FRM Exam Part I · Operational Risk

Loss Distribution Approach in Operational Risk Modeling

Updated 11 October 2026 · Fact-checked

The loss distribution approach (LDA) models operational losses with two distributions: how often losses occur (frequency, often Poisson) and how large each one is (severity, often lognormal). You aggregate them into an annual loss distribution, usually by Monte Carlo simulation. Operational VaR is a high percentile of that distribution, such as the 99.9th.

Understand Loss Distribution Approach and Modeling

Operational losses come from failed processes, people, systems or external events. Most years you see many small losses and a few huge ones. The loss distribution approach (LDA) captures this by modeling two things separately and then combining them.

Frequency is the number of loss events in a year, N. It is a discrete count, so you use a distribution such as the Poisson (mean = variance = λ) or the negative binomial (variance larger than the mean, which suits clustered events). Severity is the size of one loss, X. It is positive and right-skewed with a heavy tail, so you use the lognormal, or fatter-tailed choices such as Weibull, Pareto or a generalized Pareto fitted to the tail.

The aggregate annual loss is S = X1 + X2 + ... + XN. The usual assumptions are that severities are independent and identically distributed, and that they are independent of frequency. The distribution of S has no simple closed form, so you build it by Monte Carlo simulation: draw N, draw N severities, add them, and repeat many thousands of times. You can also use analytic methods such as recursion or Fourier transforms.

From the simulated distribution you read off the expected loss (the mean) and operational VaR, the loss at a high confidence level. The Basel advanced measurement approach used 99.9% over one year. Unexpected loss is VaR minus expected loss, and capital is meant to cover it. Because the tail is heavy and data are scarce, VaR is very sensitive to the severity tail you choose. Modelers usually fit separate distributions by business line and event type and then combine them with a dependence assumption such as a copula. Treating them as independent can understate risk.

Key formulas to remember

Poisson frequency: P(N = n) = e^(-λ) × λ^n ÷ n!; Mean and variance both equal λ. P(N = 0) = e^(-λ).
Aggregate loss: S = X1 + X2 + ... + XN; Assumes severities are i.i.d. and independent of N.
Expected aggregate loss: E[S] = E[N] × E[X]; Holds when N is independent of the i.i.d. severities.
Variance of aggregate loss: Var(S) = E[N] × Var(X) + Var(N) × (E[X])²; For Poisson frequency this simplifies to Var(S) = λ × E[X²], where E[X²] = Var(X) + (E[X])².
Lognormal severity mean: E[X] = exp(μ + σ²/2); μ and σ are the mean and standard deviation of ln X, not of X. The median is exp(μ).
Operational VaR: VaR(α) = the α-quantile of S; Basel AMA used α = 99.9% over one year.
Unexpected loss: UL = VaR(α) − E[S]; Capital is intended to cover UL.
Single-loss approximation: VaR(α) ≈ F⁻¹(1 − (1 − α) ÷ λ); F is the severity CDF. It is an approximation for heavy-tailed severity and high α, with Poisson frequency. It is not exact.

How to solve Loss Distribution Approach and Modeling questions

Most LDA questions ask you to identify the right distributions, compute a moment of the aggregate loss, or read a VaR. Work through the same sequence each time.

1Identify the frequency distribution and its parameter (usually λ for Poisson) and the time horizon, normally one year.
2Identify the severity distribution and its parameters. For a lognormal, check whether the numbers given are for ln X or for X itself.
3Compute the severity moments you need: E[X] and, for the variance of S, E[X²] = Var(X) + (E[X])².
4Aggregate. Use E[S] = E[N] × E[X] and Var(S) = E[N] × Var(X) + Var(N) × (E[X])², or λ × E[X²] for Poisson.
5For VaR, decide the confidence level and horizon. Use the simulated percentile if given, or the single-loss approximation if the question hints at it.
6Convert a severity percentile with z-values if lognormal: X = exp(μ + z × σ).
7Subtract the expected loss if asked for unexpected loss, and check that units and currency are consistent.
8Sanity check: VaR should be far above the mean, and Var(S) should exceed what frequency alone would give.

Quickest way: Moments first, tail second

When to use it: Use this when the question gives parameters and asks for expected loss, standard deviation, or an approximate high-percentile VaR.

Write E[S] = λ × E[X] immediately. This often answers part of the question.
For standard deviation with Poisson, use √(λ × E[X²]). Do not use λ × Var(X).
For VaR at 99.9% with Poisson λ, use the severity quantile at 1 − 0.001 ÷ λ.
Remember common z-values: 3.09 for 99.9%, 3.72 for 99.99%.
Eliminate answer options where VaR is below expected loss or where variance ignores frequency.

Common mistakes in Loss Distribution Approach and Modeling

Using Var(S) = λ × Var(X) for a Poisson frequency.
Students forget that the number of losses is itself random and adds variance.
Fix: Use λ × E[X²], which equals λ × (Var(X) + (E[X])²). The general form is E[N] × Var(X) + Var(N) × (E[X])².
Treating μ in a lognormal as the mean loss.
The parameters sound like the mean and standard deviation of the loss itself.
Fix: μ and σ describe ln X. The mean loss is exp(μ + σ²/2) and the median is exp(μ).
Adding severity VaR and frequency VaR separately.
Students think the two components can be stressed one after the other.
Fix: Aggregate through simulation or the formulas above, then take the percentile of S.
Confusing operational VaR with unexpected loss.
Both are called capital in different contexts.
Fix: VaR is the percentile itself. Unexpected loss is VaR minus expected loss. Read the question for which one is asked.
Assuming the 99.9% VaR is reliable because the model is fitted.
Students overlook how little tail data exist.
Fix: State that the result is sensitive to the severity tail, data scarcity and dependence assumptions. Collecting external data and scenario analysis help, but do not remove the problem.
Using the wrong time horizon for λ.
Frequency is quoted per month or per quarter in the question.
Fix: Convert λ to the VaR horizon first. A monthly λ of 2 is an annual λ of 24.

Worked examples

Example 1

A bank models one business line with Poisson frequency λ = 12 losses per year. Severity has mean $50,000 and standard deviation $80,000, independent of frequency. Find the expected annual loss and the standard deviation of annual loss.

Show the solution

Expected loss: E[S] = λ × E[X] = 12 × 50,000 = $600,000.
Second moment: E[X²] = Var(X) + (E[X])² = 80,000² + 50,000² = 6,400,000,000 + 2,500,000,000 = 8,900,000,000.
Variance of S for Poisson: Var(S) = λ × E[X²] = 12 × 8,900,000,000 = 106,800,000,000.
Standard deviation: √106,800,000,000 ≈ $326,800.

Answer: Expected annual loss is $600,000 and the standard deviation is about $326,800.

Example 2

Operational losses have Poisson frequency λ = 10 per year and lognormal severity with μ = 10 and σ = 2 (parameters of ln X, in USD). Using the single-loss approximation, estimate the 99.9% annual VaR and the unexpected loss.

Show the solution

Severity quantile level: 1 − (1 − 0.999) ÷ λ = 1 − 0.001 ÷ 10 = 0.9999.
The z-value for 0.9999 is about 3.719.
Severity quantile: ln X = 10 + 2 × 3.719 = 17.438, so X = e^17.438 ≈ $37.4 million. This is the approximate 99.9% VaR.
Mean severity: exp(10 + 2²/2) = e^12 ≈ $162,755.
Expected annual loss: E[S] = 10 × 162,755 ≈ $1.63 million.
Unexpected loss: VaR − E[S] ≈ 37.43 − 1.63 ≈ $35.8 million.

Answer: The 99.9% VaR is about $37.4 million and the unexpected loss is about $35.8 million. These are approximations because the single-loss formula is not exact.

Exam tips

Know which quantity each question wants: mean, standard deviation, VaR or unexpected loss. The options usually include the others as traps.
Expect conceptual questions on why Poisson fits frequency and why heavy-tailed distributions fit severity.
For lognormal items, always check whether the given numbers are for ln X or for X.
Understand the limits: scarce tail data, sensitivity to the severity choice and dependence between categories. These are favorite qualitative points.
A calculator is enough for the moment formulas. Keep e^x and ln handy for the lognormal and Poisson steps.

Practice questions from Operational Risk

Loss Distribution Approach and Modeling in other exams

The same ground in other exams, if you are preparing for more than one or want another angle on it.

FRM Part IILoss Data Collection and Operational Risk Capital Models

Loss Distribution Approach and Modeling: frequently asked questions

What is the loss distribution approach in operational risk?

It models loss frequency and loss severity separately, aggregates them into an annual loss distribution, and takes a high percentile as operational VaR. It is the core idea behind the Basel advanced measurement approach.

Why is Poisson used for frequency and lognormal for severity?

The Poisson fits counts of rare, independent events. The lognormal is positive, right-skewed and has a long tail, which matches operational loss sizes. Heavier tails such as Pareto are often used for extreme losses.

How is operational VaR calculated under the LDA?

You simulate many years of losses by drawing a count and then that many severities, and sum each year. You then read the chosen percentile, typically 99.9%, from the sorted results. Unexpected loss is VaR minus the mean.

Is Var(S) equal to λ times Var(X)?

No. That ignores the randomness of the number of losses. For Poisson frequency, Var(S) = λ × E[X²], which is larger than λ × Var(X).