Home  /  Tutorials  /  CAS Exam 9 / Risk Concepts and Taxonomies

CAS Exam 9 · Risk Concepts and Taxonomies · Tutorial 2 of 995

Risk, Uncertainty, and Ambiguity

Learn to identify when a risk can be quantified, when its probabilities are uncertain, and when the outcomes or model itself are poorly known.

Beginner 11 min read

What You'll Learn

  • Distinguish measurable risk, probability ambiguity, and deeper uncertainty
  • Separate uncertainty about an event from uncertainty about its probability
  • Interpret probability estimates without treating them as known facts
  • Use probability ranges and sensitivity checks when likelihoods are ambiguous
  • Recognize when a model omits outcomes or cannot support a numerical estimate
  • Identify common errors in communicating quantified and unquantified risks

Introduction

Risk analysis often begins with a description of uncertain outcomes and the exposure to those outcomes. The next question is how much can be known about them. For some situations, a probability model can be estimated and tested against relevant data. For others, the possible outcomes may be understood but their probabilities are poorly known. In the most difficult cases, decision-makers may not know all the relevant outcomes or which model is appropriate.

These distinctions matter because a numerical probability can suggest more certainty than the evidence supports. A useful analysis makes clear whether a probability is based on stable, relevant evidence, is a judgment or range, or cannot be credibly assigned at all. This tutorial develops those distinctions in a form aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: For this tutorial, measurable risk means uncertainty about outcomes that can be represented by a probability model with probabilities that are meaningfully estimable for the decision being considered. Uncertainty is the broader condition of not knowing which outcome will occur. Ambiguity is uncertainty about the probabilities, outcomes, or model itself, such that a single probability distribution is not adequately supported. These terms are used differently across disciplines, so an analysis should state its intended meaning.

Three Questions for Classifying a Situation

A practical classification asks three separate questions. First, are the relevant outcomes sufficiently well specified? Second, can their probabilities be estimated from relevant data or a defensible model? Third, are the assumptions behind that model credible for the exposure and time period being assessed?

  • Outcomes and probabilities are estimable: a model can assign probabilities to specified outcomes, with evidence supporting those assignments. The resulting probabilities remain estimates, not guarantees.
  • Outcomes are understood, but probabilities are ambiguous: the event types can be described, yet available evidence supports a range of probabilities, competing estimates, or only a qualitative likelihood.
  • Outcomes or the model are poorly known: the event set may be incomplete, or there may be no defensible basis for assigning probabilities. This is often called deep uncertainty or model uncertainty.

The categories are not permanent labels attached to an event. A situation may be measurable for one portfolio and ambiguous for another if the data quality, exposure, or model differs. It may also become more measurable as experience accumulates, or less measurable when conditions change and old data cease to be relevant.

A probability estimate should therefore be read as a statement conditional on a model and its assumptions. For example, “the annual probability is 4%” is incomplete unless the event, population, time period, data or method, and key assumptions are understood. The number does not by itself establish that the model will remain valid.

Probability Estimates and Probability Ranges

For a specified event \(A\), \(P(A)\) denotes its probability under a stated probability model. A probability is a number from 0 to 1. If outcomes are mutually exclusive and collectively exhaustive, their probabilities sum to 1. When a probability is estimated from data, it reflects both the data and the assumptions used to apply those data to the situation at hand.

A simple expected-loss calculation illustrates the role of probabilities. Suppose a loss \(L\) can take values \(L_1,\ldots,L_k\), with probabilities \(p_1,\ldots,p_k\). Under this discrete model, its expected value is the probability-weighted average:

$$ E[L]=\sum_{j=1}^{k}p_jL_j. $$

This formula requires the outcomes and their probabilities to be specified, and the probabilities to sum to 1. It does not establish that the model is appropriate or that its inputs are known precisely. If the probability is only known to lie within a range, the expected value may also be a range rather than a single well-supported number.

Key distinction: A calculation can be arithmetically exact while its inputs remain uncertain. Report the result at a precision justified by the evidence, and state whether probabilities are estimated, assumed for analysis, or not credibly available.

Worked Examples

Worked Example: Applying an Estimated Event Probability

A hypothetical insurer uses data from comparable installations to estimate that a particular type of equipment has a 4% probability of a specified failure during one year. It is considering 10 installations. For an initial screening calculation, assume the annual failures at different installations are independent and each has the same probability, \(p=0.04\). Find the expected number of failures and the probability of at least one failure. Explain what the calculation says about measurability.

Method. Treat the probability estimate as fixed for this calculation, and use the stated independence assumption. This is a measurable-risk calculation conditional on the model, not proof that the true probability is exactly 4%.

For each installation, the probability of failure is 0.04. The expected number across 10 installations is the sum of their individual expected failure indicators:

$$ E[N]=10(0.04)=0.40. $$

Here \(N\) is the number of installations that fail during the year. To calculate the probability of at least one failure, first calculate the probability that none fail. Independence gives:

$$ P(N=0)=(1-0.04)^{10}=0.96^{10}=0.6648326\ldots $$

Therefore:

$$ P(N\geq 1)=1-P(N=0)=1-0.6648326\ldots=0.3351674\ldots $$

To four decimal places, the probability of at least one failure is 0.3352, or 33.52%. The expected number, 0.40, is an average across repeated groups of 10 under the model; it is not a prediction that 0.40 of an installation will fail.

Check. The probability of at least one failure is greater than the 4% probability for a single installation, as expected when there are 10 opportunities for failure. The main trap is to present 33.52% as certain merely because the arithmetic is precise. The result depends on the estimated 4% probability, comparable exposure, and independence. If failures are dependent or the estimate is not relevant to the installations, this calculation does not apply as stated.

Worked Example: A Probability Range and an Uncertain Decision

A hypothetical infrastructure fund is evaluating a facility with an annual loss of $500,000 if a specified event does not occur and $5,000,000 if it does. Available evidence supports an annual event probability between 2% and 8%, but does not justify one value within that interval. Calculate the range of expected annual loss. If a planning threshold is $700,000, determine whether the expected-loss calculation alone puts the facility above or below the threshold.

Method. Use the two-outcome expected-loss formula, retaining the event probability \(q\) as a range. The model assumes exactly these two outcomes, and that their dollar amounts are the relevant annual losses.

For a specified \(q\), the expected annual loss is:

$$ E[L]=(1-q)(\$500{,}000)+q(\$5{,}000{,}000) =\$500{,}000+\$4{,}500{,}000q. $$

This expression increases as \(q\) increases, because the loss when the event occurs is larger than the loss when it does not. At the lower endpoint, \(q=0.02\):

$$ E[L]=\$500{,}000+\$4{,}500{,}000(0.02) =\$500{,}000+\$90{,}000 =\$590{,}000. $$

At the upper endpoint, \(q=0.08\):

$$ E[L]=\$500{,}000+\$4{,}500{,}000(0.08) =\$500{,}000+\$360{,}000 =\$860{,}000. $$

Thus the expected annual loss is between $590,000 and $860,000 under the stated assumptions. The range crosses the $700,000 threshold, so the available probability information does not establish that expected loss is either above or below the threshold.

The probability at which expected loss equals the threshold can be found by solving:

$$ \$500{,}000+\$4{,}500{,}000q=\$700{,}000, \qquad q=\frac{\$200{,}000}{\$4{,}500{,}000}=0.044444\ldots. $$

Probabilities below about 4.44% imply an expected loss below the threshold; probabilities above about 4.44% imply an expected loss above it. Since the supported interval includes values on both sides, a single point estimate chosen without justification could change the apparent conclusion. The common mistake is to select the midpoint probability, 5%, and report $725,000 as though the evidence identified that value. The correct conclusion is that the probability is ambiguous within the stated range and the threshold comparison is unresolved by this calculation.

Worked Example: A Scenario That the Baseline Model May Omit

A hypothetical insurer’s baseline model assigns a 1% annual probability to a defined equipment outage that would produce a $2,000,000 loss. An engineering review also identifies a different, emerging failure mode with a possible $8,000,000 loss. The review cannot estimate its annual probability; for sensitivity analysis only, management asks to examine probabilities from 0% to 0.5%. Assume the two loss components are distinct and additive. Calculate the baseline expected loss and the sensitivity range for the combined expected loss. State what can and cannot be concluded.

Method. Calculate the baseline expected loss from its specified probability model, then calculate a conditional sensitivity range for the additional scenario. The 0% to 0.5% interval is an analysis assumption, not an empirical estimate.

Under the baseline model, the expected loss is:

$$ 0.01(\$2{,}000{,}000)=\$20{,}000. $$

Let \(q\) be the assumed annual probability for the emerging failure mode. Its sensitivity contribution is \(q(\$8{,}000{,}000)\). At the lower bound, \(q=0\):

$$ \$20{,}000+0(\$8{,}000{,}000)=\$20{,}000. $$

At the upper bound, \(q=0.005\):

$$ \$20{,}000+0.005(\$8{,}000{,}000) =\$20{,}000+\$40{,}000 =\$60{,}000. $$

Under this sensitivity range and the additive-loss assumption, combined expected loss ranges from $20,000 to $60,000. This is not evidence that the emerging failure mode’s expected contribution lies in that range: it is the result of the assumed probability bounds. If the probability is not credibly bounded, even this range may not be informative.

The baseline calculation is conditional on the baseline model’s event definition. It does not show that every relevant failure mode has been included. The common trap is to treat a precise baseline result of $20,000 as a complete assessment of all outage risk. A model can calculate precisely for the outcomes it includes and still omit relevant outcomes or fail to represent changing conditions.

Why the Distinction Matters

Measurable risk supports quantitative comparisons, but it does not remove the need to examine evidence and assumptions. A model estimated from historical experience may be unsuitable if the exposure has changed, the data are sparse, or future conditions differ materially from the past. In those cases, a probability may be numerically available without being well supported.

When probabilities are ambiguous, ranges, alternative assumptions, and sensitivity analysis make the uncertainty visible. They should not be presented as if every value in a range were equally likely unless there is evidence for that interpretation. When outcomes or the model are poorly known, scenario descriptions and explicit data limitations may be more honest than a fabricated point probability.

A useful justification for the two-outcome formula is to represent loss as \(L_0\) when an event does not occur and \(L_1\) when it does. Let \(I\) equal 1 if the event occurs and 0 otherwise. Then \(L=L_0+(L_1-L_0)I\). Since \(I\) is 1 with probability \(q\) and 0 with probability \(1-q\), \(E[I]=q\). Taking expectations gives:

$$ E[L]=E[L_0+(L_1-L_0)I] =L_0+(L_1-L_0)q =(1-q)L_0+qL_1. $$

This derivation explains why the expected loss changes linearly with the event probability when the two loss amounts are fixed. It also shows the formula’s limits: if the outcomes are incomplete, the loss amounts are not specified, or the probability \(q\) is not supported, the expression may be mathematically valid but insufficient for a reliable conclusion.

Common Mistakes

  • Equating a probability estimate with a known probability. An estimate derived from data remains subject to data limitations and model assumptions.
  • Using a point estimate where only a range is supported. A selected midpoint is not automatically the most likely probability.
  • Confusing lack of data with zero probability. Not knowing how often an event occurs does not imply that it cannot occur.
  • Assuming a detailed model is complete. A model may omit outcomes that are difficult to anticipate or quantify.
  • Reporting precision beyond the evidence. More decimal places do not make an uncertain input more reliable.
Key takeaway: Before relying on a numerical risk result, state what outcomes the model includes, how its probabilities were obtained, and which assumptions are uncertain. Use a probability model when it is supportable; use ranges or scenarios when it is not; and identify material uncertainty that cannot be quantified credibly.

Check Your Understanding

For each problem, distinguish what is known from what is assumed, and show the relevant calculation where appropriate.

  1. A hypothetical service provider estimates a 3% probability of a defined annual system disruption from relevant experience. For six independent service units with the same probability, find the probability of at least one disruption. State one assumption needed. (Answer: \(1-0.97^6=0.1670\), or 16.70%, assuming independence and a common 3% probability.)
  2. A hypothetical facility’s annual loss is $200,000 if an event does not occur and $1,200,000 if it does. The event probability is only supportable within 1% to 4%. Find the expected-loss range. (Answer: $210,000 to $240,000.)
  3. A model assigns a 2% probability to a $3,000,000 loss, but a review identifies a separate possible $5,000,000 loss whose probability is unknown. Explain why the baseline expected loss is not a complete assessment and give the baseline figure. (Answer: baseline expected loss is $60,000; the additional loss is not quantified, so the baseline alone does not establish total expected loss.)
  4. For a two-outcome loss of $100,000 without an event and $900,000 with an event, what event probability makes expected loss equal $300,000? If evidence supports probabilities from 10% to 30%, does the range lie on one side of that expected-loss threshold? (Answer: 25%; no, the supported range includes probabilities below and above 25%, so the expected-loss range crosses $300,000.)
  5. A newly introduced technology has no relevant operating history, and reviewers cannot agree on the complete set of possible failure outcomes. Is assigning a precise annual probability necessarily a measurable-risk assessment? Explain. (Answer: No. The outcomes and model are not sufficiently established; a precise probability would require unsupported assumptions and should not be presented as well-measured.)

Previous

Next

Back to Tutorials
Clinical Biostats