Tools › Project Management › Risk › Cyber Risk Quantification


Cyber Risk Quantification

Enter your own cyber loss scenarios (events per year and loss per event, each as a 90 % range) and a control you want to test; get the expected loss per year, the loss with a 10 % and a 1 % chance of being exceeded, and the loss exceedance curve before and after the control, with the control's net benefit and return on its cost. No industry breach-cost or frequency benchmark is built in: every default is an original illustrative number, and the output is only as good as the ranges you enter.

  • Your own scenarios instead of industry averages: a seeded Monte Carlo simulation with Poisson event counts and lognormal losses fitted to your 90 % ranges, checked against the closed-form expected loss

  • Loss exceedance curve on a logarithmic loss axis before and after a control, with the 10 % and 1 % points marked and an insurance retention and limit drawn in

  • Expected loss per year, reduction, net benefit and return on the control's cost as formula, numbers and answer; the expected loss retained, insured and above the limit

  • Seeded simulation that a reviewer can repeat; each scenario's share of the expected loss; the chance of exceeding any loss amounts you name

Cyber Risk Quantification

/

Scenarios: Events per Year and Loss per Event (90 % ranges)

1
2
3

Control to Test (cuts of 0 for none)




Insurance on the Year's Total (blank for none)



Simulation





×

Cyber Risk Quantification

Cyber Risk Quantification: From Ranges to a Loss Exceedance Curve

The tool takes your own cyber loss scenarios, each as a range of events per year and a range of loss per event, and turns them into the expected loss per year and the chance of a year worse than any amount you name. It then repeats the calculation with a control in place, so the control's cost can be set against the loss it removes. Every default is an original illustrative number: no industry breach-cost or frequency benchmark is built in, and the output is only as good as the ranges you enter.

The idea: frequency times severity

NIST Special Publication 800-30 describes risk as a function of the likelihood that a threat event occurs and the adverse impact if it does. NIST Interagency Report 8286A adds the two money measures used here: the loss of one occurrence and the total loss over a year (it calls them single loss expectancy and annualized loss expectancy), and it describes estimating with ranges and repeating the calculation many times with random draws, the Monte Carlo method. Those reports give no formulas for the distributions; the model below is the standard frequency and severity model of loss mathematics, stated in full so that you can check it.

You rarely know how often an event will happen or what it will cost. You can usually give a range you are 90 % sure of: "between once in twenty years and once in two or three years", "between 40,000 and 600,000". The tool reads each range as the 5th and 95th percentiles of a lognormal distribution, a bell curve on a logarithmic scale that cannot go below zero and has a long upper tail, which is how loss sizes tend to behave.

A lognormal distribution fitted to a loss per event between 40,000 and 600,000: 90 percent of events between the two ends, 5 percent below, 5 percent above, median 154,919, mean 217,397
Figure 1: A loss per event entered as 40,000 to 600,000. The median is the geometric middle of the range, 154,919; the mean, 217,397, is higher because of the tail above the high end.

The equations

A lognormal from a 90 % range. For a range from lo to hi, with z = 1.645 (the 95th percentile of the standard normal distribution):

μ = (ln lo + ln hi) / 2,   σ = (ln hi − ln lo) / (2 z),   median = eμ = √(lo × hi),   mean = eμ + σ²/2

Events per year. The number of events of a scenario in a year is a Poisson count. With the default option the rate of that count is itself uncertain: each simulated year draws its rate from the lognormal of the frequency range, so the range expresses doubt about the true long-run rate. With the fixed-rate option every year uses the geometric mean √(lo × hi). The uncertain rate has the higher mean, eσ²/2 times the geometric mean, and the fatter tail; the tool opens on it.

Loss in a year. Each event draws its own loss from the lognormal of the loss range, and the year's loss S is the sum over all events of all scenarios. For one scenario with event rate Λ and loss per event X:

E[S] = E[Λ] × E[X],   Var[S] = E[Λ] × E[X²] + Var[Λ] × E[X]²,   with E[X²] = e2μ + 2σ²

The scenarios are taken as independent, so their means and variances add. The expected loss per year printed by the tool is this closed form; the average of the simulated years is printed beside it with its standard error, as a check on the simulation.

The simulation. The percentiles of S have no closed form, so the tool simulates 10,000 years (Advanced: 5,000 to 100,000) from a seed that it prints; the same seed and inputs give the same figures. Sorting the simulated years gives the loss exceedance curve: for each amount, the share of years in which the loss was larger. The loss with a 10 % chance of being exceeded is the 90th percentile of the years; the loss with a 1 % chance is the 99th.

The control. A control is described by its cost per year, the share of events it prevents and the share by which it shrinks each remaining loss, in the scenarios it covers. Their expected loss is multiplied by (1 − cut in events) × (1 − cut in loss size). In the simulation each covered event of the "before" year is removed with the entered probability (a thinned Poisson count is again a Poisson count, at the reduced rate) and each remaining loss is scaled down, so both curves come from the same random numbers and differ only by the control.

reduction = expected loss before − expected loss after,   net benefit = reduction − cost,   return = (reduction − cost) / cost

Insurance. A retention R and a limit M are applied to the year's total: you keep min(S, R), the insurance pays min(max(S − R, 0), M), and anything above R + M comes back to you. The tool reports the average of each part over the simulated years and the share of years in which the limit is used up.

What it assumes

  • Each entered range is a true 90 % range: the real value falls below the low end 1 time in 20 and above the high end 1 time in 20.
  • Loss sizes and uncertain event rates follow lognormal distributions; event counts are Poisson, so events of one scenario do not trigger each other.
  • Scenarios are independent of one another, and one year is independent of the next.
  • The control's cuts are certain and act evenly on every covered event; the insurance terms act on the year's total, with no exclusions, sub-limits or per-event terms.
  • Money is in one currency, in today's value, for one year.

When it misleads

  • Overconfident ranges. People asked for a 90 % range tend to give one that is too narrow. A range that is too narrow understates both the expected loss and the tail, and nothing in the arithmetic can show it.
  • Very wide loss ranges. When the high end is hundreds of times the low end, the mean of the lognormal sits far above the median and rests on events beyond the high end. The expected loss then says more about the assumed tail shape than about your estimate; the tool prints a note when a range is that wide.
  • Dependent scenarios. One intrusion can cause an outage, a data breach and a fraud at once. Treating them as independent understates the bad years. If scenarios tend to arrive together, model them as one scenario with a larger loss.
  • Tail figures from few years. The 1 % loss of 10,000 simulated years rests on the worst 100 of them. Change the seed to see how much it moves; raise the number of trials when it matters.
  • A control judged on the average alone. A control can cost more than the expected loss it removes and still cut the 1 % loss by a large amount. The tool shows both figures; which one matters is a question of how much loss the organization can absorb, which the tool does not answer.
  • Losses that are not money. Harm to people, to safety or to trust that cannot be priced is outside the model.

Common mistakes

  • Entering the smallest and largest loss ever imagined instead of a 90 % range. The ends are the 5th and 95th percentiles, not the minimum and maximum; the distribution continues beyond both.
  • Multiplying the middles. Middle frequency × middle loss is the product of two medians. The expected loss uses the means, which are higher: for a loss between 40,000 and 600,000 the median is 154,919 and the mean 217,397.
  • Entering a probability as a frequency without thought. "A 10 % chance per year" is close to 0.1 events per year, but "once every three years" is 0.33, and a frequency above 1 is not a probability at all.
  • Counting the same loss in two scenarios (the outage caused by the ransomware entered under both).
  • Reading the expected loss as a budget. In the opening example the loss is above its expected value in fewer than three years out of ten and far above it in a few; the exceedance curve, not the average, shows what a bad year costs.
  • Leaving the control's own cost out, or counting only the purchase price of a control that also takes staff time every year.
  • Applying a control to scenarios it does not touch. Mark only the scenarios the control acts on.

Checks

The lognormal parameters, means and second moments and the closed-form expected loss and standard deviation are checked against scipy; the simulated means, the share of years with a loss, the exceedance percentiles, the insurance layers and each scenario's share of the worst years are checked against an independent numpy simulation of 2,000,000 years per case, inside stated sampling tolerances. Open FAIR, the risk taxonomy of The Open Group, uses the same decomposition into loss event frequency and loss magnitude and is cited for the reader; the tool does not implement that standard.




References

Joint Task Force Transformation Initiative (2012). Guide for Conducting Risk Assessments. NIST Special Publication 800-30, Revision 1, National Institute of Standards and Technology, Gaithersburg, MD.

Stine, K., Quinn, S., Witte, G., and Gardner, R. K. (2020). Integrating Cybersecurity and Enterprise Risk Management (ERM). NIST Interagency Report 8286, National Institute of Standards and Technology, Gaithersburg, MD.

Quinn, S., Ivy, N., Barrett, M., Feldman, L., Witte, G., and Gardner, R. K. (2021). Identifying and Estimating Cybersecurity Risk for Enterprise Risk Management. NIST Interagency Report 8286A, National Institute of Standards and Technology, Gaithersburg, MD.

National Institute of Standards and Technology (2012). NIST/SEMATECH e-Handbook of Statistical Methods. Sections 1.3.6.6.9, Lognormal Distribution, and 1.3.6.6.19, Poisson Distribution, National Institute of Standards and Technology, Gaithersburg, MD.

Klugman, S. A., Panjer, H. H., and Willmot, G. E. (2019). Loss Models: From Data to Decisions, 5th ed. John Wiley and Sons, Hoboken, NJ.

The Open Group (2013). Risk Taxonomy (O-RT), Version 2.0. Open Group Standard C13K, The Open Group, Reading, UK.


DISCLAIMER: The methods and calculation on aNOWysis.com are provided as means to educate the reader. There is no warranty, expessed or implied, as to the relevance or accuracy of the methods or calculations. All calculations should be checked by a qualified analyst for appropriateness and accuracy. We will not be held liable for any damages, injuries or deaths caused by using methods or calculations from this site.