Account for the typical features of operational risk exposures and loss events, and for the management difficulties those features create.

FRM® Part II · Operational Risk and Resilience · Introduction to Operational Risk and Resilience · FRM2.ORR.1.c

The idea

Operational loss events are typically many small, frequent losses plus rare, extremely large ones. The severity distribution is therefore heavy-tailed, and a handful of events dominate the total. Expected annual loss is frequency times mean severity, summed by category, and frequency is often modelled as Poisson. Operational risk is a by-product of doing business rather than a risk taken for a reward. It is hard to diversify, and it is often idiosyncratic to a firm. Management difficulties follow. Internal data are scarce in the tail, external data need scaling, recording thresholds and reporting biases distort samples, and losses can be discovered years after they occur. A sample mean or a historical maximum is not a tail quantile. Examiners test the link between each feature and the measurement problem it causes.

Formulas

Expected annual operational loss across categories

expected annual loss = Σ (fᵢ × sᵢ)

fᵢ: the expected number of loss events a year in category i
sᵢ: the mean severity of a loss in category i

Frequency and severity are taken per category and then summed. Adding the mean severities or applying total frequency to an average severity are the usual slips, and the result says nothing about the tail.

Probability of at least one Poisson event

P(N ≥ 1) = 1 − e^(−λ × t)

λ: the arrival rate per unit of time
t: the length of the period

For large losses only, λ is the event rate times the chance that an event exceeds the size, assuming independence. Reporting λ itself, or e^(−λ × t), the chance of none, are the traps.

A practice question

From the question bank, with the reason each wrong option is wrong.

Which description best fits the typical distribution of operational loss severity?

Symmetric and bell-shaped around a stable average loss, so large losses are very unlikelyWhy not: Assumes a normal shape; operational severity is highly skewed with fat tails.
Uniform between a small floor and a fixed maximum, so no loss can exceed a known capWhy not: Ignores that there is no natural cap on the largest loss.
Many small losses and a few huge ones: a heavy right tail - correct
Large and frequent losses with only rare small ones, so the average is dominated by large eventsWhy not: Reverses the pattern: large losses are the rare ones.

Severity is heavy-tailed: most events are small and a few extreme ones dominate total losses, which is why average loss is a poor guide to capital.

A harder one

A rogue-trading fraud began in 2019 and was discovered in 2023. Which feature of operational losses does this illustrate, and which management difficulty follows?

High-frequency, low-severity losses, so the data are abundant and trends reliableWhy not: Misreads the example: it is a rare, severe, late-discovered event, and the data are not reliable.
Long lags before discovery, so recent data understate current losses and distort trends - correct
Correlation with market prices, so losses can be hedged with derivativesWhy not: Operational losses are not generally hedgeable with market instruments.
Immediate recognition, so the occurrence, discovery and accounting dates always coincideWhy not: The case shows the opposite: occurrence, discovery and accounting dates differ.

Losses are often discovered and booked long after the events occurred, so the latest periods appear cleaner than they are, which distorts trend analysis and KRI interpretation.

The hardest tier

A bank expects 8 operational loss events a year, and each has a 5% chance of exceeding USD 1 million, independently. Assuming events follow a Poisson process, the probability of at least one loss above USD 1 million in a year is closest to:

40%Why not: Treats the expected number of large losses (0.4) as if it were a probability.
67%Why not: Reports the probability of no large loss, not at least one.
33% - correct
100.0%Why not: Applies the total event rate of 8 and ignores the 5% filter.

Large losses form a Poisson process with rate 8 x 0.05 = 0.4. P(at least one) = 1 - exp(-0.4) = 0.3297, about 33%.

In the app: 6 practice questions on this outcome, spaced review that brings it back before you forget it.

Start free →