How Credit-Portfolio Granularity-Adjustment Algorithms Correct Infinite-Diversification Models: ASRF Limits, HHI, Idiosyncratic Risk, VaR/ES and Concentration Failure

Reader question: Basel-style one-factor credit models often assume a portfolio is infinitely fine-grained, so borrower-specific risk diversifies away. What mathematical correction is needed when a real portfolio contains large individual names instead?

A granularity adjustment treats finite-name concentration as a perturbation around an infinitely granular benchmark. The benchmark captures systematic credit risk. The adjustment estimates the extra tail risk created because real exposures have nonzero weights and idiosyncratic defaults do not disappear perfectly.

This article owns the finite-portfolio correction problem: obligor exposure weights + PD/LGD/default model + systematic factor structure → infinitely granular loss benchmark, conditional idiosyncratic variance, concentration-sensitive granularity correction, VaR/Expected-Shortfall diagnostics and limits of the approximation.

It does not own the entire Basel IRB formula, large-exposure regulation, sector concentration models or general portfolio optimisation. Those are separate roles in the Bukit Timah Tutor finance-and-banking-algorithms estate.

This is public mathematical and computational education. It is not financial advice or a statement about the capital adequacy of any particular bank.

1. The asymptotic idea behind one-factor credit models

Consider a portfolio loss:

L = Σ w_i LGD_iD_i,

where:

• w_i is exposure weight;
• LGD_i is loss given default;
• D_i is a default indicator.

In an Asymptotic Single Risk Factor, or ASRF, model, defaults are conditionally independent given one systematic factor X.

For sufficiently fine portfolios, borrower-specific randomness diversifies away. Conditional portfolio loss approaches its conditional expectation:

L ≈ μ(X) = E[L|X].

The tail of portfolio loss can then be derived primarily from the systematic factor rather than simulating every individual default.

2. Infinite granularity is a mathematical limit, not a description of every bank

The asymptotic result requires that no individual exposure remains material as the portfolio grows.

A common condition is conceptually:

max w_i → 0

as the number of obligors becomes large.

If one borrower is 20% of the portfolio, idiosyncratic risk from that borrower does not vanish merely because hundreds of small borrowers are added around it.

The gap between the finite portfolio and the infinitely granular limit is the problem granularity adjustment tries to approximate.

3. Conditional expected loss

Let:

q_i(X) = P(D_i=1 | X)

be obligor i‘s conditional probability of default given the systematic state.

With deterministic LGD for simplicity:

μ(X) = E[L|X] = Σ w_iLGD_iq_i(X).

In the infinitely granular limit, the loss distribution is approximately the distribution of μ(X).

This is powerful because one factor can replace thousands of borrower-level Bernoulli outcomes.

4. Finite portfolios retain conditional variance

Conditional on X, defaults are still random in a finite portfolio.

Under conditional independence and deterministic LGD:

Var(L|X) = Σ w_i²LGD_i²q_i(X)[1−q_i(X)].

The square on w_i is the important clue.

Large exposure weights contribute disproportionately to residual idiosyncratic variance.

5. HHI is the simplest concentration lens

Define the exposure Herfindahl–Hirschman Index:

HHI = Σ w_i².

For N equal-sized obligors:

w_i = 1/N

and:

HHI = N(1/N²) = 1/N.

As N → ∞, HHI → 0.

This mirrors the disappearance of conditional idiosyncratic variance in a homogeneous portfolio.

HHI is not itself a complete credit-risk model, but it is a useful first indicator of how far a portfolio is from the infinitely granular ideal.

6. Effective number of equal exposures

A useful interpretation is:

N_eff = 1/HHI.

If HHI = 0.01, the exposure concentration is comparable—purely in this weight-squared sense—to 100 equal exposures.

If HHI = 0.10, the effective number is only 10.

This does not account for different PDs, LGDs, sectors or systematic correlations, but it gives an intuitive concentration scale.

7. What the granularity adjustment is trying to approximate

Let:

VaR_α^ASRF

be the loss quantile from the infinitely granular one-factor model.

A finite-portfolio approximation has the form:

VaR_α^finite ≈ VaR_α^ASRF + GA_α.

The granularity adjustment GA_α is a higher-order correction derived from the conditional variance of finite-name loss and the local shape of the systematic-factor mapping around the target tail state.

It is not simply “VaR times HHI.” The true analytical expressions depend on the portfolio model, conditional loss distribution, factor density and derivatives of conditional moments.

8. Why the adjustment is second-order

In the asymptotic portfolio, the first-order object is the systematic conditional mean μ(X).

Finite-name randomness enters through terms involving squared exposure weights and conditional variance. In a well-granular portfolio those terms are small.

This is why granularity adjustment is naturally a perturbation method:

large systematic benchmark + small finite-name correction.

If the finite-name correction is not small, the perturbation logic itself becomes questionable.

9. A homogeneous intuition

Suppose all obligors have the same weight, PD, LGD and factor loading.

Conditional variance then scales approximately as:

Var(L|X) ∝ 1/N.

So the finite-name correction shrinks as the portfolio becomes more granular.

For heterogeneous portfolios, a few large weights can dominate the sum of squared exposures even when the nominal obligor count is large.

Ten thousand borrowers do not imply high granularity if a handful of names contain most of the exposure.

10. Why default probability alone does not determine granularity risk

Two obligors can have the same exposure weight but different PDs. The conditional Bernoulli variance term:

q(1−q)

changes with conditional default probability.

Likewise, LGD enters squared in conditional variance when deterministic LGD is used.

Therefore an exposure-only concentration measure is useful but incomplete. The true correction combines:

• exposure size;
• credit quality;
• recovery severity;
• systematic sensitivity;
• tail location.

11. The ASRF factor orientation must be explicit

Some implementations define high X as a good economy; others define high X as stress.

The target systematic state for a high loss quantile therefore depends on sign convention. A formula copied from a paper can be wrong if the factor orientation is reversed without transforming derivatives consistently.

A robust implementation stores the factor convention and tests monotonicity:

does μ(X) increase or decrease as the economy worsens?

12. VaR and Expected Shortfall corrections are not identical

VaR asks for one quantile of the finite-portfolio loss distribution.

Expected Shortfall averages losses beyond the VaR threshold.

Granularity techniques can be derived for both, but the corrections depend on different functionals of the tail.

It is unsafe to compute a VaR correction and reuse it mechanically as an ES correction.

13. Mark-to-market credit models extend the idea

The original intuition is often taught with default-only losses. Federal Reserve research by Gordy and Marrone shows that granularity-adjustment methods can also be developed for single-factor mark-to-market credit models, including ratings-based and continuous obligor-state frameworks.

That matters because credit portfolios can lose value before default through migration and spread changes.

The central mathematical idea survives:

systematic factor drives the asymptotic portfolio; finite obligors add residual idiosyncratic dispersion.

14. Inputs and outputs

Inputs can include:

• obligor-level EAD or exposure weights;
• PDs;
• LGDs or recovery distributions;
• maturity where relevant;
• factor loadings/asset correlations;
• systematic factor distribution;
• conditional default model;
• target confidence level;
• VaR or Expected Shortfall objective;
• portfolio aggregation by underlying obligor;
• sector or geographic classifications for diagnostics.

Outputs can include:

• ASRF benchmark loss;
• HHI and effective name count;
• conditional mean and variance at the tail state;
• granularity adjustment;
• finite-portfolio VaR/ES approximation;
• top-name contributions;
• upper/lower bounds where incomplete data methods are used;
• Monte Carlo benchmark error;
• model-validity warnings.

15. Obligor aggregation is essential

A borrower can appear through multiple facilities, subsidiaries or booking systems. If exposures to the same underlying obligor are treated as independent small names, HHI and conditional idiosyncratic variance are understated.

Before any granularity calculation, exposures must be aggregated to the economic obligor level required by the model.

Falsifier: merge all facilities belonging to the same obligor and recompute HHI. If concentration rises materially, the unaggregated result was misleading.

16. Exposure weights must reconcile

For a normalised portfolio:

Σ w_i = 1.

Then HHI must satisfy:

1/N ≤ HHI ≤ 1

for N positive exposures, with equality at the lower bound only for equal weights.

A result outside these bounds indicates a data or normalisation error.

17. Evidence polarity

Evidence for confidence includes reconciled obligor-level exposure data, a small maximum exposure weight, stable HHI, granularity adjustments that shrink as the portfolio is artificially subdivided into genuinely independent equal names, close agreement with finite-portfolio Monte Carlo, and robustness across nearby confidence levels and model parameters.

Evidence against confidence includes one or two dominant obligors, large correction relative to the asymptotic capital itself, failure to aggregate related facilities, strong sector dependence ignored by the single-factor model, random recoveries or fat-tail dependence outside the derivation, granularity formulas that worsen Monte Carlo approximation, or results that reverse under small changes in exposure data.

18. Counterexample: many loans but one corporate group

A bank can hold 500 facilities but only 50 economic obligors because multiple facilities belong to the same corporate groups.

Treating every facility as an independent name creates an artificially low HHI.

Falsifier: aggregate by common obligor or connected group and compare the concentration metrics and GA.

19. Counterexample: sector concentration is not name granularity

A portfolio may contain thousands of individually tiny mortgage loans in one geographic region. Name-level HHI can be very low while all borrowers share one regional housing shock.

The ASRF model may represent some common systematic risk, but a single factor can miss additional sector or geographic factors.

Falsifier: introduce sector factors or stress scenarios. If tail loss changes materially while name HHI remains small, the problem is systematic concentration rather than residual name granularity.

20. Counterexample: the correction is too large to be a correction

Granularity adjustment is an asymptotic expansion around a well-diversified portfolio.

If a single exposure is 30% of the portfolio, a small-perturbation approximation can be outside its natural operating range.

Falsifier: compare with exact or high-precision Monte Carlo simulation. If the GA error is large or non-monotone, simulate the finite portfolio directly rather than relying on the asymptotic correction.

21. Counterexample: discrete loss masses break smooth approximations

Credit loss distributions can contain probability masses because defaults are discrete. Some granularity derivations require smoothness conditions that are weaker or stronger depending on the model.

Research on the limits of granularity adjustments shows that the correction does not always improve the infinitely granular approximation, especially when regularity conditions fail.

Falsifier: compare the approximated CDF/quantile against simulation around the target confidence level. If the discrete finite distribution jumps over the target quantile, smooth asymptotic corrections can misbehave.

22. Counterexample: random LGD adds another idiosyncratic layer

If LGD is stochastic, conditional loss variance includes recovery uncertainty and potentially default–recovery dependence.

A deterministic-LGD formula can understate or distort finite-name risk.

Falsifier: simulate stochastic recoveries under the intended dependence structure and compare tail loss with the deterministic-LGD GA.

23. Counterexample: multi-factor systematic risk

The single-factor assumption is convenient because the tail state is one-dimensional. Real portfolios can depend on industry, country, commodity and macro factors simultaneously.

More general granularity theory exists, but formulas become more delicate and their validity conditions matter.

Falsifier: fit or stress a multi-factor model. If large systematic concentration remains after the name-level adjustment, the one-factor GA is addressing the wrong missing risk.

24. Monte Carlo is the most direct finite-portfolio benchmark

A validation simulation can:

1. draw the systematic factor;
2. compute conditional PDs;
3. draw borrower defaults;
4. apply LGDs;
5. sum portfolio loss;
6. repeat enough times to estimate the target tail;
7. compare finite-portfolio VaR/ES with the ASRF benchmark plus GA.

Because high-quantile credit losses are rare, simulation error must be reported. Importance sampling, saddlepoint methods or other variance-reduction techniques can help.

25. Diagnostics

• Weight reconciliation: Σw_i = 1.
• HHI bounds: verify 1/N ≤ HHI ≤ 1.
• Maximum-name test: track max w_i.
• Obligor aggregation: merge facilities to the economic borrower.
• Replication test: split a truly independent equal exposure and verify the expected granularity scaling.
• Monte Carlo benchmark: compare finite VaR/ES directly.
• Confidence-level sweep: inspect 99%, 99.9% and neighbouring tails as appropriate.
• Parameter stress: vary PD, LGD and asset correlation.
• Sector-factor test: separate name concentration from systematic concentration.
• Approximation-ratio test: flag GA that is not small relative to the ASRF benchmark.

26. Regulatory context as of 30 August 2026

The granularity-adjustment literature grew from the Basel II/III IRB architecture, whose credit VaR model uses an asymptotic portfolio assumption.

The current Basel Framework remains the authoritative source for live standards. Its Supervisory Review Process states that Pillar 2 is intended to capture risks not fully captured by Pillar 1.

The Basel credit-risk supervisory-review text explicitly notes that credit concentration risk is not addressed in the Pillar 1 credit-risk capital charge and should be considered in banks’ capital-adequacy assessment under Pillar 2. This means a granularity adjustment is best understood here as mathematical concentration-risk machinery—not as a universal claim that one particular GA formula is a mandatory Pillar 1 regulatory add-on everywhere.

27. Alternatives and complements

Direct finite-portfolio Monte Carlo avoids the asymptotic expansion but is computationally heavier in extreme tails.

Saddlepoint approximations can estimate finite-portfolio tail probabilities and quantiles using cumulant-generating functions.

CreditRisk+ constructs portfolio loss distributions through a different actuarial framework.

Multi-factor credit models address systematic sector concentration.

Large-exposure limits are ex-ante regulatory constraints rather than statistical tail approximations.

HHI and top-name limits are simpler diagnostics that can be more transparent when the portfolio is too concentrated for asymptotic theory.

28. Connections to the surrounding Bukit Timah Tutor knowledge estate

The asymptotic regulatory baseline is Basel IRB credit-risk algorithms.

The broader concentration problem is mapped in credit-portfolio concentration algorithms. This page goes deeper into the finite-name asymptotic correction itself.

Dependence modelling connects to Gaussian-copula credit-portfolio algorithms.

An alternative portfolio-loss engine is CreditRisk+ algorithms.

Tail approximation connects to saddlepoint credit-portfolio algorithms.

The full lane is indexed at Finance & Banking Algorithms | Applied Mathematics in Real Financial Systems.

29. What would falsify confidence?

Confidence should be withdrawn if exposures are not aggregated to the correct obligor; if HHI or maximum weights show the portfolio is far from granular; if the GA is large relative to the asymptotic benchmark; if Monte Carlo shows the correction worsens the approximation; if sector/multi-factor concentration dominates name concentration; if stochastic recovery or discrete-loss effects violate the derivation; or if regulatory interpretation is presented without checking the current Basel Framework and local rules.

30. Verification and update triggers

Preserve the obligor mapping, exposure weights, PDs, LGDs, factor loadings, systematic-factor convention, confidence level, ASRF benchmark, HHI, conditional variance, GA formula version and Monte Carlo benchmark with each calculation.

Revalidate after material portfolio acquisitions, large-name changes, obligor mergers, PD/LGD methodology changes, asset-correlation changes, recovery-model changes or factor-model redevelopment. Trigger review when the maximum exposure or HHI moves beyond the range covered by the original approximation validation, or when sector/geographic concentration becomes more important than name granularity.

31. Primary and high-quality references

• Michael B. Gordy and Eva Lütkebohmert, Granularity adjustment for Basel II, Deutsche Bundesbank Discussion Paper, 2007.
• Michael B. Gordy and Eva Lütkebohmert, Granularity Adjustment for Regulatory Capital Assessment, International Journal of Central Banking, 2013.
• Michael B. Gordy and James Marrone, Granularity Adjustment for Mark-to-Market Credit Risk Models, Federal Reserve FEDS, 2010.
• Basel Committee on Banking Supervision, Studies on credit risk concentration: an overview of the issues and a synopsis of the results from the Research Task Force project, 2006.
• Jean-David Fermanian, The limits of granularity adjustments, Journal of Banking & Finance, 2014.
• Basel Committee on Banking Supervision, Supervisory review process, current Basel Framework; verified 30 August 2026.

Educational boundary: Granularity adjustment is an asymptotic approximation for residual finite-name risk. It is most informative when the portfolio is already reasonably granular and its assumptions are independently validated.