Default Dependence and Credit Portfolio Tail Losses

Why Default Dependence Matters

A credit portfolio can have an accurately estimated expected loss and still be exposed to substantial uncertainty in its total loss. The reason is that obligors do not necessarily default independently. A recession, commodity-price collapse, or other common economic condition can worsen many obligors’ credit quality at once. Defaults then cluster, and the portfolio’s upper tail can be materially heavier than an independence model suggests.

This tutorial models that clustering with a shared systematic driver. It compares independent defaults with defaults that are independent only after conditioning on the driver, and shows how this distinction affects portfolio tail losses. The emphasis is on a one-period portfolio with fixed exposures and loss given default; rating migration and individual loss severity are held fixed rather than modeled in detail. The material is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: For obligor $i$, let $D_i=1$ if the obligor defaults during the horizon and $D_i=0$ otherwise. Its default probability is $p_i=P(D_i=1)$. If its exposure at default is $EAD_i$ and its loss-given-default fraction is $g_i$, its loss is $L_i=EAD_i g_i D_i$. Portfolio loss is $L=\sum_i L_i$. A systematic factor $Z$ is a shared random variable that affects multiple obligors’ conditional default probabilities.

In an independence model, the indicators $D_i$ are mutually independent, with their specified marginal probabilities. In a factor model, default indicators may instead be independent conditional on $Z$. This means that once the factor value is fixed, each obligor’s conditional default probability is specified and the default events are independent; before observing $Z$, their outcomes are generally dependent.

A simple factor can have two states, such as adverse and ordinary conditions. A more detailed model can use a continuous factor, for example a standard normal variable. In either case, the key modeling step is to make conditional default probabilities respond to the same driver. A model must state the factor distribution and the conditional probabilities; saying only that “defaults are correlated” does not fully specify a portfolio loss distribution.

Dependence From a Shared Factor

Write $q_i(Z)=P(D_i=1\mid Z)$. By the law of total probability, the marginal probability is $p_i=E[q_i(Z)]$. If two defaults are conditionally independent given $Z$, then $P(D_i=1,D_j=1\mid Z)=q_i(Z)q_j(Z)$. Taking expectations gives

$$ P(D_i=1,D_j=1)=E[q_i(Z)q_j(Z)] $$

Subtracting $p_i p_j=E[q_i(Z)]E[q_j(Z)]$ shows that the covariance of the default indicators is the covariance of their conditional probabilities:

\operatorname{Cov}(D_i,D_j) =E[q_i(Z)q_j(Z)]-E[q_i(Z)]E[q_j(Z)] =\operatorname{Cov}(q_i(Z),q_j(Z)).

Thus, if a factor tends to raise both conditional probabilities, the indicators tend to have positive covariance. For two obligors with the same marginal default probability $p$, the correlation is $\operatorname{Cov}(D_i,D_j)/[p(1-p)]$, because a Bernoulli indicator has variance $p(1-p)$. Conditional independence does not mean unconditional independence.

Worked Example: Default Correlation Under a Two-State Factor

Two obligors have equal exposures. During a one-year horizon, an adverse factor state occurs with probability $0.10$. Given that state, each obligor has default probability $0.40$; otherwise, each has default probability $0.05$. The defaults are conditionally independent given the state. Find the marginal probability, joint-default probability, and default-indicator correlation. Ignore interest and assume no recoveries are needed for this indicator calculation.

Method. Average the conditional probabilities to find the marginal probability, then use conditional independence to calculate the joint probability. Finally, calculate covariance and divide by the indicator variance.

For either obligor, the marginal default probability is

$$ p=0.10(0.40)+0.90(0.05)=0.040+0.045=0.085. $$

The joint-default probability is a weighted average of the conditional joint probabilities:

$$ P(D_1=1,D_2=1)=0.10(0.40^2)+0.90(0.05^2) =0.016+0.00225=0.01825. $$

The covariance is $0.01825-(0.085)^2=0.01825-0.007225=0.011025$. Each indicator has variance $0.085(1-0.085)=0.085(0.915)=0.077775$. Therefore,

\operatorname{Corr}(D_1,D_2)=\frac{0.011025}{0.077775}=0.141755\ldots\approx 0.1418.

The results are marginal probability $0.085$, joint probability $0.01825$, and correlation $0.1418$ to four decimal places. As a check, the joint probability exceeds $p^2=0.007225$, as expected under positive dependence. Common mistake avoided: conditional independence does not permit replacing the unconditional joint probability by $p^2$.

How Dependence Changes Portfolio Loss Tails

Expected loss depends on the marginal default probabilities, not on dependence. With fixed $EAD_i$ and $g_i$,

E[L]=\sum_i EAD_i g_i p_i.

The variance and tail of $L$, however, depend on joint default behavior. In particular,

\operatorname{Var}(L) =\sum_i \operatorname{Var}(L_i) +2\sum_{i<j}\operatorname{Cov}(L_i,L_j).

This identity follows by expanding the variance of a sum: each individual variance appears once, and each pairwise covariance appears twice. With fixed nonnegative loss amounts, positive default-indicator covariance produces positive loss covariance. The resulting distribution can place more probability on both unusually small and unusually large losses than an independence model with the same marginals.

For a loss variable $L$, define $\operatorname{VaR}_{\alpha}(L)$ as the smallest loss $x$ for which $P(L\le x)\ge\alpha$. For discrete losses, an upper-tail average must account for the probability mass at the quantile. We use the quantile-integral definition of TVaR:

\operatorname{TVaR}_{\alpha}(L) =\frac{1}{1-\alpha}\int_{\alpha}^{1}\operatorname{VaR}_{u}(L)\,du.

This definition averages the worst $1-\alpha$ fraction of outcomes, taking only the needed fraction of probability mass at a discrete boundary. It is not always the same as simply averaging losses strictly greater than $\operatorname{VaR}_{\alpha}$.

Worked Example: Independent and Correlated Losses for Four Obligors

Consider four equal obligors, each with a hypothetical one-year loss of $100 if it defaults and zero otherwise. There is no interest adjustment. Compare the 95% VaR and TVaR under (a) independent defaults with marginal probability $0.085$, and (b) the two-state factor in the previous example, where all four defaults are conditionally independent given the state. Let $K$ be the number of defaults, so $L=100K$.

Method. In case (a), use the binomial distribution $K\sim\operatorname{Binomial}(4,0.085)$. In case (b), mix the conditional binomial distributions: $K\mid\text{adverse}\sim\operatorname{Binomial}(4,0.40)$ and $K\mid\text{ordinary}\sim\operatorname{Binomial}(4,0.05)$, with state probabilities $0.10$ and $0.90$.

(a) Independent defaults. The probability of no defaults is $P(K=0)=0.915^4=0.7009456$. The probability of exactly one default is $P(K=1)=4(0.085)(0.915^3)=0.2604607$. Hence $P(K\le1)=0.9614063$, so the 95% VaR is $100: the cumulative probability below $100 is $0.7009456$, and at $100 it exceeds $0.95$.

For TVaR, the worst 5% includes all outcomes with at least two defaults and enough of the one-default probability mass to make 5%. The probability with at least two defaults is $1-0.9614063=0.0385937$; the remaining probability needed from the $K=1$ mass is $0.0500000-0.0385937=0.0114063$. The unconditional expected number of defaults contributed by $K\ge2$ is $E[K]-P(K=1)=4(0.085)-0.2604607=0.0795393$. Including the selected one-default mass gives a tail loss numerator of $100(0.0795393+0.0114063)=\$9.0946$. Thus $\operatorname{TVaR}_{0.95}=9.0946/0.05=\$181.89$, rounded to cents.

(b) Factor-dependent defaults. The mixed probabilities for zero and one default are

P(K=0)=0.10(0.6^4)+0.90(0.95^4)=0.7460156, $$ $$ P(K=1)=0.10\{4(0.4)(0.6^3)\}+0.90\{4(0.05)(0.95^3)\}=0.1888875.

Therefore $P(K\le1)=0.9349031$. The probability of exactly two defaults is $0.10(0.3456)+0.90(0.0135375)=0.0467438$, giving $P(K\le2)=0.9816469$. The 95% VaR is consequently $200.

For TVaR, the probability of three or four defaults is $0.0157875+0.0025656=0.0183531$. The remaining $0.05-0.0183531=0.0316469$ probability is taken from the $K=2$ mass. The unconditional default-count contribution from $K\ge3$ is $3(0.0157875)+4(0.0025656)=0.0576250$. Adding $2(0.0316469)=0.0632938$ for the selected two-default mass gives a tail loss numerator of $100(0.1209188)=\$12.0919$. Therefore $\operatorname{TVaR}_{0.95}=12.0919/0.05=\$241.84$, rounded to cents.

Both models have the same expected loss, $4(0.085)(\$100)=\$34$, but their tail measures differ: VaR rises from $100 to $200 and TVaR from $181.89 to $241.84. These calculations check against the larger probability of clustered defaults under the factor model. Common mistake avoided: a 95% VaR of $200 does not mean every outcome in the worst 5% is exactly $200; TVaR must include the more severe three- and four-default outcomes.

Gaussian One-Factor Approximation

For a large, well-diversified portfolio of similar obligors, the Gaussian one-factor model gives a useful approximation to the loss distribution. This is an asymptotic model: the portfolio is assumed to have many small, comparable exposures, so idiosyncratic variation averages out. It is not an exact finite-portfolio formula, and it can be inappropriate when a few exposures dominate.

Let $Z\sim N(0,1)$ be the systematic factor, and let each obligor’s latent asset variable be $A_i=\sqrt{\rho}Z+\sqrt{1-\rho}\epsilon_i$, where the $\epsilon_i$ are independent standard normal variables, also independent of $Z$. Default occurs when $A_i\le c$, with $c=\Phi^{-1}(p)$; this gives each obligor marginal default probability $p$. Here $\rho$ is the asset correlation, not the default-indicator correlation. For a large homogeneous portfolio, the conditional default fraction approaches

q(Z)=\Phi\left(\frac{\Phi^{-1}(p)-\sqrt{\rho}Z}{\sqrt{1-\rho}}\right).

If the loss-given-default fraction is $g$, the portfolio loss fraction approaches $gq(Z)$. A high loss corresponds to a low factor value. The resulting upper-tail quantile is

\operatorname{VaR}_{\alpha}(\text{loss fraction}) \approx g\Phi\left( \frac{\Phi^{-1}(p)+\sqrt{\rho}\Phi^{-1}(\alpha)} {\sqrt{1-\rho}} \right).

Set the default threshold.
Since $P(A_i\le c)=p$, the threshold is $c=\Phi^{-1}(p)$.

Choose the factor quantile for an upper loss quantile.
The loss increases as $Z$ decreases. The factor value at the upper $\alpha$ loss quantile is $Z=\Phi^{-1}(1-\alpha)=-\Phi^{-1}(\alpha)$.

Substitute and scale by severity.
Putting this factor value into $q(Z)$ changes $-\sqrt{\rho}Z$ to $+\sqrt{\rho}\Phi^{-1}(\alpha)$. Multiplying the default fraction by $g$ gives the loss-fraction approximation.

Worked Example: A Large Portfolio’s 99% Loss Quantile

A hypothetical insurer holds a very large, granular portfolio of similar corporate exposures. Each obligor has one-year default probability $p=0.02$, asset correlation $\rho=0.15$, and loss-given-default fraction $g=0.60$. Estimate the 99% portfolio loss fraction using the Gaussian one-factor asymptotic model. The horizon is one year, and the calculation is undiscounted.

Method. Apply the large-portfolio formula, first finding the standard normal quantiles and then transforming the conditional default fraction into a loss fraction.

The required values are $\Phi^{-1}(0.02)\approx-2.0537$, $\Phi^{-1}(0.99)\approx2.3263$, $\sqrt{0.15}\approx0.3873$, and $\sqrt{0.85}\approx0.9220$. The numerator inside the standardization is

-2.0537+(0.3873)(2.3263)\approx-2.0537+0.9010=-1.1527.

Dividing gives $-1.1527/0.9220\approx-1.250$. Therefore the conditional default fraction at the 99% loss quantile is approximately $\Phi(-1.250)=0.1056$. The loss fraction is $0.60(0.1056)=0.06336$, or approximately $6.34\%$ to two decimal places in percentage terms.

As a check, the expected loss fraction is only $gp=0.60(0.02)=0.012$, or $1.20\%$. The larger 99% quantile reflects the adverse systematic factor, not an increase in the marginal default probability. Common mistake avoided: using $\Phi^{-1}(1-\alpha)$ with a plus sign in the numerator would move the factor in the wrong direction and understate the upper loss quantile.

Interpretation and Exam Traps

Dependence modeling matters because portfolio capital is driven by the distribution of aggregate losses, not only by their mean. A factor model supplies a mechanism for joint deterioration and allows a direct comparison with an independence benchmark. Its conclusions depend on assumptions: the factor distribution, the conditional default probabilities, conditional independence, exposure sizes, and loss severities. A single pairwise correlation does not by itself determine all higher-order joint default probabilities or the tail of a large portfolio.

Confusing conditional and unconditional independence. In a factor model, defaults can be independent after conditioning on the systematic driver while remaining positively dependent overall.
Using $p_i p_j$ as a joint probability automatically. That product applies to independent defaults, not generally to defaults sharing a factor.
Assuming dependence changes expected loss. With fixed marginal default probabilities and severities, dependence changes variance and tail risk, but not $E[L]$.
Mixing up correlation parameters. In the Gaussian one-factor construction, $\rho$ is the correlation of latent asset variables. It is not generally equal to the correlation of default indicators.
Treating an asymptotic result as exact. The homogeneous large-portfolio quantile is an approximation and may not capture finite-name effects or concentrated exposures.

Key takeaway: A common systematic factor can leave marginal default probabilities and expected loss unchanged while increasing clustered defaults and portfolio tail losses. Calculate dependence from conditional probabilities, then evaluate the aggregate loss distribution using the model’s stated assumptions.

Check Your Understanding

For each problem, identify the relevant dependence assumption before calculating. Amounts refer to one-period losses unless stated otherwise.

A shared adverse state has probability $0.20$. Two conditionally independent obligors each default with probability $0.30$ in that state and $0.04$ otherwise. Find their marginal default probability and default-indicator correlation. (Answer: Marginal probability $=0.0920$; correlation $=0.1295$.)
Three independent obligors each have default probability $0.10$ and incur a loss of $50 on default. Find the 90% VaR and TVaR. (Answer: VaR $=\$50$; TVaR $=\$64.50$.)
A two-state factor has adverse-state probability $0.05$. Each of two obligors has conditional default probability $0.50$ in the adverse state and $0.02$ otherwise, with conditional independence. Find the marginal default probability and default-indicator correlation. (Answer: Marginal probability $=0.0440$; correlation $=0.2602$.)
A large homogeneous portfolio has $p=0.05$, asset correlation $\rho=0.20$, and loss-given-default fraction $g=0.50$. Use the Gaussian one-factor approximation to estimate its 95% loss fraction. (Answer: Approximately $0.0774$, or $7.74\%$.)

What You'll Learn