Bayesian estimation of incident rates for risk registers
Risk analysts need incident frequencies but rarely have enough data. Starting from an expert's range, each year's incidents update the estimate, and the uncertainty is carried through to expected loss. Six illustrative asset classes show the effect on prioritisation.
"How many ransomware incidents should we expect on our endpoints next year?" A sound answer combines two limited sources of knowledge. Experts have an informed view, from industry reports and experience, but no precise number. The organisation's own history has precise numbers, but only a handful of them: two years of data, three incidents.
Classical statistics forces a choice between the two. Bayesian statistics combines them, weighting each source by its information content. For rates of events, the arithmetic is as simple as for proportions in sequential Bayesian updating of a failure rate: two additions.
The Poisson model for counts of rare events goes back to Poisson (1837) and to Bortkiewicz's 1898 study of deaths by horse kick in the Prussian army. The Gamma distribution is its conjugate prior, a pairing described in every Bayesian textbook, for example Gelman and colleagues (2013). In cybersecurity, Hubbard and Seiersen (2023) argue for this approach: calibrated expert ranges as priors, updated with an organisation's own incident data, feeding Monte Carlo loss models such as the loss exceedance curves of risk matrices and loss exceedance curves. Open FAIR's "loss event frequency" is the quantity being estimated.
From an expert's range to a prior
The expert provides a 90% range, for example: "I'm 90% sure our web applications see between 1 and 6 incidents a year." This range is converted into a Gamma(, ) distribution for the rate whose 5th and 95th percentiles match those numbers. The ratio of the two percentiles fixes the shape ; one of them then fixes the rate . A wide ratio means a small : a vague prior that the data will quickly override.
The prior can be read as experience: Gamma(, ) is worth about incidents observed over years.
Updating with data
Observe incidents over years. With a Poisson model the likelihood is proportional to , and the posterior is
The incidents are added to and the years to . The posterior mean is a weighted average of the expert's mean and the observed rate , with weights and : the expert's equivalent years of experience and the observed years.
Motor insurers have priced policies on this principle for decades. A new driver is charged roughly the average for drivers with the same profile, which plays the role of the prior. Each claim-free year then lowers the premium, and after many years the driver's own record dominates. Actuaries call this credibility weighting, and for claim counts its standard form is the same Gamma-Poisson model.
Outputs for the risk owner
Two outputs are relevant for a risk register.
- A frequency with its uncertainty: the posterior mean and a 90% interval, rather than a single "likely" label.
- Predictions: integrating over the uncertainty in , the number of incidents next year follows a negative binomial distribution, and the probability of a quiet year is with the posterior parameters , .
Multiplying the rate by the average loss per incident gives the expected annual loss per asset class, which informs where mitigation funding is allocated, provided its uncertainty is reported with it.
The figure below holds six illustrative asset classes. The numbers are illustrative and are not real data.
Selecting an asset class draws its prior and posterior at the top. The bubble chart places all six by incident rate and loss per incident, with bubble area proportional to expected annual loss, and the treemap shows each class's share of the total.
Suggested settings:
- Web applications. The expert's range of 1 to 6 per year becomes a Gamma prior, and 9 incidents in 2 years pull the posterior up to about 3.9 per year, with a 90% interval of about 2.3 to 5.9. The chance of a quiet year is about 3%.
- No incidents observed. Choose "OT network": no incidents in 2 years. The posterior moves down from the expert's guess but keeps a long right tail; two quiet years do not prove a rate of zero. With 10 years the tail shrinks further.
- Vague prior. For "Cloud accounts", widen the expert's range to 0.05-20. The prior becomes nearly flat and the posterior is driven almost entirely by the data. Narrow it to 1-1.5 and the same data barely move the belief.
- Reprioritisation. Raise the observed incidents for "Suppliers" from 3 to 15. Its bubble moves right and grows, and its share of the treemap increases. The investment ranking changes on the basis of evidence.
Derivation of the update. The Gamma(, ) prior density is proportional to , and the Poisson likelihood of events in time is proportional to . Their product is , the kernel of Gamma(, ). For the predictive probability of no event in the next year, average over the posterior:
using the Gamma integral .
This is the quantitative core of a risk register that is updated with evidence. Each year's incident log updates every frequency. Experts carry the weight where data are thin, the data take over where they are plentiful, and the uncertainty is carried forward rather than reduced to a label. It also makes the risk owner's decision easier to audit, since every number can be traced to a stated range and a counted history, in the spirit of the role of information security in risk decisions.
References
- S. D. Poisson, Recherches sur la probabilité des jugements en matière criminelle et en matière civile, Bachelier, 1837.
- L. von Bortkiewicz, Das Gesetz der kleinen Zahlen, Teubner, 1898.
- A. Gelman, J. B. Carlin, H. S. Stern, D. B. Dunson, A. Vehtari and D. B. Rubin, Bayesian Data Analysis, 3rd ed., CRC Press, 2013.
- D. W. Hubbard and R. Seiersen, How to Measure Anything in Cybersecurity Risk, 2nd ed., Wiley, 2023.
- The Open Group, Risk Taxonomy (O-RT) and Risk Analysis (O-RA) standards (Open FAIR).