Why Likelihood Categories Need a Defined Basis
A likelihood label such as “possible” is not meaningful on its own. One assessor may mean a chance of at least one event in the next year; another may be thinking about the next five years. The event itself may also be unclear: “system disruption” could mean any interruption, or only one lasting more than four hours. If the event, time horizon, or exposure basis changes, the assigned category may change even when the underlying risk has not.
A useful likelihood scale therefore begins with an operational event statement and a fixed time horizon. It then assigns that event to one of a set of clearly bounded probability categories. The categories are a communication device: they summarize an assessment, but do not replace the assumptions and evidence behind it. The scale below is illustrative. An organization should choose bands that suit its decisions and use them consistently.
Specify the Event, Horizon, and Exposure
Before assigning a label, write down what counts as an occurrence. Include a threshold when the event could otherwise be interpreted in several ways. For example, “a vendor outage that prevents claims processing for more than four consecutive hours” is more assessable than “vendor problems.” State whose or what exposure is being assessed, such as one processing platform or one insured location. A probability per platform-year is not automatically comparable with a probability per policy-year.
For the scale below, the horizon is one year and the event is whether at least one qualifying occurrence takes place during that year. The labels and thresholds are hypothetical design choices, not universal standards. “Expected occurrences per 100 exposure-years” is included as a frequency-style interpretation: if each exposure-year has the same probability \(p\), the expected number of occurrence-years in 100 such years is \(100p\). It does not mean events will occur at regular intervals.
| Category | Annual probability of at least one occurrence | Frequency-style wording per 100 exposure-years |
|---|---|---|
| Rare | \(0 \leq p < 0.01\) | Fewer than 1 occurrence-year expected |
| Unlikely | \(0.01 \leq p < 0.05\) | At least 1 but fewer than 5 occurrence-years expected |
| Possible | \(0.05 \leq p < 0.20\) | At least 5 but fewer than 20 occurrence-years expected |
| Likely | \(0.20 \leq p < 0.50\) | At least 20 but fewer than 50 occurrence-years expected |
| Almost certain | \(0.50 \leq p \leq 1\) | At least 50 occurrence-years expected |
The half-open intervals make the scale mutually exclusive: a probability of exactly 5% is Possible, not Unlikely; exactly 20% is Likely, not Possible. The bands are also exhaustive for probabilities from zero to one. These properties prevent assessors from assigning the same value to two categories or leaving an otherwise valid probability unclassified.
From Event Data or Frequency to a Category
If the assessment starts with a probability estimate, compare it directly with the stated bands, using the specified horizon. If it starts with a frequency rate, convert to a probability only under an appropriate model. One common assumption for a count of events over time is a Poisson process with constant rate \(\lambda\) per exposure-year and independent increments. Under that model, the count \(N(t)\) in \(t\) years has a Poisson distribution with mean \(\lambda t\).
The probability of no events during one year is \(P(N(1)=0)=e^{-\lambda}\). Therefore, the probability of at least one event is the complement:
The complement works because “no occurrences” and “one or more occurrences” are mutually exclusive and cover all possible counts. This conversion is not automatic for every risk. If the rate changes over time, events are dependent, or the event definition changes, the Poisson assumptions may not be suitable. Use a model consistent with the available evidence and disclose its limitations.
Worked Example: Converting an Event Rate to an Annual Category
Hypothetical scenario. A risk assessor estimates that a single claims platform experiences a qualifying data-transfer failure at a constant Poisson rate of \(0.08\) events per platform-year. A qualifying event is a failure that delays a scheduled transfer by at least six hours. Assign an annual likelihood category for at least one qualifying event.
Method and assumptions. Use the Poisson model with \(\lambda=0.08\) per platform-year. The event definition and one-platform, one-year exposure basis match the scale.
First calculate the probability of zero qualifying events in one year:
Then take the complement:
Rounded to four decimal places, the annual probability is \(0.0769\), or \(7.69\%\). Since \(0.05 \leq 0.0769 < 0.20\), the assigned category is Possible. The frequency-style interpretation is \(100(0.0768837)=7.68837\) occurrence-years expected per 100 comparable platform-years; this is an expectation, not a schedule of failures.
Check. The probability is between zero and one, and \(7.69\%\) falls within the Possible band. A common mistake is to call the rate \(0.08\) itself an 8% probability. The correct probability is \(1-e^{-0.08}=7.69\%\), because a rate counts events and can exceed one, whereas an occurrence probability cannot.
Assigning Categories Consistently
Consistency requires more than using the same labels. Assessors should use the same definition of the event, horizon, exposure unit, and probability interpretation. Where several risks are being assessed, record these items alongside the category. Otherwise, apparent differences in likelihood may simply reflect different questions. The category should be assigned from the estimate using the written boundaries, rather than selected first and justified afterward.
The following steps provide a practical sequence. They also make the assessment easier to review if estimates change or another assessor needs to reproduce the classification.
Specify the occurrence threshold and the unit exposed. Avoid combining distinct events into a broad phrase if they could have different probabilities.
State whether the question concerns a month, year, or another period. Use the same horizon as the category scale, or convert only under stated assumptions.
Clarify whether the estimate is the probability of at least one event, an expected count, or another frequency measure. Do not compare unlike quantities.
Use the available evidence and a suitable method, then apply the category boundaries literally, including their endpoints.
Keep the event definition, time horizon, exposure unit, estimate, assumptions, and category together so the label remains interpretable.
Worked Example: Applying a Boundary Correctly
Hypothetical scenario. A risk owner assesses whether a newly configured payment interface will produce at least one duplicate settlement during the next year for one business unit. The documented estimate is exactly \(20\%\). Assign the category using the stated scale.
Method. Confirm that the estimate concerns at least one occurrence, that the horizon is one year, and that the exposure unit is one business unit. These match the scale. No conversion is needed.
The relevant boundaries are Possible for \(0.05 \leq p < 0.20\) and Likely for \(0.20 \leq p < 0.50\). The estimate is \(p=0.20\), so it meets the lower boundary of Likely:
The correct category is Likely. As a frequency-style interpretation, the estimate corresponds to \(100(0.20)=20\) occurrence-years expected per 100 comparable business-unit-years.
Check. Exactly 20% is included in Likely and excluded from Possible by the written inequalities. The common mistake is to treat both adjacent ranges as inclusive at 20%, creating an overlap. Explicit endpoint rules remove that ambiguity.
Changing the Horizon
A multi-year probability cannot be placed directly on a one-year scale. Conversion requires assumptions about how occurrence behaves over time. For example, if each year has the same probability \(p\) of an occurrence and yearly outcomes are independent, then the probability of no occurrence in \(n\) years is \((1-p)^n\). The probability of at least one occurrence is its complement:
This follows because no event over the full horizon requires no event in every year; independence makes the probability of that joint outcome the product of the yearly no-event probabilities. If the annual probabilities vary, or years are dependent, this formula does not apply as written. A one-year category should not be assigned by comparing a multi-year probability directly with annual bands.
Worked Example: Annualizing a Five-Year Probability
Hypothetical scenario. For one insured facility, an engineering assessment estimates a \(30\%\) chance of at least one qualifying refrigeration-control failure over the next five years. The organization uses a one-year likelihood scale. For this conversion only, assume the annual probability is constant and occurrences in different years are independent. Find the annual probability and category.
Method. Let \(p\) be the annual probability of at least one qualifying failure. The five-year probability is \(0.30\), so the five-year probability of no failure is \(1-0.30=0.70\). Under the stated independence and constant-probability assumptions:
Take fifth roots and solve:
The annual probability is \(0.0689\), or \(6.89\%\), rounded to four decimal places. Since \(0.05 \leq 0.0689 < 0.20\), the annual category is Possible.
Check. Substituting the unrounded annual probability gives \(1-(1-0.0688501)^5 \approx 0.3000\), recovering the stated five-year estimate. The common mistake is to divide \(30\%\) by five and report \(6\%\). That shortcut does not solve the independent-year relationship and is not the exact conversion.
Frequency Language and Common Traps
Frequency wording can make a category easier to interpret, but it must retain a clear denominator. “One event every 20 years” may be understood as a return period, a historical average, or a probability claim. It is not automatically a statement that the annual probability is \(5\%\). A Poisson rate of one event per 20 years would imply an annual probability of at least one event of \(1-e^{-1/20}\), approximately \(4.88\%\), under the Poisson assumptions. The model and interpretation matter.
For recurring events, report both the expected count and the probability of at least one event when each is relevant. Under a Poisson model with rate \(\lambda\), the expected number of events in one year is \(\lambda\), while the probability of at least one is \(1-e^{-\lambda}\). These are different quantities. A probability-based category scale applies to the latter, not to the raw expected count.
The most common errors are practical rather than computational:
- Leaving the horizon implicit. “Likely” without “during the next year” cannot be compared reliably with an annual scale.
- Using an undefined event. A broad label can mix minor incidents with threshold events that matter for the assessment.
- Comparing different exposures. A probability per location-year and a probability per policy-year need not be comparable.
- Confusing rate and probability. A rate may exceed one event per year; a probability cannot exceed one.
- Ignoring boundaries. A value at a band endpoint belongs wherever the stated inequalities assign it.
- Converting horizons without assumptions. Independence and a constant annual probability must be justified before using the annualization formula.
A well-designed likelihood scale makes disagreement more informative. If two assessors choose different categories, the event definition, evidence, time horizon, or assumptions can be examined directly. The label alone is not the assessment; it is a concise summary of a clearly specified judgment. This approach is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.
Check Your Understanding
For each problem, use the category boundaries and endpoint conventions stated in this tutorial unless the problem specifies otherwise.
- A hypothetical risk has an estimated annual probability of \(0.009\) of at least one qualifying occurrence for one facility. What category applies? (Answer: Rare.)
- A hypothetical system has a Poisson event rate of \(0.30\) qualifying outages per system-year. Find the annual probability of at least one outage and its category. (Answer: \(1-e^{-0.30}=0.2592\), Likely.)
- A hypothetical risk estimate is a \(12\%\) chance of at least one event in the next year for one operating unit. State the category and the expected occurrence-years per 100 comparable unit-years. (Answer: Possible; 12 occurrence-years expected per 100 unit-years.)
- A hypothetical facility has a \(40\%\) probability of at least one event over four years. Assuming independent years with a constant annual probability, find the annual probability and category. (Answer: \(1-0.60^{1/4}=0.1199\), Possible.)
- A hypothetical count model has a Poisson rate of \(1.2\) events per year for one processing center. State the expected event count, the probability of at least one event, and the category under the probability scale. (Answer: Expected count \(=1.2\) events per year; probability \(=1-e^{-1.2}=0.6988\); Almost certain.)