Advanced Credit VaR

1 The Question Behind Credit VaR

A lender can price every loan for its average expected loss and still be unprepared for a recession. The missing question is not simply “How likely is this borrower to default?” It is “How large could the portfolio loss become when many things go wrong together?”

Credit Value-at-Risk translates a joint loss distribution into a decision threshold. Its usefulness depends on what happens before that threshold is calculated: how we define loss, estimate default probabilities, model dependence, allow recoveries to deteriorate and recognise concentrations.

Learn by doing: Try three simple interactive caselets: Credit VaR, WCDR and capital →

What you will learn

• Distinguish expected loss, loss volatility, VaR, expected shortfall and economic capital.
• Derive the one-factor credit model and interpret its assumptions.
• Explain why concentration, uncertain recoveries and migration risk change the answer.
• Design a simulation, allocate risk and challenge a model with scarce default data.

Prerequisites: basic probability, normal distributions and present value. Read §§2–6 first; use the later sections as an advanced reference. All numerical portfolios are constructed teaching examples, not estimates of an actual bank.

2 Define the Loss Before Measuring It

Fix a one-year horizon \(T\). For a default-only portfolio, a useful starting point is

\(D_i\) is one if borrower \(i\) defaults by the horizon and zero otherwise. \(EAD_i\) is the amount exposed at default; \(LGD_i\) is the fraction lost after recovery. Express both consistently, including the treatment of recovery timing and costs. If exposure and recovery are random, their joint behaviour with default matters.

For a mark-to-market model, define loss relative to a stated benchmark, for example \(L=V_0-V_T-C_T\), with intervening cash flows \(C_T\) expressed in compatible units. A forward-value benchmark would instead include carry. Neither convention is inherently wrong; mixing them is.

A percentile is only meaningful after “loss” has a precise definition.

3 One Portfolio, Five Different Risk Numbers

Our running portfolio has 1,000 equal loans of ₹10 lakh each: total exposure ₹100 crore. Every borrower has one-year PD \(p=2\%\), and LGD is 45%. One default therefore costs ₹4.5 lakh, or ₹0.045 crore. We first assume independent defaults, then relax that assumption.

Expected loss is \(EL=E[L]\). With fixed exposures and severities, \(EL=\sum_i EAD_iLGD_iPD_i\). This sum does not require independent defaults. The loss variance does depend on the joint probabilities:

Here \(p_{ij}=P(D_i=1,D_j=1)\). Under independence the covariance terms vanish. The running portfolio then has standard deviation \(0.045\sqrt{1000(0.02)(0.98)}=\mathrm{INR}\,0.1992\) crore.

The term unexpected loss is used both for loss standard deviation and for capital above EL. They are not interchangeable. Also, the simple EC expression is an economic modelling convention; it is not a complete regulatory capital calculation.

4 Discrete Tails: Where VaR Can Mislead

Credit losses often arrive in jumps. A loan may pay in full or lose a large amount; there need not be a smooth bell-shaped distribution. This changes both quantile calculation and diversification behaviour.

Expected shortfall avoids this particular defect. Its general definition works even with probability mass at the VaR threshold:

For the two-loan portfolio, the worst 1% includes all ₹200 and ₹100 outcomes plus a small amount of the zero-loss outcome. Thus \(ES_{99\%}=(200\times0.000025+100\times0.00995)/0.01=\mathrm{INR}\,100\), even though VaR is zero.

Zero VaR does not mean zero risk. It may mean the default lies just beyond the selected confidence level.

5 How a Common Factor Connects Defaults

Defaults can cluster without one borrower directly causing another to fail. A recession, funding shock or sector downturn can weaken them together. The one-factor Gaussian model represents this common exposure as

\(Z\) is a common standard-normal economic factor. The \(\varepsilon_i\) are independent standard-normal borrower shocks, independent of \(Z\). Lower \(X_i\) means poorer credit quality. Because \(X_i\) remains standard normal, each borrower’s unconditional PD stays equal to \(p\).

For distinct borrowers, \(\operatorname{Corr}(X_i,X_j)=\rho\). This is latent asset correlation, not default-indicator correlation. The latter is

Once the common factor is fixed at \(z\), borrowers are conditionally independent and their conditional default probability becomes

The PD at \(Z=0\) need not equal the unconditional 2% PD. Averaging a nonlinear conditional probability across all economic states is different from evaluating it at the average factor.

6 Deriving the Vasicek Loss Quantile

Imagine splitting the same total exposure across infinitely many tiny borrowers while retaining their common factor. Conditional on the economy, individual surprises average out. The portfolio default fraction converges to \(p(Z)\), and its loss rate converges to \(LGD\times p(Z)\).

Bad credit outcomes correspond to low \(Z\). The upper \(\alpha\)-quantile of loss therefore uses the lower \(1-\alpha\) quantile of the economic factor. Since \(\Phi^{-1}(1-\alpha)=-\Phi^{-1}(\alpha)\),

This is an asymptotic benchmark for the 1,000-loan portfolio, not its exact finite-portfolio quantile. The derivation removes idiosyncratic risk; a finite model must retain it.

Integrating the same asymptotic losses over the worst 0.1% of economic states gives \(ES_{99.9\%}\approx\mathrm{INR}\,9.4232\) crore. ES exceeds VaR because it measures severity beyond the threshold, not just entry into the tail.

The relationship between conditional stressed losses, expected loss and regulatory capital is explained in the Basel Committee’s IRB explanatory note. The calculation here deliberately excludes regulatory maturity adjustments, floors and jurisdiction-specific implementation.

7 Explore PD, LGD and Dependence

The model is most useful when we can explain what moves its answer. Change one assumption at a time before combining shocks. All values in this lab use the asymptotic formula; changing the inputs does not change the number of borrowers because that formula has no finite borrower-count parameter.

With fixed PD and dependence, doubling EAD or LGD doubles loss VaR. Changing PD or correlation works through a nonlinear probability transformation. At \(\rho=0\), the asymptotic model’s loss becomes deterministic and equals EL; this does not mean a finite independent portfolio has zero loss volatility.

8 Concentration: When Diversification Stops Working

A portfolio of 1,000 facilities is not necessarily a portfolio of 1,000 independent borrowers. Ten facilities to the same group can default together. Start with obligor aggregation, then assess sector, region, collateral and common funding exposures.

For equal-PD independent borrowers and fixed LGD, exposure weights \(w_i=EAD_i/EAD_{\rm total}\) give

The exposure Herfindahl index \(H=\sum_iw_i^2\) measures name concentration, and \(1/H\) is an effective number of equal exposures. Neither measure captures sector correlation on its own.

In a multi-factor model, \(X_i=\boldsymbol\beta_i^\top\mathbf Z+\sqrt{1-\boldsymbol\beta_i^\top\Sigma_Z\boldsymbol\beta_i}\,\varepsilon_i\), with valid non-negative residual variance. Shared sector loadings produce additional clustering. A portfolio heavily exposed to one sector cannot diversify that sector factor by merely adding more names within it.

Granularity adjustments approximate the residual idiosyncratic risk missing from an asymptotic model. For heavily concentrated books, direct portfolio modelling is often more informative than relying on a small correction. Gordy and Marrone’s research develops this idea for mark-to-market credit models.

9 Copulas: Dependence Beyond One Correlation

Individual PDs specify how often each borrower defaults; they do not specify how often borrowers default together. A copula provides that joint structure. For uniforms \(U_i\) and default rule \(U_i<p_i\), the two-name joint-default probability is \(C(p_i,p_j)\).

The Gaussian copula is a convenient benchmark. For correlations strictly between −1 and 1, its asymptotic tail-dependence coefficients are zero. This does not imply independence at realistic default thresholds: finite-threshold joint default probabilities can still substantially exceed \(p_ip_j\).

A Student-t construction introduces a shared random scale. Generate a correlated normal vector \(\mathbf Y\), an independent \(S\sim\chi^2_\nu\), and set \(T_i=Y_i/\sqrt{S/\nu}\), \(U_i=t_\nu(T_i)\). Default when \(U_i<p_i\). Using the t CDF and its corresponding thresholds preserves marginal PDs.

With discrete default indicators, the underlying copula is not uniquely identified from the binary outcomes alone. Defaults are sparse, and fitting a complex dependence model can create false precision. Compare plausible alternatives and report sensitivity instead of treating one calibrated parameter as a fact.

For a fuller treatment of the probability transforms and tail coefficients, continue with the Academy’s Copulas & Tail Dependence guide.

10 Random LGD, Drawdowns and Wrong-Way Risk

In a downturn, more borrowers can default at the same time that collateral prices fall and undrawn commitments are used. Modelling these three effects separately can miss their interaction.

The general single-name mean loss is \(E[EAD\,LGD\,D]\), not automatically \(E[EAD]E[LGD]PD\). For fixed EAD, it becomes \(EAD\times PD\times E[LGD\mid D=1]\). A recovery average taken across non-default states is the wrong conditioning set.

For a revolving facility, an illustrative exposure model is \(EAD=\text{drawn}+CCF\times\text{undrawn}\). With ₹6 crore drawn, ₹4 crore undrawn and CCF 50%, EAD is ₹8 crore. If stressed utilisation implies CCF 90%, exposure rises to ₹9.6 crore before any change in LGD.

Wrong-way risk describes adverse dependence between exposure and counterparty credit quality. A derivative may become more valuable to the bank precisely when the counterparty is less able to pay. Simulating market exposure independently of credit quality can therefore understate counterparty losses.

A bad year may contain more defaults, larger exposures and smaller recoveries—all at once.

11 Migration Risk and Horizon Revaluation

A borrower need not default to produce a loss in economic value. CreditMetrics-style models represent future ratings, revalue surviving exposures and assign recovery values on default. The original CreditMetrics technical document is a foundational reference.

Consider a deliberately small transition matrix with states Good, Weak and Default:

Under a time-homogeneous Markov assumption, the two-year matrix is \(P^2\). The Good-to-Default probability is \(0.90(0.01)+0.09(0.05)+0.01(1)=2.35\%\). This counts every intermediate state.

The Good-to-Good entry is \(0.90^2+0.09(0.05)=81.45\%\), not \(0.90^2=81\%\): a borrower can deteriorate and recover. A shorter-period matrix must also be valid; an arbitrary fractional matrix power can produce inadmissible entries. A continuous-time generator \(Q\), where valid, gives \(P(t)=e^{Qt}\).

A fixed rating curve does not capture all spread risk. Spreads can widen while ratings remain unchanged. An extended model needs systematic spread shocks, rating-specific movements and internally consistent default valuations. Avoid recording the same credit deterioration once as a migration loss and again as an unrelated default loss.

12 CreditRisk+ and Model Choice

An actuarial model starts from default frequency rather than changes in market value. For a large set of small default probabilities, a Poisson approximation makes the aggregate count tractable. If default intensity itself fluctuates with shared economic conditions, a gamma mixing distribution yields a negative binomial count.

For a gamma shape \(k\) and scale \(\theta\), \(E[\Lambda]=k\theta\) and \(Var(\Lambda)=k\theta^2\). Shared intensity factors connect borrowers and sectors. A compound loss calculation then converts counts into money, allowing different exposure bands.

CreditRisk+ is associated with analytical recursion and related numerical techniques; Monte Carlo is not a defining requirement. Likewise, a multi-name migration model may use simulation without making simulation a separate economic theory. The choice of loss definition and joint assumptions matters more than the model label. Kurth and Tasche develop analytical VaR and ES contributions within CreditRisk+.

13 A Reproducible Monte Carlo Blueprint

Simulation turns assumptions into a sample of portfolio losses. It does not validate those assumptions. For the finite, homogeneous one-factor portfolio, there is a particularly efficient construction:

1. Draw one common factor \(Z\) for the scenario.
2. Calculate the conditional PD \(p(Z)\).
3. Draw \(N\mid Z\sim Binomial(n,p(Z))\).
4. Set \(L=(EAD_{\rm total}/n)LGD\times N\).
5. Repeat independently across scenarios; sort the losses.

This retains finite-name randomness without drawing 1,000 separate borrower shocks in every scenario. Heterogeneous portfolios require borrower-level indicators, grouped conditional calculations or another valid aggregation scheme.

# Reproducible teaching example: Python + NumPy + SciPy
import numpy as np
from scipy.special import ndtr, ndtri
rng = np.random.default_rng(20261003)
n, scenarios = 1000, 1_000_000
p, rho, lgd, ead, alpha = .02, .15, .45, 100., .999
z = rng.standard_normal(scenarios)
p_cond = ndtr((ndtri(p) - np.sqrt(rho)*z) / np.sqrt(1-rho))
defaults = rng.binomial(n, p_cond)
loss = ead / n * lgd * defaults       # units: ₹ crore
ordered = np.sort(loss)
q = ordered[int(np.ceil(alpha*scenarios))-1]
# Use only the required mass at q; do not average all ties.
es = (np.mean(loss * (loss > q))
      + q * (np.mean(loss <= q) - alpha)) / (1-alpha)
print(loss.mean(), q, es, q-loss.mean())

A checked run of this seeded example produced mean loss ₹0.9003 crore, 99.9% VaR ₹8.0100 crore and ES ₹9.5154 crore. These are finite-portfolio Monte Carlo estimates, not exact population values; small differences across runs or library versions are normal. The asymptotic VaR benchmark is ₹7.9348 crore.

At one million scenarios, only about 1,000 lie in a 0.1% tail. At 10,000 scenarios, only about ten do. More simulated rows reduce sampling error; they do not fix biased PDs or an inappropriate dependence structure.

For rare tails, importance sampling draws more adverse factors and reweights each scenario by its likelihood ratio. Those weights must enter both the loss CDF and ES calculation. Failing to reweight turns a numerical technique into a different, stressed probability model.

Check several seeds or independent batches, benchmark simple cases against exact binomial calculations, and record quantile uncertainty. For discrete loss distributions, do not assume a smooth density or apply continuous quantile standard-error formulas without checking their conditions.

14 Excel Blueprint: Build the Model in Cells

You can reproduce the asymptotic example before building a large simulation. Enter probabilities as decimals, and use crore consistently for every monetary value.

For one finite-portfolio simulation row, set borrower count in B7 to 1000. Use D2 for the macro factor, E2 for conditional PD, F2 for default count and G2 for loss:

D2 = NORM.S.INV(MAX(1E-12,MIN(1-1E-12,RAND())))
E2 = NORM.S.DIST((NORM.S.INV($B$3)-SQRT($B$5)*D2)
     /SQRT(1-$B$5),TRUE)
F2 = BINOM.INV($B$7,E2,RAND())
G2 = $B$2/$B$7*$B$4*F2

Fill down for independent scenarios. Excel’s RAND is volatile: freeze a copy of the simulated values for audit and comparison. For \(M\) equally weighted losses, use SMALL(loss_range,ROUNDUP(alpha*M,0)) for the lower empirical quantile. Interpolating spreadsheet percentiles can generate a loss that no discrete scenario actually has.

If VaR is stored in J2, calculate discrete ES with the strict-greater-than loss sum plus the required threshold mass:

=(SUMIF(loss_range,">"&J2,loss_range)/COUNT(loss_range)
 +J2*(COUNTIF(loss_range,"<="&J2)/COUNT(loss_range)-alpha))
 /(1-alpha)

15 Risk Contributions and Lending Decisions

A portfolio total cannot tell a credit committee which loan consumes the most risk capacity. A loan with a small standalone PD can have a large contribution if it is concentrated in the portfolio’s vulnerable sector.

For a differentiable, positively homogeneous risk measure \(R(\mathbf w)\), Euler allocation defines

Under appropriate continuity and regularity conditions, VaR contributions are related to conditional exposure losses at the portfolio VaR boundary, and ES contributions to conditional losses in the portfolio tail. With discrete defaults, differentiability and conditioning at the boundary require care; ES allocations must consistently distribute any threshold probability mass.

Incremental risk is different: it is the change in portfolio risk after adding or removing an actual exposure. Revalue the entire portfolio under the changed composition. It is not generally equal to the loan’s standalone VaR or to a derivative-based contribution.

Use consistent horizons, tax treatment and capital definitions when comparing RAROC with a hurdle rate. Risk contributions support limits and pricing; they should not hide model uncertainty inside a single precise percentage.

16 Stress Testing and Reverse Stress Testing

VaR answers a probability question under a model. Stress testing asks what a specified adverse scenario would do. Reverse stress testing starts with a failure threshold—such as losses exceeding available capital—and searches for combinations of assumptions that breach it.

The changes are deliberately illustrative. A defensible scenario needs an economic narrative: weaker revenues affect PD, collateral declines affect recoveries, and a common sector shock affects clustering. Avoid increasing parameters arbitrarily and describing the result as a calibrated forecast.

For reverse stress testing, suppose available capital is ₹10 crore and a separately available loss allowance absorbs ₹0.9 crore. The loss-absorption threshold is ₹10.9 crore under that simplified balance-sheet convention. Search PD–LGD–dependence combinations whose loss quantile crosses it, then examine whether the resulting scenarios are economically plausible.

Stress results are conditional “what if” answers. Their severity does not by itself establish their probability.

17 Physical Risk, Pricing and Regulatory Capital

Credit work uses several probability concepts. Mixing them can create a mathematically polished but economically inconsistent model.

A credit spread is not simply an observed physical PD multiplied by LGD. Under restrictive assumptions, \(s\approx\lambda^{Q}LGD\) is a useful pricing intuition, where \(\lambda^{Q}\) is a risk-neutral hazard rate. Liquidity, premia and other valuation effects can also affect observed spreads.

A coherent migration simulation may use physical probabilities to generate future credit states and market-consistent pricing curves to value each state. This is a deliberate separation of scenario likelihood from valuation, not an accidental mixing of measures.

Similarly, point-in-time PD responds to current conditions, while through-the-cycle estimates smooth cyclical variation. Applying a full recession shock to a PD already conditioned on that recession can double-count stress. Document the starting calibration and exactly what the scenario changes.

The simplified Vasicek formula explains one mathematical building block of IRB capital. It does not replace current rules. For implementation, consult the Basel risk-weight functions and relevant local rules. Accounting expected-credit-loss provisions, economic capital and regulatory capital serve related but different purposes.

18 Validation When Defaults Are Rare

At a 99.9% one-year threshold, a correctly specified continuous model expects only 0.001 exceedances per year. Over 20 independent years, the expected number is 0.02; the probability of seeing no exceedances is approximately \(0.999^{20}=98.02\%\). A spotless history therefore provides little direct evidence that the far tail is correct.

Validation must work below the final portfolio number:

• PD calibration: compare forecasts and realised default rates by grade, vintage and cycle.
• Discrimination: check whether the model ranks relative credit quality; good ranking alone does not prove calibrated probabilities.
• Severity: compare predicted and realised recoveries, including timing, costs and incomplete workouts.
• Dependence: examine clustering, factor exposure and alternative plausible dependence models.
• Implementation: benchmark limiting cases, data transformations and aggregation.
• Use: test whether limits and decisions remain sensible under adverse but plausible assumptions.

Pluto and Tasche’s low-default portfolio research discusses conservative estimation when observations are scarce. External data, expert judgement and pooled experience can help, but require comparability checks.

Separate simulation error from parameter uncertainty and model risk. A million scenarios can make a wrong assumption look very stable. Report sensitivity ranges and explain what evidence would change the model.

19 Formula Sheet and Common Mistakes

1. Calling VaR the maximum possible loss. It is a quantile; VaR alone does not describe the severity beyond it.
2. Subtracting EL twice. First identify whether “Credit VaR” already means quantile minus mean.
3. Using asset correlation as default correlation. The mapping depends on PD and the joint distribution.
4. Treating many facilities as many independent obligors. Aggregate connected exposures.
5. Using an asymptotic formula for a concentrated book. Retain finite-name and sector risk.
6. Averaging every loss at or above VaR for discrete ES. Include exactly the required tail probability mass.
7. Taking powers of individual migration probabilities. Multiply the full matrix.
8. Scaling annual default VaR by square-root-of-time. Default jumps and migration dynamics do not generally justify it.
9. Equating pricing PD with real-world PD. State the measure and calibration purpose.
10. Claiming more simulations prove model accuracy. They address numerical precision, not economic validity.

20 Knowledge Check and Takeaways

Try each question before opening the answer.

§ Sources & Further Reading

Original explanations and constructed numerical examples. The following primary and research sources support the modelling framework and provide deeper treatment.

1. Basel Committee (2005), An Explanatory Note on the Basel II IRB Risk Weight Functions — conditional losses and the ASRF foundation.
2. Basel Framework, CRE31: IRB risk-weight functions — regulatory specification; consult the effective version and local implementation.
3. Gupton, Finger & Bhatia (1997), CreditMetrics Technical Document — migration-based credit portfolio valuation; archived by Yale.
4. Kurth & Tasche (2002/2003), Credit Risk Contributions to Value-at-Risk and Expected Shortfall — analytical contributions in CreditRisk+.
5. Gordy & Marrone (2010), Granularity Adjustment for Mark-to-Market Credit Risk Models — residual idiosyncratic risk and concentration.
6. Pluto & Tasche (2004/2005), Estimating Probabilities of Default for Low Default Portfolios — inference with few or no defaults.
7. The Mountain Path Academy, All About Credit VAR (teaching workbook, supplied for this article). Cases A–F identify the worksheet and cell ranges used. Numerical results were independently checked; explanatory qualifications distinguish raw statistics, finite portfolios and asymptotic assumptions.

↗ Continue Learning