Model Development: Building Risk Models

Developing an IRB-compliant PD model follows a structured process: First, data quality and representativeness of historical default time series are assessed, typically spanning at least five years. This is followed by risk driver selection through univariate and multivariate analyses. Modeling typically uses logistic regression, supplemented by gradient boosting for nonlinear relationships. The model is then calibrated to deliver point-in-time or through-the-cycle estimates. Before submission to the supervisory authority, the model undergoes independent validation including backtesting, discriminatory power analysis (Gini/AUROC) and calibration tests.

LGD models (Loss Given Default) estimate the loss rate upon default, incorporating collateral values, recovery proceeds and resolution timelines. They often use two-stage models: first classifying between total loss and partial recovery, then estimating the recovery rate via regression. EAD models (Exposure at Default) forecast the exposure amount at the point of default, considering credit line utilization and conversion factors. Unlike PD models that deliver point estimates of default probability, LGD and EAD models require distribution modeling and are more dependent on macroeconomic downturn scenarios.

Regulatory authorities require formal approval for IRB models based on CRR/CRD requirements. Key requirements include: representative data foundations with sufficient observation periods, transparent methodology with documented assumptions, regular independent validation by a unit separate from development, ongoing performance monitoring with defined thresholds and a model risk management framework. Institutions must also comply with EBA guidelines on PD and LGD estimation and demonstrate the use test, meaning actual use of models in credit decisions and risk management.

Integrating machine learning into IRB models requires a hybrid approach ensuring interpretability and regulatory acceptance. Proven methods include: gradient boosting (XGBoost, LightGBM) as challenger models for benchmarking, SHAP values and LIME for explaining nonlinear predictions, ML-based feature engineering to identify new risk drivers that feed into interpretable models, and ensemble methods combining logistic regression with tree-based approaches. Thorough documentation following SR 11‑7 and EBA requirements is essential for regulatory approval.

Market risk model development covers Value-at-Risk and Expected Shortfall models accounting for nonlinear market dynamics. Methodologies include parametric approaches (variance-covariance), historical simulation and Monte Carlo simulation. Advanced models use GARCH processes for time-varying volatilities, regime-switching models for different market phases, copula methods for complex dependency structures and Extreme Value Theory for tail risks. In the FRTB context, we develop both standardized approach and IMA models with risk factor eligibility tests and P&L attribution.

Data quality is the foundation of every reliable risk model. We systematically verify: completeness of historical default time series spanning five to ten years, sample representativeness across all portfolio segments, consistency of definitions across source systems, correct default definition per CRR Article 178, and appropriate risk driver granularity. Automated data validation routines identify outliers and inconsistencies. A data governance framework with defined data ownership structures and regular quality reviews ensures ongoing data quality.

External consulting for model development provides methodological breadth from numerous projects across different institutions, current knowledge of regulatory developments such as CRR III and EBA guidelines, independent perspective on existing model landscapes and proven methodologies that shorten development timelines. ADVISORI combines over eleven years of risk modeling experience with expertise from more than 520 projects. Our consultants understand both supervisory requirements and the practical challenges of integrating models into existing IT infrastructures and risk management processes.