Credit Risk Fundamentals › Limitations and Practical Considerations

The PD-LGD-EAD framework, while foundational, has important limitations that you must understand. No model perfectly captures reality, and awareness of these limitations is essential for sound risk management.

Parameter estimation uncertainty affects all components. PD estimates rely on historical data that may not reflect future conditions, particularly for low-default portfolios where statistical significance is limited. If only 1 in 500 investment-grade borrowers defaults annually, it takes many years to accumulate enough default observations for reliable estimation. LGD exhibits high variability and depends on factors that are difficult to observe in advance, such as the economic environment at the time of default and the specific circumstances of each bankruptcy. EAD for commitments and derivatives involves behavioral assumptions that may not hold under stress, as borrowers may act differently in crisis conditions than in normal times.

Correlation estimation is perhaps the weakest link in portfolio credit models. Default correlations are difficult to estimate empirically because defaults are rare events. We simply do not observe enough joint defaults under varying conditions to estimate correlations with precision. The correlations that matter most, those during systemic crises, are precisely the ones for which we have the least data, since such crises occur infrequently. The Basel framework addresses this by prescribing conservative correlation assumptions, but model risk remains substantial.

Point-in-time versus through-the-cycle tension creates practical challenges. Accounting standards (IFRS 9, CECL) require forward-looking expected credit loss provisions based on current conditions, while long-term business decisions benefit from cycle-average estimates. Institutions must maintain multiple PD frameworks for different purposes, reconciling potentially conflicting signals from each approach.

Model validation is constrained by the rarity of default events. Unlike market risk models that can be backtested against daily P&L, credit models produce predictions over long horizons with limited observable outcomes. This makes it difficult to distinguish between model accuracy and good fortune. A model that has performed well for ten years may simply have been lucky not to encounter conditions that reveal its flaws.

Despite these limitations, the EL = PD × LGD × EAD framework gives a disciplined structure for credit risk assessment. It forces explicit consideration of the main drivers of credit losses and allows consistent comparison across exposures. The next chapter extends this base to more sophisticated modeling approaches, including structural models and reduced-form credit risk models.