Implementation Model Monitoring
● IRB Framework ● Key challenges
● Data preparation ● Future Trends
● Model Methodology
Introduction to Basel Guidelines
Basel II - 2004 Basel III - 2010 Basel IV - 2017
● Improve risk sensitivity by ● Enhancement in capital ● More risk-sensitive
redefining risk and liquidity standard. standardized approach.
management standards.
● Pillar 1 - Minimum capital ● CET1 capital requirement ● Restrictions on using
requirement for credit, raised to 4.5% of RWA A-IRB models for large
market and operational from 2%, capital buffer to corporate, sovereign and
risk. be kept. banks.
● Pillar 2 - Supervisory ● 3% leverage ratio to be ● 72.5% cap introduced for
review process on risk
assessment and capital maintained to prevent A-IRB model for RWA.
adequacy framework. excessive borrowing.
● Stricter boundary between
● Pillar 3 - Banks need to ● Liquid assets to be kept for banking and trading book,
disclose risk exposures, 30 days stress, 1 year introduction of FRTB.
capital structure, and risk
assessment methods. stable funding needed.
Key Components of IRB credit risk modelling
● PD : Likelihood borrower defaults within 1 year.
● LGD : % of exposure lost if borrower defaults.
● EAD : Outstanding exposure at time of default.
● M : Effective maturity of loan/exposure(Per guideline, always between 1-5 years)
● Regulatory capital requirement : This is defined as the formula below -
where G is inverse normal cdf, R is asset correlation, b is maturity adjustment function. Using
this formula, conservative capital is calculated at 99.9% confidence level.
● Two levels: Foundation IRB (regulators provide LGD, EAD) vs Advanced IRB (bank estimates
all inputs).
IRB Framework
Standardized Approach Internal-Rating Based Approach
● Fixed risk weights ● Bank estimates PD, LGD, EAD, M
● Based on external ratings ● Risk-sensitive
● Simple & transparent ● Complex, model-driven
● Suitable for small/medium banks ● Suitable for large banks
● Less risk-sensitive, rating agency ● Costly, huge inconsistency across
reliance banks, risk of tweaking the model
assumption, data or methodology to
report lower RWA.
Data Preparation
Loan-Level
Data Calculate
Financial
Ratios
Default data
Data Sources
EDA, Data
Perform Data
Cleaning &
Checks
Standardize
Macroeconom
ic data
Derive default
definition
Develop the
External data
model
Model Methodology : Segmentation of Assets
IRB Approach
A-IRB and
Retail IRB
F-IRB
Residential Mortgage Institution Exposure
Revolving credit retail exposure Corporate Exposure
Other exposures Large financial Exposure
Model Methodology : PD/LGD
Input Data
Processed Data
Calculate PIT CDR via Calculate PIT Loss Severity via
logistic/linear regression model linear/fractional logit model
Convert PIT Loss Severity to
Convert PIT CDR to TTC PD
downturn LGD
PD Segmentation LGD Segmentation
Pre-Implementation Model Testing
● Diagnostic testing for estimation method to ensure the models are suitable for usage.
● Perform backtesting on in-sample and out-of-sample data separately.
● Sensitivity analysis on key categorical and numerical risk parameters.
● Compare with the capital required per standardized approach and calculate regulatory amount
of capital saved per IRB approach.
Post-Implementation Model Monitoring
● Check PD and LGD buckets PSI(Population Stability Index)
● Check data relevancy by comparing the distribution of key variables in the data.
● Check if PD or LGD is underestimated in consecutive quarters in recent times.
● If either of these tests fail, the model needs to be re-calibrated with latest data.
Key Challenges
● Data Quality & Availability
● Mixing internal and external data without any bias
● Risk Estimates sensitive to economic cycle
● Huge cost associated with recalibration of the model
Future Trends
● Machine Learning methods(Xgboost etc) for PD estimation.
● Scenario-based LGD modelling just like other time-sensitive scenarios.
● Climate risk integration
Thank you all!
