Why a Reserve Range Needs Interpretation
A reserve estimate is often presented as a central amount accompanied by a low and a high amount. The three numbers can help a reader understand the estimate’s context, but the bounds do not explain themselves. A range might summarize alternative plausible estimates, a judgment about uncertainty, or percentiles from a probability model. Those interpretations are not interchangeable.
The first task is to identify what the amounts represent and whether they share the same measurement basis. For example, a range should be understood in relation to its valuation date and whether it is gross, ceded, or net of reinsurance. Its bounds also need a stated basis: they might be outputs from selected assumptions or quantiles from a modeled distribution. Without that information, a reader cannot infer what the range says about the likelihood of future outcomes.
What the Bounds Communicate
A range can communicate that the selected reserve is not a uniquely certain amount. Its width shows the span between the reported bounds, while its position relative to the central estimate shows whether the central estimate is centered in that span. These are descriptive features of the reported amounts; neither feature assigns probabilities to possible outcomes.
For a central estimate \(C\), low bound \(L\), and high bound \(H\), the absolute distances from the central estimate are \(C-L\) and \(H-C\). Relative distances can be expressed as percentages of \(C\), provided \(C\) is positive. A useful comparison is:
These calculations describe the range’s geometry. If the upper distance exceeds the lower distance, the reported interval is asymmetric around the central estimate. That asymmetry may reflect how the bounds were selected, but it does not establish that adverse outcomes are more likely.
Worked Example: Reading an Asymmetric Reported Range
A hypothetical insurer reports a net reserve central estimate of $101 million, with a low bound of $92 million and a high bound of $118 million. All three amounts are for the same valuation date and reporting basis. Calculate the distances and describe what the range supports.
Step 1: Calculate the lower distance. The difference between the central estimate and the low bound is \(101-92=9\) million dollars. Relative to the central estimate, this is \(9/101=0.0891089\), or 8.91% when expressed to two decimal places.
Step 2: Calculate the upper distance. The difference between the high bound and the central estimate is \(118-101=17\) million dollars. Relative to the central estimate, this is \(17/101=0.1683168\), or 16.83% to two decimal places.
Step 3: Calculate the total width. The range width is \(118-92=26\) million dollars.
Interpretation. The high bound is $17 million above the central estimate, while the low bound is $9 million below it. The interval is asymmetric around $101 million. These facts do not show that the reserve has a 50% chance of being above or below the central estimate, nor do they establish any other probability. The method used to select the bounds would be needed for that interpretation.
Common mistake avoided. Calling $101 million the midpoint would be incorrect: the midpoint is \((92+118)/2=105\) million dollars. A selected central estimate need not equal the midpoint of its reported range.
A Range Is Not Automatically a Probability Interval
A probability interval requires a probability model for the quantity of interest and a rule for selecting its endpoints. For example, the endpoints might be the 10th and 90th percentiles of a specified distribution. In that case, the interval’s probability content can be determined from the distribution and the percentile convention. A range selected by judgment or by a set of alternative estimates has no such probability content unless a probabilistic interpretation is separately established.
The distinction matters in reserve reporting. “The estimate is between these two values under the selected range construction” is not equivalent to “there is an 80% probability that the ultimate liability is between these two values.” The second statement makes a probabilistic claim, and requires a model and assumptions that support it. A central estimate also does not automatically identify a particular probability statistic: depending on its definition, it could represent a mean, median, or selected actuarial estimate.
Percentile Intervals Under a Continuous Model
Suppose \(R\) is a modeled reserve amount with cumulative distribution function \(F_R(r)=P(R\leq r)\). Use the quantile convention \(Q(p)=\inf\{r:F_R(r)\geq p\}\). If the distribution is continuous and the selected quantiles satisfy \(F_R(Q(p))=p\), the probability between the 10th and 90th percentiles follows directly from the CDF values.
The endpoint inclusions do not change this result for a continuous distribution because the probability of either exact endpoint is zero. The calculation depends on the stated continuous model and the percentile endpoints. It does not establish that an independently reported reserve range has 80% probability.
Worked Example: An 80% Interval Under a Normal Model
For a hypothetical portfolio, suppose the modeled ultimate net reserve \(R\), measured in millions of dollars, follows \(R\sim N(100,10^2)\). Thus, the model assumes a normal distribution with mean $100 million and standard deviation $10 million. Find the interval from the 10th to the 90th percentile and verify its probability under this model.
Step 1: Identify the standard-normal quantiles. For the standard normal distribution, the 10th and 90th percentile values are approximately \(-1.2816\) and \(1.2816\). The model is continuous, so its percentile interval has probability equal to the difference between the corresponding CDF values.
Step 2: Convert each quantile to reserve dollars. For this normal model, a percentile corresponding to standard-normal value \(z\) is \(100+10z\), in millions of dollars. The lower endpoint is \(100+10(-1.2816)=87.184\) million. The upper endpoint is \(100+10(1.2816)=112.816\) million.
Step 3: Report and check. Rounding only the final endpoints to two decimal places gives $87.18 million to $112.82 million. Under the specified model, the probability is \(\Phi(1.2816)-\Phi(-1.2816)\), approximately \(0.9000-0.1000=0.8000\). Thus the interval has 80.00% probability under this continuous normal model.
Common mistake avoided. Using \(1.645\) as the standard-normal multiplier would give approximately $83.55 million to $116.45 million. Those are the 5th and 95th percentiles, an approximately 90% central interval under this model, not the requested 10th-to-90th percentile interval. The percentile levels must match the requested interval.
Why Atoms Require an Actual Probability Check
A distribution with atoms assigns positive probability to one or more exact values. Then an interval with percentile endpoints need not have probability equal to the difference between the percentile labels. The endpoint convention matters because including a value with positive probability can add a jump in the CDF. With the quantile convention \(Q(p)=\inf\{r:F_R(r)\geq p\}\), calculate the probability of the stated interval from the distribution itself rather than assuming it equals 80%.
Worked Example: Percentile Endpoints With a Discrete Reserve Distribution
A hypothetical simplified reserve model assigns probabilities as follows: \(R=80\) million with probability 0.10, \(R=100\) million with probability 0.80, and \(R=120\) million with probability 0.10. Use \(Q(p)=\inf\{r:F_R(r)\geq p\}\). Find the 10th and 90th percentile endpoints, calculate the probability in the closed interval between them, and identify the central estimate if it is defined as the mean.
Step 1: Build the cumulative probabilities. At $80 million, \(F_R(80)=0.10\). At $100 million, \(F_R(100)=0.10+0.80=0.90\). At $120 million, \(F_R(120)=1.00\).
Step 2: Apply the stated quantile convention. The first value at which the CDF reaches at least 0.10 is $80 million, so \(Q(0.10)=80\). The first value at which it reaches at least 0.90 is $100 million, so \(Q(0.90)=100\). The closed percentile interval is therefore $80 million to $100 million.
Step 3: Calculate the interval probability directly. Both reserve values $80 million and $100 million lie in the closed interval. Its probability is \(0.10+0.80=0.90\), or 90%, not 80%. The atom at $80 million is included at the lower endpoint, and the atom at $100 million is included at the upper endpoint.
Step 4: Calculate the mean. The mean is \(80(0.10)+100(0.80)+120(0.10)=8+80+12=100\) million dollars. Under the specified definition, the central estimate is $100 million, which is the upper endpoint of this interval.
Check. The probabilities sum to \(0.10+0.80+0.10=1.00\), and the interval probability is \(0.90\). The percentile labels alone do not justify reporting an 80% probability for this closed interval.
Scenario-Based Bounds and Clear Reporting
Another common construction is to select bounds under specified alternative assumptions or estimation approaches. Such bounds can be useful for showing the span of estimates considered, but the resulting range is not a probability interval just because its endpoints are called “low” and “high.” If no probability distribution or probability weights are assigned to the alternatives, the range does not say how likely any point within it is.
Worked Example: Interpreting a Scenario-Based Range
A hypothetical insurer reports a central net reserve estimate of $133 million. Its reporting team describes $125 million and $151 million as low and high bounds representing two selected alternative estimation bases. No probability distribution or weights are assigned to those alternatives. Calculate the range’s midpoint and distances from the central estimate, then state what can be concluded.
Step 1: Calculate the midpoint. The midpoint is \((125+151)/2=138\) million dollars.
Step 2: Compare the central estimate with the bounds. The central estimate is \(133-125=8\) million dollars above the low bound and \(151-133=18\) million dollars below the high bound. It is $5 million below the midpoint.
Step 3: Interpret the construction. The central estimate lies within the reported range, and the range is not symmetric around it. Because the two bounds are alternative estimates without assigned probabilities, the information does not support a probability that the ultimate reserve will fall within the range. It also does not establish that $138 million is a better central estimate merely because it is the midpoint.
Common mistake avoided. Treating the low and high alternatives as equally likely would require an explicit model or probability assignment. The fact that there are two alternatives does not make each one a 50% outcome.
Common Interpretation Traps
- Equating range bounds with confidence limits. Bounds are not confidence limits unless a statistical construction and its assumptions justify that description.
- Assuming the central estimate is the midpoint. The central estimate may be selected using a different criterion and can sit off-center within the range.
- Reading asymmetry as likelihood. A longer distance above the estimate is a geometric feature, not proof that higher outcomes are more probable.
- Ignoring the distribution type. For a continuous model, a central 10th-to-90th percentile interval has 80% probability when the quantiles are correctly defined. For a distribution with atoms, calculate the probability of the actual interval using the stated quantile convention.
- Mixing measurement bases. Bounds and central estimates should be compared only after confirming that they refer to the same valuation date and liability basis.
For clear reporting, name the central estimate, state the range and its basis, and explain whether the bounds are judgmental or probabilistic. If they are probabilistic, identify the modeled quantity, distribution or modeling approach, and endpoint convention. This makes the range’s meaning explicit without making a probability claim that its construction does not support.
This interpretation is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus: candidates should distinguish a reserve amount from the uncertainty information attached to it, and should match probability statements to the assumptions and calculations that support them.
Check Your Understanding
For each problem, interpret the stated range using only the information provided.
- A hypothetical insurer reports a central estimate of $250 million, with bounds of $232 million and $279 million on the same basis. Calculate the lower and upper distances as percentages of the central estimate, to two decimal places. Is the central estimate the midpoint? (Answer: Lower distance $18 million, or 7.20%; upper distance $29 million, or 11.60%; midpoint $255.5 million, so no.)
- A modeled reserve is continuous and normal with mean $60 million and standard deviation $8 million. Using standard-normal 10th and 90th percentile values of approximately \(-1.2816\) and \(1.2816\), calculate the interval endpoints to the nearest cent and state its probability under the model. (Answer: $49.75 million to $70.25 million; 80%.)
- A discrete reserve distribution has values $40 million, $70 million, and $100 million with probabilities 0.20, 0.70, and 0.10, respectively. With \(Q(p)=\inf\{r:F_R(r)\geq p\}\), find the 20th and 80th percentile endpoints and the probability in the closed interval between them. (Answer: Endpoints $40 million and $70 million; interval probability 90%.)
- A report gives a central estimate of $133 million and a scenario-based range of $125 million to $151 million, with no probabilities assigned to the scenarios. Calculate the midpoint and explain whether the report establishes the probability that the reserve falls within the range. (Answer: Midpoint $138 million; no probability can be inferred from the stated information.)
- A report calls its low and high bounds the 5th and 95th percentiles of a specified continuous distribution. What is the probability between those endpoints, and what additional information would be needed before treating a separate judgmental range as having that probability? (Answer: 90% under the stated continuous distribution; the separate range needs a specified probabilistic construction and endpoint basis.)