1Write a portfolio credit model as a Bernoulli mixture and use conditional independence to compute its loss distribution.
2Compute the corporate book's loss distribution exactly, with and without a common factor, and read its mean, standard deviation, VaR and expected shortfall.
3Show how a pool of equal loans converges to the Vasicek limit as it grows, and why the binomial does not.
4Explain why one common factor shapes the tail, and build the whole lending book's distribution from names and large pools.
Why it matters
RiskAtoms Bank plc lends £5,072m and expects to lose £22.77m a year. In one year in a hundred its lending book loses more than £86.4m; in one year in a thousand, more than £135.2m, 6 times the average. Nothing in the expected loss tells you that. The tail comes from one fact: borrowers share an economy. L3.3.11: Unexpected loss and default correlation showed it for twelve names. This lesson gives the general recipe and runs it on the whole book.
Booked lending bookRiskAtoms Bank plc's lending book at 31 December 2025
The portfolio is RiskAtoms Bank plc's booked lending book from M3.3: Credit risk, unchanged: 12 corporate names (£587.0m of exposure at default) and 8 retail segments of 152,100 accounts (£4,485.0m), £5,072.0m in all, with an expected loss of £22.77m a year. The bank and its borrowers are invented. PDs, LGDs and exposures are the booked ones; the corporates share the booked asset correlation of 0.2, and each retail segment takes the Basel retail correlation of its class. The booked figures of earlier modules stand; model figures sit beside them.
One recipe: the Bernoulli mixture
Every portfolio credit model in this module has the same skeleton. Each borrower has a default indicator, one if it defaults this year (a Bernoulli variable; M-C04: The binomial distribution for the binomial count of many). Given the state of the economy, borrowers default independentlyConditional independence: two events are independent once a third quantity (here the economy) is known, though they move together when it is not., each with a conditional PD that rises in bad years. The economy is random, so averaged over it the defaults are not independent. That is a Bernoulli mixture modelBernoulli mixture model: a model in which, given some common factors, borrowers default independently, each with its own conditional probability; the factors are random, so the defaults are not.. Frey and McNeil (2003) showed that the threshold models banks use, in which a firm defaults when its asset value falls below a level, can be written this way, and that the mixing distributionMixing distribution: the distribution of the common factor, or of the default rate it produces; it decides the shape of the loss tail., how the conditional PD varies with the economy, is what decides the tail.
The bank's choice is the one-factor model of L3.3.11: Unexpected loss and default correlation: one standard normal factor Z, and each name's conditional PD Φ((Φ−1(PD) − √ρ Z)/√(1 − ρ)), with asset correlation ρ = 0.2 for the corporates. In a year as bad as the worst in a thousand (Z = −3.09), Bramwick Stores plc's PD of 1.13% becomes 15.8%; in a good year (Z = 1) it is 0.11%. Conditional independence turns a hard problem, the joint behaviour of many correlated names, into an easy one repeated: independent names, one economy at a time.
The corporate book, exactly
Twelve names have 212 = 4,096 patterns of default. For each value of Z the probability of every pattern is a product; integrating over Z (Simpson's rule on 800 intervals) and adding up the patterns by loss gives the whole distribution. With no factor the integral is not needed:
Exact over all 4,096 outcomes. The mean, sd, VaR and P(no loss) equal the booked figures of L3.3.11: Unexpected loss and default correlation (£44.38m and £56.54m at 99.9%); expected shortfall is new here.
Corporate book, one year
Independent
One factor, ρ 0.2
Expected loss
£2.81m
£2.81m
Standard deviation
£6.18m
£6.86m
P(no loss)
77.6%
79.6%
99% VaR
£34.20m
£34.20m
99.9% VaR
£44.38m
£56.54m
99% expected shortfall
£37.14m
£42.94m
99.9% expected shortfall
£49.03m
£63.71m
Unexpected loss at 99.9% (VaR − EL)
£41.57m
£53.72m
The factor leaves the mean alone, makes loss-free years more common, and pushes the 99.9% loss up by £12.2m. Expected shortfallExpected shortfall (ES) at q: the average loss in the worst 1 − q of years, with the probability at the VaR point split so it counts exactly 1 − q. asks the next question: how bad is the average year beyond the credit VaRCredit VaR at q: the smallest loss that is not exceeded with probability at least q over the year.? At 99.9% it is £63.71m. On a stepped distribution ES needs care: the outcome sitting at the VaR point is split, so exactly the worst 0.1% of years is averaged. Acerbi and Tasche (2002) showed that this version stays coherent for any distribution.
Why one factor shapes the tail
Take the losses above £34.2m, beyond the 99% point. With the factor they happen in 0.70% of years against 0.29% without it. More telling is when they happen. Split the probability by the year's Z:
Exact: the tail probability given Z, integrated over the stated range of Z (400 intervals).
Share of P(loss > £34.2m) from years with…
One factor
Independent
Z in its worst 0.1%
5.4%
0.1%
Z in its worst 1.0%
25.4%
1.0%
Z in its worst 5%
55.5%
4.8%
Z in its worst 10%
72.0%
9.8%
Without the factor the big losses are spread over years in proportion: 4.8% of them come from the worst 5% of years. With it, 55.5% do, and 25.4% from the worst 1%. The tail is a story about bad years. That is why the regulatory formula of L6.4.02: The asymptotic single risk factor model can fix the economy at one bad value and read the loss off.
A pool of equal loans and its limit
The same recipe on 100 equal loans with PD 3.5% and ρ = 0.1 (L3.3.11: Unexpected loss and default correlation's band B pool) gives a default correlation of 0.0207. Small as it is, it raises the standard deviation of the count from 1.84 to 3.21 and moves the 99.9% count from 10 to 21 defaults. Now let the pool grow:
PD 3.5%, ρ 0.1; exact counts, the 99.9% count divided by the pool size.
Loans
99.9% default rate, independent
99.9% default rate, one factor
10
30.00%
40.00%
100
10.00%
21.00%
1,000
5.40%
19.20%
10,000
4.08%
18.97%
Infinite (the limit)
3.50%
18.95%
Independent defaults diversify: the bad-year rate falls towards the PD. Correlated defaults do not: the rate settles on 18.95%, the 99.9% point of the Vasicek large-pool limitLarge-pool (Vasicek) limit: the distribution of the default rate of an infinitely large pool of equal loans, in which only the common factor is left.. In a big enough pool the borrowers' own luck cancels and the default rate is the conditional PD itself, a function of Z alone (Vasicek, 1991 and 2002). The limit has a price: it depends entirely on the mixing distribution. A beta distribution for the default rate with the same PD and the same default correlation puts the 99.9% rate at 16.87%, not 18.95%. Two models that agree on the first two moments disagree about the tail.
The whole book
The bank's retail book is 8 segments of 3,100 to 52,000 loans each, so each one is a pool at its limit: given Z it loses EAD × LGD × p(Z), with the Basel retail correlation of its class. The twelve names stay exact. Given Z, the book's loss is one of 4,096 corporate outcomes plus the pools' smooth conditional loss, so P(L ≤ x) solves, for each outcome, for the Z at which the pools fill the gap.
Names exact, pools at their large-pool limit, Z on 1,600 intervals; no outcome dropped.
Whole lending book, one factor
Exact
Expected loss
£22.77m
Standard deviation
£16.91m
99% VaR
£86.36m
99.9% VaR
£135.24m
99% expected shortfall
£107.41m
99.9% expected shortfall
£158.30m
Retail drives this tail. At Z's 1-in-1,000 value the pools alone lose £99.3m and the corporates' conditional expected loss is £27.2m: £126.5m together. The exact 99.9% loss is £8.7m higher, because twelve lumpy names do not behave like their averages. That gap is the subject of L6.4.02: The asymptotic single risk factor model.
P(Yi=1∣Ψ)=pi(Ψ),Y1,…,Yn independent given ΨP(Y=y)=∫i=1∏npi(ψ)yi(1−pi(ψ))1−yidF(ψ),L=i=1∑nYiLGDiEADione factor: pi(z)=Φ(1−ρiΦ−1(PDi)−ρiz),Z∼N(0,1)ρD=p(1−p)Var(p(Z))=p(1−p)Φ2(Φ−1(p),Φ−1(p);ρ)−p2(two loans in one pool)large pool: nDn→p(Z),P(p(Z)≤x)=Φ(ρ1−ρΦ−1(x)−Φ−1(p))VaRq=min{x:P(L≤x)≥q},ESq=1−q1(E[L1{L>VaRq}]+VaRq(P(L≤VaRq)−q))whole book: P(L≤x)=y∑∫z≥zy∗(x)P(Y=y∣z)φ(z)dz,r(zy∗)=x−l(y)
Yi is name i's default indicator, Ψ the common factors (here one, Z), F their distribution and φ its density; Φ is the standard normal cdf and Φ2 the bivariate one; Dn the number of defaults among n equal loans; l(y) the corporate loss of outcome y and r(z) = Σs EADs LGDs ps(z) the pools' conditional loss, which falls in z. Integrals in z run over [−8, 8] (Simpson, 800 intervals; the whole book on a 1,600-interval grid). Sources: Vasicek (1987, 1991, 2002); Frey and McNeil (2003); McNeil, Frey and Embrechts (2015), chapter 11; Acerbi and Tasche (2002) for ES.
Try it
Start with the pool of 100: correlation doubles its bad-year count. Switch to 10,000 loans and watch the independent bars collapse onto the PD while the correlated ones match the grey Vasicek limit. Then switch to the corporate book and slide ρ from 0 to 0.5: the 99% loss sits on Bramwick Stores plc until ρ passes about 0.3, while the 99.9% loss and ES climb.
Try itBuild a loss distribution from a mixture
What to model
Loans in the pool
Expected defaults (both)3.5
Default correlation of two loans0.0207
Standard deviation: independent / one factor1.8 / 3.2
P(no default): independent / one factor2.84% / 13.0%
Grow the pool with ρ above zero: the one-factor bars settle onto the grey Vasicek bars, while the binomial bars pile up at the PD. Idiosyncratic luck diversifies away; the shared factor does not. Rates above 40% fall in the last bin.
Worked example
The corporate book's 1-in-1,000 loss
From conditional PDs to the 99.9% loss
Step 1 of 4
1
Fix the economy
Take one bad value, Z = −3.0902. Each name's conditional PD follows; for Bramwick Stores plc and Graybeck Plastics Ltd:
and 57.6% for Graybeck Plastics Ltd (PD 11.31%). Given Z the twelve names are independent.
Regulatory lens
The IRB formula is the large-pool limit of this model at 99.9% over one year, applied loan by loan (Basel Framework CRE31; the BCBS explanatory note of July 2005): L6.4.02: The asymptotic single risk factor model derives it and measures what it leaves out. Where a bank's portfolio is less diversified than the formula assumes, the explanatory note expects it to be addressed under Pillar 2, and banks' own economic capital models are full loss distributions like the ones on this page (L6.4.08: Economic capital and risk-adjusted performance). Expected shortfall is the market risk measure of the Basel trading book rules (M6.2: Market risk modelling); for credit, regulatory capital still rests on a quantile.
Spark Atoms
✦Spark AtomA memo, fifteen years early+
Oldřich Vasicek worked out the loss distribution of a loan portfolio in a memo dated 12 February 1987, two years before he co-founded KMV, and its large-pool limit in a KMV memo in August 1991. The paper most people cite, “The distribution of loan portfolio value”, appeared in Risk in December 2002 as “Loan portfolio value”, by which time the Basel Committee was building its capital formula on the same result.
✦Spark AtomSmall changes, large tails+
Frey and McNeil (2003) put it in one line: small changes to the structure of a model, or to the parameters describing dependence, can have a large impact on the loss distribution, and in particular on its tail. Once PD and default correlation are fixed, they found, the form of the mixing distribution matters little until beyond the 99% point: the beta and probit-normal pools above agree on both and still differ by 2.1 points at 99.9%.
Key takeaways
A Bernoulli mixture model makes defaults independent given common factors; averaging over the factors makes them correlated.
Conditional independence turns the twelve-name book into 4,096 products and one integral: 99.9% £56.54m with ρ 0.2, £44.38m without, as booked in L3.3.11: Unexpected loss and default correlation.
The factor puts the tail into bad years: 56% of the corporate losses beyond 99% come from the worst 5% of years.
A growing pool tends to the Vasicek limit (18.95% at 99.9% here); without correlation it tends to the PD.
The mixing distribution decides the tail: a beta mixture with the same PD and default correlation gives 16.87%.
The whole book under one factor: VaR 99.9% £135.2m, ES £158.3m; retail pools drive it.
Reaction
Check what stuck
Question 1 · objective 1
In a Bernoulli mixture model, what is independent?
Question 2 · objective 3
A very large pool has PD 2% and asset correlation 0.15, so a default correlation of 0.0243. What is the standard deviation of its yearly default rate, in percentage points? (Two decimals.)
points
Question 3 · objective 2
The corporate book's 99% loss is £34.2m with or without the factor, yet its 99.9% loss is £44.38m independent and £56.54m with ρ 0.2. Why does the factor matter only further out?
Question 4 · objective 2
In a model the 1% worst years are three equally likely outcomes beyond VaR, with losses of £40m, £50m, £90m. What is the 99% expected shortfall, in £m?
£m
Question 5 · objective 4
Retail is 88% of the book's exposure, in 152,100 accounts. Why can the model treat each retail segment as a pool at its large-pool limit but must treat the twelve corporates name by name?
Glossary
Bernoulli mixture model
Bernoulli mixture model: a model in which, given some common factors, borrowers default independently, each with its own conditional probability; the factors are random, so the defaults are not.
Conditional independence
Conditional independence: two events are independent once a third quantity (here the economy) is known, though they move together when it is not.
Mixing distribution
Mixing distribution: the distribution of the common factor, or of the default rate it produces; it decides the shape of the loss tail.
Large-pool (Vasicek) limit
Large-pool (Vasicek) limit: the distribution of the default rate of an infinitely large pool of equal loans, in which only the common factor is left.
Credit VaR
Credit VaR at q: the smallest loss that is not exceeded with probability at least q over the year.
Expected shortfall
Expected shortfall (ES) at q: the average loss in the worst 1 − q of years, with the probability at the VaR point split so it counts exactly 1 − q.