Semi-Quantitative Risk Scoring, Decoded

The scoring sheet inside almost every risk register: rate likelihood and impact from 1 to 5, multiply, band the result. It is quick and it ranks risks better than words do — as long as you know what the numbers are not.

1–25likelihood (1–5) × impact (1–5)
14different scores are possible — not 25
12of 25 cells land in a typical “Low” band (1–7)
33×the expected loss a lower-scoring risk can carry (worked below)

NIST SP 800-30 Rev. 1 defines a semi-quantitative assessment as one that uses “bins, scales, or representative numbers whose values and meanings are not maintained in other contexts.” That last clause is the whole story. A risk scored 20 is worse than one scored 10 on your register, but it is not twice as bad, and a 20 on someone else’s register may mean something else entirely. Used with that in mind, the method is the best tool there is for ranking a long list of risks quickly. Used as arithmetic, it misleads.

It sits between the two other approaches NIST describes: qualitative (words combined by table) and quantitative (numbers that keep their meaning, like dollars per year — see Inherent vs Residual Risk).

1 · Bins, Scales and Representative Values

NIST gives three forms a semi-quantitative value can take, and its own tables use two of them side by side with the qualitative words:

LevelBin (0–100)Representative value (0–10)
Very High96–10010
High80–958
Moderate21–795
Low5–202
Very Low0–40

The same bins and values appear in NIST SP 800-30 Rev. 1 Tables G-2 to G-4 (likelihood), H-3 (impact) and I-3 (risk). Note the bins are deliberately uneven: “Moderate” spans 21–79.

NIST’s explanation of why this helps is precise: bins “translate easily into qualitative terms” (a 95 reads as very high), and they allow comparison within a level — the gap between 70 and 71 is insignificant, the gap between 36 and 70 is significant. It is just as precise about the cost: “the role of expert judgment in assigning values is more evident than in a purely quantitative approach”, and each bin “needs to be clearly defined and/or characterized by meaningful examples.”

2 · The 5×5: What Multiplying Actually Gives You

Most registers use the simplest form: a 1–5 scale for each factor and a product.

Risk score = Likelihood (1–5) × Impact (1–5)
Likelihood ↓  /  Impact →12345

Bands used here: 1–7 Low, 8–14 Medium, 15–25 High — a common illustrative convention, not a standard.

Three properties of that grid are worth knowing before you trust it:

  • Only 14 scores exist. The products of 1–5 and 1–5 are 1, 2, 3, 4, 5, 6, 8, 9, 10, 12, 15, 16, 20 and 25. Nothing ever scores 7, 11, 13, 14, 17–19 or 21–24, so a band edge at “above 13” is the same as one at “15 or more”.
  • The bands are lopsided. With 1–7 / 8–14 / 15–25, twelve cells are Low, seven Medium and six High. Where you draw the band edges decides more of the outcome than any individual rating.
  • Different risks tie. 5 × 1 and 1 × 5 both score 5: a near-certain nuisance and a rare catastrophe land in the same place and get the same treatment.

3 · Worked Example: Ransomware via Remote Access

The series’ standard risk, on a typical word-anchored 1–5 scale:

StateLikelihoodImpactScoreBand
Before MFA4 — likely this year5 — severe20High
MFA on the VPN2 — unlikely5 — severe10Medium
Plus tested offline backups2 — unlikely3 — moderate6Low

This is what semi-quantitative scoring does well: each control is visible as a move on one axis, and the register can be sorted. What it cannot say is whether going from 20 to 6 was worth what MFA and backups cost — for that, the same risk is worked in money in Inherent vs Residual Risk.

Try it: 5×5 risk score

Score20
Band
Same score from

The selected cell is outlined in the score map above.

4 · Where It Breaks: Cox’s Critique

The best-known analysis of the problem is Louis Anthony (Tony) Cox Jr., “What’s Wrong with Risk Matrices?”, Risk Analysis 28(2), 2008. His findings, in brief: a matrix can give the same rating to risks that differ enormously in size (range compression), can give a higher rating to a quantitatively smaller risk (risk inversion), and in some situations performs worse than choosing at random. Here is inversion with numbers. Define the scales with real ranges:

ScoreLikelihood (per year)Impact (per event)
5over 90%over $10M
450–90%$1M–$10M
310–50%$100k–$1M
21–10%$10k–$100k
1under 1%under $10k
RiskReal figuresScoreExpected annual loss
A — data-centre loss8% a year, $50M if it happens2 × 5 = 10$4,000,000
B — laptop theft wave60% a year, $200k if it happens4 × 3 = 12$120,000

▸ B outranks A on the register, yet A’s expected loss is about 33 times larger. Nothing was rated wrongly — the scales put every likelihood from 1% to 10% in one step and every loss above $10M in another, and multiplication of ordinal labels cannot recover what was thrown away.

Ordinal means ordered: 4 is above 2, but not “twice” 2. Multiplying ordinal labels is a convenience, not a measurement — which is exactly what NIST’s “not maintained in other contexts” warns.

5 · Making a 5×5 Honest

  1. Give every level a numeric range — a frequency per year and a loss per event, like the table above. Words alone (“likely”, “major”) drift between assessors; see Qualitative Risk Matrix.
  2. Space the ranges by a constant factor, ideally ×10 on both axes (frequency: under once a century, once per 10–100 years, once per 1–10 years, 1–10 a year, over 10 a year; loss: under $10k up to over $10M). Then each point of likelihood plus impact is roughly ten times the expected loss, and the sum ranks risks in the right order of magnitude.
  3. Keep both numbers visible next to the score, so a 5 × 1 and a 1 × 5 are never read as the same risk.
  4. Add a severity override — any impact rated 5 is reviewed by an executive whatever its score, the same idea as FMEA’s severity-first rule (see FMEA Risk Priority Number).
  5. Graduate the top of the list — for the five or ten risks where money for controls is real, estimate loss in dollars instead of scoring it.

Bottom line

  • Scores rank; they do not measure — NIST: values and meanings are “not maintained in other contexts”.
  • A 5×5 has 14 possible scores and its band edges decide more than its ratings.
  • Ties and inversions are built in — Cox (2008); keep likelihood and impact visible.
  • Anchor levels to numbers spaced by ×10, and add L + I rather than trusting L × I.
  • Use scores to sort the register, money to decide the spend.

Sources & method. Definition, bins, representative values and the quoted guidance from NIST SP 800-30 Rev. 1, Guide for Conducting Risk Assessments (September 2012), §2.3.2 and Tables G-2 to G-4, H-3 and I-3. L.A. Cox Jr., “What’s Wrong with Risk Matrices?”, Risk Analysis 28(2):497–512 (2008), summarised, not quoted. Score counts are arithmetic on the 5×5 grid; band edges, scale ranges and both example risks are illustrative. Original analysis; no third-party graphics reproduced.

← Back to Cyber security