Articles

Cyber Risk Loss Exceedance Curve: How to Build and Read It

Monte Carlo trials, percentiles as return periods, why a severe scenario can show a $0 90th percentile, gross and net curves, and risk appetite tested as points on the curve.

A loss exceedance curve answers the question boards actually ask about cyber risk: what is the chance that we lose more than a given amount in one year? It plots, for every loss threshold, the probability that the annual loss exceeds it. Read from left to right, it shows how often small losses happen and how fast the probability falls as the amounts grow. This guide builds the curve from a Monte Carlo simulation of five scenarios at Halden Marsh Industries, Inc., a fictional manufacturer, reads it gross and net of insurance, explains why one scenario's 90th percentile is $0, and turns a board's risk appetite into two points that can be tested against the curve. Gross of insurance, the portfolio exceeds $10,000,000 in 16.0 percent of simulated years and $40,000,000 in 1.2 percent.

Worked in full in The Cyber Risk Quantification Handbook by Julian R. Sterling, with every figure reproduced in live Excel workbooks.Get the DRM-free PDF, $19.99 →

From trials to a curve

The input is a column of simulated annual losses. In the case, each scenario runs 1,000 trials, one trial being one simulated year, drawn from stored uniform random numbers (seed 2027) so the results can be reproduced to the dollar. How each trial is built, from threat event frequency to loss magnitude, is set out in the guide on how to quantify cyber risk with FAIR. The curve itself takes three steps:

  1. Add the five scenario losses trial by trial to get 1,000 portfolio years.
  2. Choose thresholds: $1,000,000, $5,000,000, $10,000,000 and so on up to a level no trial reaches.
  3. For each threshold, count the trials above it and divide by 1,000.

Plot probability against threshold, usually with a log scale on the loss axis, and you have the curve. Nothing in it is averaged away: every trial contributes a point.

Percentiles are return periods

A percentile is the same information read the other way. The 90th percentile is the loss exceeded in 10 percent of years, or about one year in 10; the 95th is the one-in-20 year; the 99th the one-in-100 year. Percentiles here use linear interpolation between neighboring trials (the PERCENTILE.INC convention). For the Halden Marsh portfolio, gross of insurance:

The mean, $5,016,300, sits far above the median. That gap is the signature of a skewed risk, and it is why an ALE alone misleads.

Why S1's 90th percentile is $0

The ransomware scenario, S1, has the largest expected loss of the five, yet its 90th percentile is $0. This surprises every first-time reader, and it is correct. S1 produced no loss event in 907 of 1,000 simulated years, one event in 90 years and two events in 3 years. Since 90.7 percent of years are zero, the value below which 90 percent of years fall is zero as well.

The tail starts just beyond. The 95th percentile is $14,546,698, the 99th $33,292,684 and the worst simulated year $77,804,295. A rare, severe scenario is invisible at the one-in-10 level and dominant at the one-in-20 level. The free Cyber Risk Monte Carlo Model, prefilled with S1, reports the same split: a loss in 9.3 percent of years, a 6.6 percent chance of exceeding $10,000,000 and a 1.6 percent chance of exceeding $25,000,000.

IdYears with no loss90th percentile99th percentile
S190.7 percent$0$33,292,684
S250.3 percent$1,794,669$4,301,692
S389.6 percent$1,056,417$12,113,483
S479.3 percent$2,702,221$11,738,271
S581.5 percent$3,023,021$18,052,478

Payment fraud, S2, is the opposite profile: a loss in about half of all years, but a 99th percentile of only $4,301,692. Ranking scenarios on one percentile, or on the mean alone, would hide that difference. Report at least the ALE and two tail points per scenario.

The portfolio curve, gross and net

The same count can be run on any version of the annual loss: gross, net of the insurance program, or gross after proposed controls. Because every version reuses the same stored random numbers, the differences come only from the change in inputs.

ThresholdGrossNet of current insuranceGross with all three controls
$1,000,00057.4 percent47.6 percent43.3 percent
$5,000,00028.7 percent13.6 percent21.5 percent
$10,000,00016.0 percent6.9 percent9.0 percent
$25,000,0003.1 percent1.4 percent1.1 percent
$40,000,0001.2 percent0.4 percent0.2 percent
$60,000,0000.4 percent0.1 percent0.1 percent
$100,000,0000.0 percent0.0 percent0.0 percent

Read the columns as two different tools. Insurance lowers the curve most in the middle, between $5,000,000 and $10,000,000, where losses sit above the retention and inside the limit. The three controls, which together cost $2,180,000 a year, do less in the middle but more in the far tail: at $25,000,000 they bring the probability to 1.1 percent against 1.4 percent for insurance. Insurance pays for losses; controls make them smaller or rarer.

Percentiles do not add

A frequent error is to build a portfolio tail by adding scenario tails. The five scenario 95th percentiles sum to $36,295,514; the portfolio 95th percentile is $20,906,669. Bad years in different scenarios rarely coincide, so the one-in-20 year of the portfolio is far smaller than the sum of five one-in-20 years. ALEs, by contrast, do add: the five simulated ALEs sum exactly to the portfolio ALE.

The case simulates scenarios independently. If a common cause linked them, for example one intrusion that both encrypts plants and steals personal data, bad years would cluster and the portfolio tail would grow toward the sum. Where such a link is plausible, model it as its own scenario rather than trusting the diversification.

How far to trust the tail

A reference run of 100,000 trials gave an ALE of $4,846,290, a 95th percentile of $21,295,978 and a 99th percentile of $40,335,385. The 1,000-trial figures differ by 3.5 percent on the mean, 1.8 percent on the 95th and 12.1 percent on the 99th. The one-in-100 point rests on about ten simulated years, so treat it as indicative and say so in the report.

Presenting the curve without misleading

A curve is easy to draw and easy to misread. Four habits keep it honest in front of a committee:

The last habit is what makes a curve useful for decisions: when a control or a policy term moves it, the movement is real.

Risk appetite as points on the curve

A statement such as "low appetite for cyber risk" cannot be tested. Points on the curve can. Halden Marsh's board set two tolerances for 2027: the probability of an annual cyber loss above $15 million no more than 10 percent, and above $40 million no more than 1 percent. Each is a threshold and a maximum probability, so the test is a lookup:

Tolerance pointGrossNet of current insuranceGross with all three controls
Above $15 million (limit 10 percent)9.6 percent4.1 percent5.1 percent
Above $40 million (limit 1 percent)1.2 percent0.4 percent0.2 percent

Gross of insurance, the one-in-100 tolerance is breached. The current insurance program brings both points inside tolerance, and so do the three controls on their own. That result frames the next decision: whether to rely on transfer, on controls or on both, which the guide on how to choose a cyber insurance retention and limit works through. The appetite wording, the board report and the full curve are in The Cyber Risk Quantification Handbook.

Questions practitioners ask

What does a loss exceedance curve show?

For each loss amount, the probability that the annual loss exceeds it. At Halden Marsh, gross of insurance, the portfolio exceeds $1,000,000 in 57.4 percent of simulated years, $10,000,000 in 16.0 percent and $40,000,000 in 1.2 percent.

How many trials does the curve need?

Enough for the point you report. With 1,000 trials the case's mean and 95th percentile sit within a few percent of a 100,000-trial run, but the 99th percentile differs by 12.1 percent because it rests on about ten simulated years.

Why can a scenario's 90th percentile be zero?

Because it has no loss in more than 90 percent of years. S1, ransomware, has a loss in only 93 of 1,000 simulated years, so its 90th percentile is $0 while its 95th percentile is $14,546,698.

How do I set a risk appetite on the curve?

State a threshold and a maximum probability, then read the curve. Halden Marsh's board allows at most 10 percent above $15 million and 1 percent above $40 million; gross of insurance the second point reads 1.2 percent and fails, net it reads 0.4 percent and passes.

Read the full handbook

This guide is one arc of The Cyber Risk Quantification Handbook. The book follows one company from its heat map to five quantified scenarios: calibrated estimates, the FAIR model, Monte Carlo simulation, the portfolio loss exceedance curve against board tolerance, the return on controls, the cyber insurance renewal and an incident materiality decision.

Paperback PDF edition, $19.99 Book page and companion workbooks

Free download: Cyber Risk Monte Carlo Model (Excel).

Also on this site

Illustrative figures computed by the book model; the company, its scenarios and every estimate are fictional. Rules described in principle as read in November 2026. Education only, not insurance, legal or security advice.