Margin of Conservatism Explained: Reducing Capital Inflation in IRB Models

Zanders guided research by students at Erasmus University Rotterdam into emerging modeling techniques for the Margin of Conservatism (MoC) categories, and the results point to real opportunities for banks. The research has shown promising new techniques, particularly for MoC categories A and B, and for the aggregation of MoC into a single total figure. In this article, we examine how each MoC category is defined, how banks approach them today, and what innovative techniques are available to improve the balance between conservatism and capital requirements.

Category A - Data & Methodology Deficiencies

MoC addresses situations where known issues in data or model design could bias risk estimates. For example, a bank might discover gaps or errors in historical data or realize that a simplifying assumption in the model (like a coarse segmentation or incomplete risk drivers) introduces a bias. According to EBA's definitions, MoC A should cover "data and methodological deficiencies" identified in the modeling or calibration process (EBA, 2017).

If you find a material deficiency in your data or method, you should first try to address it directly via Appropriate Adjustments (AA) to your data or model (e.g. imputing missing values). However, if some bias or shortcoming remains unaddressed, you then quantify a conservative add-on to compensate for the residual uncertainty. The aim is to offset any potential underestimation of risk caused by known issues.

Current Market Practices

Banks often rely on expert judgment and simple scenario analyses to set MoC A adjustments. For instance:

1- If certain older data is deemed less reliable, common practices are to:

• Benchmark model results using a cleaner subset of data (or external data) to gauge the impact. The difference might be added as MoC A.

• Model the upper confidence bound of the parameter estimate and consequently set the add-on equal to the distance between the point estimate and the upper bound.

2- In some instances, analysts may even apply a qualitative add-on to risk parameters based on their expert judgment, sometimes amounting to a flat increase or a scaling factor.

These simpler approaches are easy to communicate but can be somewhat simple and subjective, making it hard to defend the resulting values to model validation and the regulator.

Emerging Techniques

Recent research has helped to quantify MoC A more objectively. One such is the probabilistic quantification approach proposed by (Biche, 2022). They establish a closed-form expression that does not rely on an arbitrarily predefined significance level, to model uncertainty arising from missing values in the calibration sample. Instead, the Biche-approach defines a quantification of MoC A which is derived from a purely probabilistic perspective, increasing consistency across institutions.

Another direction we have explored is a framework based on influence functions, a statistical technique introduced in (Hampel, 1974), that examines how small changes in individual data points or model inputs affect the final estimate. By measuring influence on model output, one can identify where data issues or outliers might be biasing the estimates and then derive an MoC A to counteract that bias. This method allows the user to identify observations that drive MoC adjustments without repeated model re-estimation.

Conclusion

While market practices still lean on simple and judgment-driven approaches, these are increasingly challenged from a regulatory perspective. Emerging quantitative techniques offer a more structured and transparent way to capture identified biases, helping institutions move toward more unified MoC A calibration.

Category B – Changes in Environment & Policies

Category B MoC addresses structural breaks or shifts in legal- and/or macro-economic policy. These breaks between the past and the present (or expected future) would imply that your historical calibration data is (partially) out-of-date. In practice, such shifts can stem from any relevant changes to underwriting standards, risk appetite, collection and recovery policies, or any other source of additional uncertainty about the future. Even if your data and model are flawless, the future may not resemble the past.

Suppose your bank significantly tightened its lending standards, or a new economic regime (like a financial crisis or pandemic) emerged. Your PD, LGD or EAD models, built on the last decade of data, might now be too optimistic or not fully relevant for today's portfolio and/or economic state. Regulation requires you to ensure that if tomorrow's world differs from yesterday's, your risk estimates have a buffer to cover that gap. This is achieved through MoC B.

One key challenge is that Category B issues often overlap with Category C or even Category A. For instance, a structural break (Category B) usually reduces the usable data length, which in turn increases estimation error (Category C). Some banks attempt to model these overlaps explicitly. Nevertheless, modelling the co-movement of the different MoC categories is still a novel area for most banks. Section 5 covers this topic in more detail.

Current Market Practices

Banks usually handle MoC B by blending statistical tests and scenario analysis:

1- Some use structural break tests on historical default or loss data to detect if and when behavior changed (for example, a Chow test might reveal a significant shift in default rates after a certain year or policy change). If a break is detected, this flags a deficiency.

2- Others apply simple ratio comparisons (scaling factors): e.g. comparing post-change default rates to pre-change. If the model calibration sample underrepresents a recent adverse period, they might scale up PDs or LGDs by the observed difference (or by an expert-determined factor) to be conservative.

Emerging Techniques

Advanced frameworks propose quantifying MoC B via scenario modeling: first create scenarios that reflect new processes or extreme macroeconomic conditions, then measure how much PD, LGD or EAD might change under those scenarios. The weighted differences from the base forecast can serve as the MoC B. This is more dynamic and connects modeling directly to the business.

Conclusion

While current practices often rely on relatively simple diagnostics and scaling approaches, these can struggle to fully capture shifts. More advanced, scenario-based frameworks offer a forward-looking and economically grounded alternative, enabling banks to align their MoC B quantification more closely with evolving portfolio characteristics and external conditions.

Category C – General Estimation Error

Finally, Category C MoC addresses the inherent statistical uncertainty present in risk parameter estimates. Regulation states that MoC C "should reflect the dispersion of the distribution of the statistical estimator" (EBA, 2017). As econometric models are estimated on finite data samples, there will always be a margin of error surrounding the estimated PD, LGD, or EAD. This error, also referred to as noise, is most prevalent in low-default portfolios.

Current Market Practices

There is a spectrum of approaches that each have pros and cons; the most common techniques are summarized in Table 1.

Many banks choose a confidence level (α) and compute MoC C so that the final estimate corresponds to a high-end confidence bound. For example, for PD one could use a binomial distribution of default rates to determine that the upper bound of the 95% confidence interval is X% higher than the observed long-run average default rate (LRADR). MoC C is then set equal to the difference between the two.

A simplified approach is the k·σ ("k-sigma") method: choose a multiplier k for the standard error (σ) of the risk parameter, and set MoC C = k × σ. For example, if a PD's standard error is 0.2 percentage points, applying k=1.96 (approx. 95% confidence) leads to a MoC of 0.39 percentage points.

Emerging Techniques

Category C is the most well-developed MoC category as it relates to the general estimation uncertainty, present in every econometric model. Hence, many quantification techniques have already been developed and are already in use.

Conclusion

Unlike Categories A and B, MoC C is methodologically well established, leading to widespread use of standardized techniques such as confidence intervals and k‑sigma approaches. The key challenge therefore lies less in innovation and more in selecting appropriate confidence levels and ensuring a consistent level of conservatism across portfolios.

Technique	Pros	Cons
Parametric distribution	Assume defaults or losses follow a known statistical model. These often give higher, more stable MoCs and never result in zero or negative MoC.	Relies on model assumptions.
Empirical variance estimation	This is model-free and often simply uses the sample standard deviation of historical default rates.	Tends to give volatile results if data is scarce or correlated.
Resampling/Bootstrapping	Flexible (doesn't impose a parametric form) and ensures MoC is always positive.	Can be computationally heavy and may not improve accuracy if data is very limited.

Aggregation – Beyond Simple Summation

After estimating MoC A, B, and C, how should banks combine them?

The simplest (and, in regulatory landscape, the default) approach is to sum the margins over the categories: MoC_Total = MoC_A + MoC_B + MoC_C. This straightforward sum ensures a conservative outcome and is in line with: "Institutions should quantify the final MoC as the sum of MoC A, MoC B and MoC C" (EBA, 2017). But this conservative simplicity comes at a cost.

Key issues with simple summation

1- Double Counting: The categories are not guaranteed to be truly independent. For example, a structural break (Category B) often reduces the data sample, increasing the estimation error (Category C). Simply adding margins implicitly assumes worst-case alignment of all uncertainties, that every deficiency, future change, and statistical error all push the risk estimate in the same adverse direction simultaneously. However, in practice some uncertainties might overlap or partially offset.

2- Confidence Level Mismatch: If each margin is set to cover a high percentile of uncertainty, then summing them could yield a combined cushion far beyond that percentile. If MoC A and B each target an approx. 90% confidence level, adding them to MoC C could result in overall conservatism beyond 99.9%, which is more than intended. This excess conservatism means higher capital requirements.

Emerging Techniques

At Zanders, we are conducting ongoing research to explore methods to aggregate MoC more analytically. One idea is to treat overall MoC as a target confidence interval problem: define the desired overall safety level (e.g. 90% or 95% confidence that the true risk parameter is below [PD+MoC]) and then calculate MoC collectively rather than category by category. This might involve joint simulations or analytical combination of uncertainties:

1- Integrated Models: Simulate scenarios that incorporate data issues, policy changes, and statistical noise together. Compute the required MoC as the difference between the base estimate and a desired/high percentile of the simulated risk parameter distribution. This would inherently account for any interdependence between MoC A, B, and C factors (avoiding simple over-addition).

2- Variance-Covariance Aggregation: If one can estimate the variance contribution of each category and their correlations, the total uncertainty could be aggregated using statistical rules (like summing variances for independent factors, or more general formulas when not independent). Such approaches yield a smaller composite MoC than raw summation, aligning with a specified confidence level.

3- Calibration of k-Factor: Another approach (as suggested by some industry studies) is to adjust the k multiplier in the k·σ method to implicitly capture some effects of A and B. For instance, a higher k can be chosen in segments known to have Category A/B issues, rather than adding separate chunks.

Conclusion

These advanced methods can be complex to implement and justify. Regulators have so far preferred simplicity to ensure consistency across banks (hence the official recommendation of summation). The onus is on banks to demonstrate that any alternative approach is sound, transparent, and doesn't understate capital.

Toward a more calibrated Margin of Conservatism

The Margin of Conservatism is a crucial but challenging element of IRB credit risk modelling. It ensures a safety net for risk estimation by addressing known data issues (A), evolving business environments (B), and statistical uncertainties (C) that could otherwise lead to underestimation of risk. As we have seen, implementing MoC requires balancing rigor vs. simplicity: applying enough conservatism to satisfy regulatory standards and absorb model risk, but not so much that it overrules true risk differentiation or double counts uncertainties.

The state of practice is still evolving. Banks have adopted a variety of methods, from straightforward confidence interval calculations to the innovative use of influence functions, structural break analyses, and bootstrap simulations, each with its pros and cons. Meanwhile, supervisors, like the ECB, have pointed out inconsistencies between the way banks have implemented the MoC, encouraging them to refine their frameworks.

A key open question remains: what's the best way to aggregate different MoC components without introducing unintended excess conservatism? Forward-looking institutions are exploring ways to jointly model categories A, B, and C into a unified measure that respects their overlaps.

In the end, the MoC discipline forces modelers to explicitly confront model risk within Pillar 1 capital. By continuously improving how we quantify and aggregate these conservatism margins, the banking industry can meet regulatory expectations while preserving the integrity of their models. If you're an IRB modeler or validator, now is a great time to re-examine your MoC approach – ensuring it captures all relevant uncertainties without sinking your model's risk differentiation. Embracing rigorous yet practical techniques for MoC can turn a source of capital inflation into a well-calibrated guardrail for robust, credible credit risk models.

Are you interested in how you could leverage these methodologies to enhance your Credit Risk modeling approach? Contact Kyle Gartner, John de Kroon or Kasper Wijshoff for more information.

References

BCBS. (2004). International Convergence of Capital Measurement and Capital Standards. Bank for International Settlements.

Biche, E. (2022). A Probabilistic Approach to Quantify Margin of Conservatism for Data Deficiencies in a Calibration Sample for Credit Scoring Models.

EBA. (2017). Guidelines on PD estimation, LGD estimation and the treatment of defaulted exposures. European Banking Authority.

Hampel, F. (1974). The influence curve and its role in robust estimation.