RiskFree guide

Market and credit risk

Measure risk like a bank risk team: VaR (value at risk) and its blind spots, backtesting, expected shortfall and stress tests, counterparty exposure and CVA (credit valuation adjustment), expected loss and the other risks that sink banks.

Read first: Statistics and machine learning

7 concepts8 min read25-question test, from a bank of 36Updated

Concept 1 of 71 min

Computing VaR: parametric, historical and combined

The rule
A normal VaR is the multiplier × volatility × value, 1.65 at 95% and 2.33 at 99%, and it is a threshold, not the largest possible loss.
Worked examples

A parametric VaR, in millions

Given
Portfolio250
Daily volatility0.8%
Multiplier at 99%2.33
  1. VaR = 2.33 × 0.008 × 250 = 4.66

A 1-day 99% VaR of 4.66 million: the loss stays below it 99 days out of 100, and can be far larger on the 100th.

From 1 day to 25

Given
1-day VaR2,000,000
Horizon25 days
  1. 25-day VaR = 2,000,000 × √25 = 10,000,000

Risk grows with the square root of time, provided the days are independent.

2 uncorrelated desks

Given
Desk A VaR6
Desk B VaR8
Correlation0
  1. Sum of squares = 62 + 82 = 100
  2. Combined VaR = √100 = 10

Together the desks need 10, not the 14 of adding them up.

The picture
Formula
VaR=z × σ × value
VaR over n days=1-day VaR × √n
uncorrelated VaR=√(VaR₁² + VaR₂²)
historical VaR at 99% over N days ≈ the (N × 1%)th worst loss
Why it holds

It scales with √days for independent days. Historical simulation replays past moves on today's positions. VaRs of desks with uncorrelated P&L (profit and loss) combine like volatilities, so a hedge can lower the total.

Scaling linearly assumes every day's loss adds up perfectly. At 99% over 500 days, about 5 days lie beyond the VaR, so it is roughly the 5th worst loss; conventions differ slightly.

Adding VaRs is the case of perfect correlation; uncorrelated books combine as the square root of the sum of squares. A new trade with its own VaR can still reduce the desk's VaR if it offsets existing risk.

Losses beyond the threshold can be much larger, which is the point of expected shortfall.

Watch out
  • Scaling VaR linearly with time.
  • Adding the VaRs of uncorrelated desks.
  • Calling VaR the worst loss possible.
In the interview
  1. How would you compute a historical VaR?

  2. What does a 1-day 99% VaR of 10 million mean, and what does it not tell you?

  3. Explain VaR to a board member who has never heard of it, and tell them what it does not say.

    A complete answer covers

    • It is a loss we expect to exceed only rarely, say 1 day in 100.
    • It comes from past market moves applied to today's positions.
    • It does not say how bad losses are beyond that level.
    • So we also run stress tests and look at expected shortfall.

Also in our technical interviews, on camera, with a score.

Concept 2 of 71 min

Where VaR fails: tails, liquidity, options and backtesting

The rule
A normal-distribution VaR understates fat tails, assumes positions can be sold within the horizon and misses the non-linear losses of short options.
Worked example

Backtesting a 99% VaR

Given
VaR level99%
Tail1%
Days observed250
Exceptions seen7
Amber zone5 to 9
Red zone from10
  1. Exceptions expected = 250 × 0.01 = 2.5
  2. 7 sits between 5 and 9: the amber zone

Almost 3 times the expected exceptions: amber, short of the red zone.

The picture
Formula
Expected exceptions=days × (1 − confidence)
99% over 250 days: 2.5 expected; green 0 to 4, amber 5 to 9, red 10 or more
Why it holds

Short options, out-of-the-money ones included, need full revaluation and gamma limits. Backtesting counts exceptions: at 99% over 250 days, 10 or more is Basel's red zone.

Large moves happen far more often than the normal predicts. Selling a big block of an illiquid bond takes time and moves the price.

A desk that is flat delta but short gamma loses on large moves in either direction, which a delta limit does not see.

About 2.5 exceptions are expected in 250 days at 99%; Basel's traffic lights run green up to 4, amber from 5 to 9 and red from 10.

Watch out
  • Trusting a normal VaR on fat-tailed returns.
  • Valuing an illiquid position as if it could be sold at once.
  • Limiting an options desk on delta alone.
In the interview
  1. Your VaR model had 9 exceptions this year at 99%. What do you do?

    A complete answer covers

    • Recognise it is well above the 2 or 3 expected: the model understates risk.
    • Check the exceptions: fat tails, missing risk factors, stale data, option effects.
    • Fix the model or add capital and tighten limits meanwhile.
    • Report it to model validation and the regulator as required.

Also in our technical interviews, on camera, with a score.

Sources: Basel Framework, MAR32.9

Concept 3 of 71 min

Expected shortfall, stressed VaR and stress tests

The rule
Expected shortfall averages losses beyond the VaR; Basel uses it at 97.5%, close to a 99% VaR under normality and larger with fat tails.
Worked example

VaR and expected shortfall (ES)

Given
Daily losses observed500
Confidence99%
Tail1%
Worst loss12
2nd worst9
3rd worst8
4th worst7
5th worst6
  1. Days in the tail = 500 × 0.01 = 5
  2. 99% VaR: about 6, the 5th worst
  3. ES = (12 + 9 + 8 + 7 + 6) ÷ 5 = 8.4

VaR says where the tail starts, 6; expected shortfall says how deep it goes, 8.4.

The picture
Formula
ES=average of losses beyond VaR
normal 97.5% ES ≈ 2.34σ against 99% VaR ≈ 2.33σ
liquidity horizons: 10, 20, 40, 60, 120 days
Why it holds

Basel's liquidity horizons run from 10 to 120 days.

Stressed VaR is calibrated on a past crisis, stress tests price extreme scenarios, and a reverse stress test starts from the loss that would break the firm. VaR says where the tail starts; expected shortfall says how bad it is on average, which is why the Fundamental Review of the Trading Book moved to it.

A 10-day horizon is too short for exotic or illiquid positions. Stressed VaR was added after 2008 so capital does not fall when recent markets are calm.

Stress tests go beyond the data, such as a crash or a rate shock; a reverse stress test exposes hidden vulnerabilities by asking what would sink the firm.

Watch out
  • Saying expected shortfall is smaller than VaR.
  • Thinking stress tests replace VaR or limits.
  • Starting a reverse stress test from a historical scenario.
In the interview
  1. Two portfolios have the same VaR. Why might one be far riskier?

  2. Why did regulators move from VaR to expected shortfall for market risk?

  3. Why did regulators move trading book capital from VaR to expected shortfall?

    A complete answer covers

    • VaR ignores how large losses are beyond the threshold.
    • Expected shortfall averages the tail, so fat tails raise capital.
    • At 97.5% it is close to a 99% VaR in calm, normal markets.
    • It came with longer liquidity horizons for illiquid risks.

Also in our technical interviews, on camera, with a score.

Sources: Basel Framework, MAR33.3

Concept 4 of 71 min

Counterparty risk: exposure, netting, CVA and wrong-way risk

The rule
Counterparty exposure is today's positive value plus potential future exposure minus collateral, and netting offsets trades with the same counterparty.
Worked examples

Exposure to one client, in millions

Given
Value today+8
Add-on (future exposure)3
Collateral6
  1. Exposure = 8 + 3 − 6 = 5

The bank could lose up to about 5 million if the client defaults.

Netting with one counterparty

Given
Trade 1+7
Trade 2−4
Trade 3+2
  1. Netted = 7 − 4 + 2 = 5

Netting cuts the exposure to 5, instead of the 9 owed on the positive trades alone.

The CVA, in millions

Given
Expected exposure10
Default probability2%
Loss given default60%
  1. CVA ≈ 10 × 0.02 × 0.6 = 0.12

The trade is worth about 120,000 less because the client may default.

The picture
Formula
Exposure=max(value, 0)+PFE−collateral
netted exposure=max(Σ values, 0)
CVA ≈ EE × PD × LGD
Why it holds

CVA lowers value by the expected cost of default, about expected exposure × PD (probability of default) × LGD (loss given default), so an uncollateralised BB-rated client costs more than a AAA bank. Wrong-way risk grows as the counterparty weakens.

A derivative worth +5 today can be worth much more tomorrow, so exposure adds what it could become. Without netting, the bank would be exposed to its positive trades and still owe on its negative ones.

An uncollateralised swap with a weak name carries a larger CVA from day one than the same swap with a strong, fully collateralised bank. Protection on a bank bought from a closely tied bank is worth least exactly when it is needed.

Watch out
  • Leaving the potential future exposure out.
  • Adding gross positive values when trades are netted.
  • Buying protection from a seller tied to the reference name.
In the interview
  1. A corporate client wants a large uncollateralised interest rate swap. What would you consider before agreeing?

    A complete answer covers

    • The exposure today and its potential growth over the life of the swap.
    • The client's credit quality and the CVA to charge.
    • Netting with other trades, collateral or break clauses to reduce exposure.
    • Any wrong-way risk between the swap and the client's health.

Also in our technical interviews, on camera, with a score.

Concept 5 of 71 min

Expected loss, capital and concentration

The rule
Expected loss is probability of default × loss given default × exposure at default, covered by pricing and provisions; capital absorbs unexpected losses.
Worked example

Expected loss on a loan, in millions

Given
Loan50
Probability of default2%
Loss given default40%
  1. Loss if it defaults = 50 × 40% = 20
  2. Recovered = 50 − 20 = 30
  3. Expected loss = 0.02 × 20 = 0.4 a year

The 400,000 a year goes into the loan's margin; capital is there for the losses beyond it.

The picture
Formula
EL=PD × LGD × EAD
Why it holds

Expected losses are a cost of doing business; capital is for the bad years. One downturn hits many loans in the same industry together, so sector and single-name limits control it.

Day-to-day costs and dividends are not what capital is for.

Pricing and provisions cover the losses a bank expects on average, while capital covers the unexpected ones at a high confidence level, which is why it is sized on a bad year rather than a normal one.

Watch out
  • Saying capital covers expected losses.
  • Multiplying exposure by the probability of default alone.
  • Ignoring correlation within an industry.
In the interview
  1. Why do lenders hold capital if expected losses are already covered by provisions?

  2. How do you compute the expected loss on a loan, and what is it used for?

  3. A loan officer argues that a sector is safe because each borrower has a low default probability. What is missing from that view?

    A complete answer covers

    • Defaults in one sector tend to happen together.
    • Concentration raises unexpected losses, which capital must cover.
    • Look at the share of the book and at stress scenarios for the sector.
    • Set limits by sector and by name.

Also in our technical interviews, on camera, with a score.

Concept 6 of 71 min

Operational, model, basis and banking-book risks

The rule
Rogue trading is operational risk, from failed people, processes and controls, while model risk is the loss from a wrong or misused model.
Worked example

Hedging with index futures

Given
Portfolio€80 million
Beta0.9
Future size€0.2 million
  1. Futures = 80 × 0.9 ÷ 0.2 = 360

Selling 360 futures of €200,000 each removes the market risk; buying them would double it.

Formula
Futures to sell=value × β ÷ contract notional
Why it holds

A government bond future hedging a corporate bond leaves basis risk. Banking-book rate risk is measured on EVE (economic value of equity) and NII (net interest income).

An equity hedge sells value × beta ÷ contract notional in futures. Markets caused the losses in a rogue trading case, but the failure was in controls.

Mis-specified models, bad data or use outside their limits misprice or mis-hedge; independent validation addresses it.

A rate hedge leaves the bond's credit spread unhedged. EVE looks at the present value of all cash flows, NII at earnings over the next year or two, and a shock can hit one and not the other.

Ignoring beta leaves part of the market risk; buying futures would double it.

Watch out
  • Calling rogue trading market risk.
  • Thinking a government bond future hedges a corporate bond fully.
  • Ignoring beta when sizing an index hedge.
In the interview
  1. A pricing model has been used for years without problems. Why would you still have it reviewed?

    A complete answer covers

    • Markets change and assumptions can stop holding.
    • It may now be used for products it was not built for.
    • Data feeding it may have changed or degraded.
    • Independent validation and limits on use are what control model risk.

Also in our technical interviews, on camera, with a score.

Concept 7 of 71 min

AdvancedWhen VaR punishes diversification

The rule
VaR can rise when 2 risks are combined, so it can penalise diversification, while expected shortfall never exceeds the sum of the parts.
Worked examples

VaR punishes the pair

Given
Bonds2
Defaultsindependent
Loss on default100 each
Default chance4% each
Survival chance96% each
VaR confidence95%
  1. Alone: 4% is below 5%, so the VaR is 0
  2. At least one default = 1 − 0.962 = 7.84%
  3. Together: 7.84% is above 5%, so the VaR is 100

The pair's VaR of 100 is more than the sum of the parts, 0 and 0.

Expected shortfall does not

Given
Tail averagedworst 5%
1 bond alone4% at 100
Both default0.16% at 200
One defaults4.84% at 100
  1. One bond = 4 × 100 ÷ 5 = 80
  2. Pair's tail = 0.16 × 200 + 4.84 × 100 = 516
  3. The pair = 516 ÷ 5 = 103.2

The pair's 103.2 stays below the 160 of the 2 bonds apart: diversification is rewarded.

The picture
Formula
Expected shortfall of a sum ≤ sum of the expected shortfalls
VaR of a sum can exceed the sum of the VaRs
Why it holds

Take 2 independent bonds, each with a chance of default smaller than the 5% tail of a 95% VaR. Alone, each bond's VaR is 0, because its default sits outside the worst 5% of outcomes.

Together, the chance that at least one defaults rises above 5%, so the portfolio's VaR jumps to a full loss: more than the sum of the parts.

Expected shortfall averages the worst 5% of outcomes, so it sees each bond's default even alone, and together the tail mixes double and single defaults, which keeps it below the sum of the parts. VaR ignores what happens beyond its threshold; expected shortfall looks there and rewards diversification.

That is one reason regulators moved to expected shortfall, and why tail risk in credit portfolios is measured with it.

Watch out
  • Assuming VaR always falls when risks are diversified.
  • Reading a VaR of 0 as no risk when the loss sits just beyond the threshold.
  • Adding expected shortfalls and calling it the portfolio's: that is only an upper bound.
In the interview
  1. Two independent loans each have a 95% VaR of zero. Can their portfolio have a positive VaR?

    A complete answer covers

    • Yes: each default chance can sit below 5% while the chance of at least one sits above.
    • VaR then rises with diversification, so it is not subadditive.
    • Expected shortfall averages the tail and never exceeds the sum of the parts.
    • That is one reason the trading-book rules moved to expected shortfall.

Also in our technical interviews, on camera, with a score.

Recap

  • Normal VaR = z × σ × value (1.65 at 95%, 2.33 at 99%); scale by √days; VaRs combine like volatilities, not by addition.
  • VaR is a threshold, not a worst case; it misses fat tails, illiquidity and the gamma of options.
  • Expected shortfall averages the tail; Basel uses it at 97.5% with liquidity horizons of 10 to 120 days.
  • Counterparty exposure = positive value + potential future exposure − collateral, after netting; CVA prices the default risk.
  • Expected loss = PD × LGD × EAD, covered by pricing and provisions; capital covers the unexpected.
Now test it

25 questions on Market and credit risk

Timed, like the real one, drawn from a bank of 36. Every answer explained.

Test yourself →