On this page
Default correlation measures how strongly borrowers’ default events occur together over a specified horizon. For a credit portfolio, it helps explain why losses may cluster even when individual default probabilities look modest. The measure is useful, but a single estimate can mislead when joint defaults are scarce, borrower exposures differ, or the model assumes a simple dependence pattern. A candidate should calculate the pairwise quantity and then question what it can say about the portfolio tail.
Start with the events, not a correlation label
For each borrower, define a default indicator that is one if default occurs within the selected horizon and zero otherwise. The marginal default probability describes each borrower separately. The joint default probability describes the chance that both default during that same horizon. If the borrowers were independent, the joint probability would equal the product of their individual probabilities. A larger joint probability indicates positive default association. The distinction matters because a loan book can contain individually low-risk names that are exposed to the same industry shock or economic contraction. The BIS discussion of credit portfolio diversification links unexpected loss to exposure size, variation in default rates and default dependence across segments.
The standard pairwise default-correlation formula is the covariance of the two binary default indicators divided by their standard deviations. In words, subtract the independence benchmark from the joint default probability, then divide by the square root of the product of each marginal probability times its complement. The formula uses probabilities on the same horizon and under comparable definitions of default. Mixing a one-year probability for one borrower with a multi-year joint estimate for the pair does not produce a meaningful correlation. Neither does substituting a correlation between bond price changes or equity returns: those are continuous market variables, whereas default indicators record discrete events.
A worked example with invented probabilities
Assume two loans have illustrative one-year default probabilities of 10% each. Suppose the illustrative probability that both default in the year is 4%. Independence would imply a 1% joint probability, because 10% multiplied by 10% is 1%. For each binary indicator, the variance is 10% multiplied by 90%, or 0.09 in decimal units. The covariance is 0.04 minus 0.01, or 0.03. Dividing 0.03 by 0.09 gives a default correlation of one-third, approximately 33.3%. These numbers are constructed for arithmetic clarity, not estimated from borrowers or observed market data.
The four mutually exclusive outcomes provide a useful check. Both borrowers default with probability 4%. Only the first defaults with probability 6%, and only the second with 6%, because each borrower’s total default probability is 10%. Neither defaults with probability 84%. The four figures add to 100%, and each borrower still has a 10% marginal default probability. Under independence, the comparable probabilities would be 1%, 9%, 9% and 81%. Positive dependence has moved probability away from the middle, one-default outcomes, towards the two extremes. This is why a portfolio loss distribution can change markedly even though the expected number of defaults has not changed.
If both illustrative loans have equal exposure and equal loss severity, the expected number of defaults is 0.2 under either dependence assumption: 10% for the first plus 10% for the second. The likelihood of losing on both loans, however, rises from 1% under independence to 4% in the correlated case. The expected loss is unchanged because expectations add across loans; the shape of outcomes is different. A manager concerned with a high loss quantile therefore cares about dependence. It would be wrong to claim that more correlation mechanically raises expected loss when marginal probabilities, exposure and loss severities are fixed. The Basel Committee explanation of expected and unexpected loss distinguishes those concepts and discusses the systematic factor affecting many borrowers.
Why one number is a fragile portfolio description
Joint defaults are difficult to estimate precisely because they are less frequently observed than individual defaults. If a historical window contains no severe downturn, a calculated pairwise measure may say little about a downturn. If a short stressed window dominates the sample, the estimate may be unstable in another direction. Borrowers may enter and leave the portfolio, credit standards may change, and default definitions or observation horizons may differ. The estimated correlation can therefore reflect the sample construction as much as a persistent economic relationship. A candidate should ask what population and period generated the estimate, whether the period included shared stress and how sensitive the result is to reasonable alternative windows.
A large book also has far more possible patterns of defaults than any practical table of pairwise estimates can describe. With three borrowers, for example, there are eight possible default-or-survival configurations. Pairwise correlations constrain some relationships but do not by themselves determine the probability of a three-name default cluster. For many borrowers, separate estimates for every pair become noisy and hard to maintain. Models commonly introduce factors or common-correlation assumptions to make the problem tractable. Those assumptions are simplifications, not facts about every pair of names. A concentrated portfolio in one industry can behave differently from a geographically and economically diverse one even if a summary statistic looks similar.
Credit exposures also need not be fixed at the time default arrives. A borrower may draw a facility before distress, collateral values may fall, or recoveries may deteriorate when many firms fail together. A binary default correlation says when default indicators move together; it does not fully specify exposure at default, loss given default, mark-to-market movements before default, or settlement risk. That is a crucial boundary when applying a simple calculation to derivatives, guarantees or revolving credit. The model can be internally correct and still omit an important source of loss. The risk question is therefore broader than finding a correlation coefficient that appears precise to two decimal places.
How diversification changes the question
Adding more small loans can reduce the importance of any one borrower’s idiosyncratic failure. It cannot make a shared recession or sector shock disappear. The Basel Committee’s single-factor model note explains why individual risks tend to cancel across many small exposures while systematic risk remains. This is a model insight, not a promise that a particular portfolio is diversified. If many names rely on the same demand driver, source of funding or collateral market, the apparent spread across legal entities may conceal a common exposure. Conversely, a portfolio with a large single borrower has concentration risk even if pairwise correlations are low.
When reviewing a credit-risk result, separate three statements. The marginal default probabilities and loss severities drive the average loss under the stated assumptions. Dependence affects how often many losses arrive together. Exposure concentration determines how much any particular default event matters. This separation avoids two common mistakes: treating default correlation as the whole credit-risk model, and treating a low expected loss as evidence that the extreme-loss tail is small. The GARP Part II study materials place credit-risk measurement within a wider risk-management curriculum, which is the proper setting for the number.
A practical exam reading order
A final limitation is that default correlation is conditional on how the event is defined. A missed payment, a bankruptcy filing and a rating agency default classification may not enter the same data set at the same time. Likewise, two loans that mature on different dates can produce an awkward common horizon. If the estimate comes from traded credit prices, it may reflect risk premia and changing market expectations as well as realised events. These differences are reasons to document a model input and test alternatives, not reasons to discard dependence entirely. In a decision meeting, a transparent range of plausible tail outcomes is more informative than an unexplained point estimate. That is particularly true when the outcome turns on a rare combination of defaults that the historical sample has scarcely observed.
First identify the default horizon and the two marginal probabilities. Then find the joint probability and compare it with the independence product before calculating the standardized correlation. Check that the resulting four outcome probabilities are non-negative and sum to one. Next ask whether the question concerns average loss, tail loss, or the reliability of the estimated input. If a proposed conclusion jumps from pairwise correlation straight to a precise portfolio capital figure, look for assumptions about exposure size, recovery, common factors and calibration. The FRM® programme page shows the broader study context, while Omni’s how-it-works page explains the available learning route.
Sources
- FRM Study Materials | GARP · garp.org · Retrieved
- Bank Mergers, Diversification and Risk | BIS · bis.org · Retrieved
- An Explanatory Note on the Basel II IRB Risk Weight Functions | BIS · bis.org · Retrieved



