Why Interval Shape Matters
A reserve estimate and its standard error summarize a predictive distribution, but they do not describe its shape. A symmetric normal prediction interval may be a useful approximation when prediction errors are reasonably symmetric. It can be misleading when plausible outcomes cluster near zero and a smaller number of outcomes lie far to the right. In that setting, the distance from the estimate to a lower percentile need not resemble the distance to an upper percentile.
The issue is especially visible when an unpaid-claim reserve is small relative to the uncertainty or when the reserve is constrained to be nonnegative. A symmetric interval can extend below zero, even though the modeled reserve cannot. Simply replacing the negative endpoint by zero does not establish that the resulting interval has the intended predictive coverage. A predictive distribution obtained by simulation offers another way to estimate interval endpoints: take the desired quantiles of the simulated outcomes.
What the Normal Approximation Assumes
If the prediction error \(R-\widehat{R}\) is approximately normal, centered at zero, and has estimated mean squared prediction error \(\widehat{\operatorname{MSEP}}\), the usual approximate central 95% interval is \(\widehat{R}\pm 1.96\sqrt{\widehat{\operatorname{MSEP}}}\). The square root supplies a prediction-error scale; the factor 1.96 gives equal tail probabilities under a standard normal model. This approximation concerns the shape of the predictive error, not merely the calculation of its variance.
The symmetry follows directly from the normal quantiles. If \(Z\sim N(0,1)\), then \(P(-1.96\leq Z\leq 1.96)\) is approximately 0.95, with equal tail probabilities of about 0.025. Rescaling and shifting \(Z\) gives equal distances from the center:
That reasoning relies on a normal approximation to the predictive error. It does not follow from knowing only its mean and variance. A right-skewed distribution can have a long upper tail and a short lower tail, so its central quantiles are not generally equally spaced around its mean. If \(R\) is a gross unpaid-claims reserve, it is also ordinarily nonnegative. A normal interval with a negative lower endpoint therefore signals that the approximation may be poorly suited to the modeled outcome; it is not evidence that negative gross claims are possible.
Reading Simulation-Based Predictive Intervals
Suppose a predictive simulation produces outcomes \(R_1,\ldots,R_B\), each generated under the assumptions chosen for the reserve analysis. Sort them from smallest to largest. The empirical quantile at probability \(p\) is the outcome at the first sorted position whose cumulative proportion is at least \(p\). For example, with 1,000 outcomes, the empirical 2.5th and 97.5th percentiles are found around sorted positions 25 and 975 under the inverse empirical CDF convention.
A simulation-based interval reflects the shape represented by those outcomes. It can be asymmetric, and its lower endpoint can respect a nonnegative constraint if every simulated reserve is nonnegative. Its interpretation still depends on the simulation model: the outcomes must represent the uncertainty the analyst intends to include. A simulation that omits parameter uncertainty, for example, cannot show the effects of that omitted uncertainty merely by using many runs. Nor does a large number of runs correct misspecified assumptions. Simulation estimates quantiles of the specified predictive distribution; it does not validate that distribution.
Worked Example: A Long Right Tail and a Negative Normal Endpoint
A hypothetical portfolio has a simulated future gross reserve \(R\), measured in thousands of dollars. For this example, assume no discounting or interest adjustment, and no mortality component. The complete predictive distribution is \(P(R=0)=0.50\), \(P(R=20)=0.30\), and \(P(R=80)=0.20\). Compare its central 95% interval with a normal interval based on its mean and standard deviation.
Method. Calculate the predictive mean and variance, form the normal approximation, and then read the exact equal-tailed quantiles from the stated distribution. This comparison avoids treating a symmetric interval as if its shape followed from the mean and variance alone.
The mean is \(E[R]=0(0.50)+20(0.30)+80(0.20)=0+6+16=22\) thousand dollars. The second moment is \(E[R^2]=0^2(0.50)+20^2(0.30)+80^2(0.20)=0+120+1{,}280=1{,}400\). Thus \(\operatorname{Var}(R)=1{,}400-22^2=1{,}400-484=916\), and the standard deviation is \(\sqrt{916}=30.265\) thousand dollars (rounded to three decimals).
The normal 95% interval is \(22\pm1.96(30.265)=22\pm59.319\), giving endpoints \(-37.319\) and \(81.319\) thousand dollars. The exact cumulative probabilities are \(F_R(0)=0.50\), \(F_R(20)=0.80\), and \(F_R(80)=1.00\). Therefore \(Q(0.025)=0\) and \(Q(0.975)=80\), so the equal-tailed interval is \([0,80]\) thousand dollars.
Check. The normal interval is centered at 22, while the exact interval is not: the mass at zero and the upper-tail outcomes make the distribution asymmetric. Both results use the same predictive distribution’s moments where relevant, but the normal interval has an impossible negative endpoint for this nonnegative reserve. All displayed interval endpoints are rounded to three decimals.
Small Reserve Bases and Discrete Outcomes
A small reserve base can make the normal approximation particularly fragile when the number of material future claims is limited. The predictive outcomes may then be discrete, with a substantial probability of no further payment and a small probability of a large payment. Those features can create a large coefficient of variation, pronounced right skewness, and a sizable point mass at the lower bound. Multiplying the reserve and its standard deviation by a common currency scale does not, by itself, change the distribution’s shape. The concern is the underlying predictive pattern, which small or sparse portfolios may expose more clearly.
Worked Example: A Small Reserve With Rare Large Payments
Consider a hypothetical small portfolio whose future gross reserve \(R\), in thousands of dollars, has the following predictive distribution: \(P(R=0)=0.80\), \(P(R=5)=0.15\), and \(P(R=25)=0.05\). Assume no discounting or interest adjustment and no mortality component. Compare the normal approximation with the exact central 95% interval.
Method. Use the distribution’s moments for the normal interval and its cumulative probabilities for the exact quantiles. The key trap is to infer that a small mean implies small uncertainty.
The mean is \(E[R]=0(0.80)+5(0.15)+25(0.05)=0+0.75+1.25=2\) thousand dollars. The second moment is \(E[R^2]=0+25(0.15)+625(0.05)=3.75+31.25=35\). Hence \(\operatorname{Var}(R)=35-2^2=31\), and \(\operatorname{SD}(R)=\sqrt{31}=5.568\) thousand dollars (rounded to three decimals).
The normal interval is \(2\pm1.96(5.568)=2\pm10.913\), or approximately \([-8.913,12.913]\) thousand dollars. For the exact distribution, \(F_R(0)=0.80\), \(F_R(5)=0.95\), and \(F_R(25)=1.00\). Both \(Q(0.025)\) and \(Q(0.975)\) are 0 and 25, respectively. The exact interval is therefore \([0,25]\) thousand dollars.
Check. A reserve mean of only 2 thousand dollars does not prevent a 25-thousand-dollar outcome in this model. The normal approximation’s lower endpoint is negative, while its upper endpoint is below the possible 25-thousand-dollar outcome. The exact equal-tailed interval respects the distribution’s point mass and upper tail. Endpoints are rounded to three decimals.
Calculating Percentiles From Simulated Outcomes
The same quantile method applies when the predictive distribution is represented by simulation rather than given as a probability table. The sample mean and standard deviation can still be calculated, but simulation also makes the empirical distribution available for direct percentile estimation. With discrete outcomes, quantiles may repeat or jump across a wide range: there may be no simulated value between two adjacent outcomes. Accordingly, empirical interval endpoints need not vary smoothly as the number of simulations or the requested confidence level changes.
Worked Example: Comparing a Simulated Interval With a Normal Interval
A hypothetical reserve simulation produces 1,000 future gross reserve outcomes, measured in thousands of dollars: 50 outcomes equal 0, 750 equal 10, 150 equal 20, and 50 equal 100. Assume no interest adjustment or mortality component. Calculate the empirical central 95% interval and compare it with the moment-based normal interval.
Method. Treat the listed counts as the empirical predictive distribution. First calculate its mean and variance; then use the specified inverse empirical CDF convention for its quantiles. This avoids replacing a simulated skewed distribution with a symmetric one without checking the consequences.
The mean is \((50(0)+750(10)+150(20)+50(100))/1{,}000=(0+7{,}500+3{,}000+5{,}000)/1{,}000=15.5\) thousand dollars.
The second moment is \((50(0^2)+750(10^2)+150(20^2)+50(100^2))/1{,}000=(0+75{,}000+60{,}000+500{,}000)/1{,}000=635\). Thus the empirical variance is \(635-15.5^2=635-240.25=394.75\), and the standard deviation is \(\sqrt{394.75}=19.868\) thousand dollars (rounded to three decimals).
The normal interval is \(15.5\pm1.96(19.868)=15.5\pm38.941\), or approximately \([-23.441,54.441]\) thousand dollars. In the sorted simulation, positions 1 through 50 are zero, 51 through 800 are 10, 801 through 950 are 20, and 951 through 1,000 are 100. Position \(\lceil0.025(1{,}000)\rceil=25\) has value 0; position \(\lceil0.975(1{,}000)\rceil=975\) has value 100. The empirical interval is \([0,100]\) thousand dollars.
Check. Fifty of the 1,000 simulated outcomes equal 100, so the upper empirical percentile lies in that group. The normal upper endpoint, 54.441, does not reach those outcomes, and the normal lower endpoint is negative. This does not prove that the empirical interval will be accurate in a new simulation; it shows how the interval changes when the simulated distribution, rather than a normal shape assumption, determines its endpoints. Interval endpoints are rounded to three decimals.
Common Exam Trap: Clipping the Normal Interval
A frequent mistake is to calculate a normal interval with a negative lower endpoint and then replace that endpoint by zero, while leaving the upper endpoint unchanged. The result may be operationally convenient, but clipping changes the interval without recalculating its predictive probability. It is not automatically a central 95% interval, nor does it inherit 95% coverage from the original normal calculation. The correct interpretation depends on the predictive distribution and the interval method chosen.
For example, in the first worked example, clipping the normal interval \([-37.319,81.319]\) to \([0,81.319]\) does not give the exact central 95% interval. The exact interval from the specified predictive distribution is \([0,80]\). The difference is not a rounding issue; it comes from using a normal shape that the distribution does not have.
A reliable comparison should state the target interval level, the predictive distribution being represented, and the quantile convention used for simulation outcomes. It should also keep the uncertainty target clear: a prediction interval for future reserve outcomes is not merely an interval for the estimated expected reserve. The former describes possible future outcomes under its predictive assumptions; the latter would describe uncertainty about a mean or parameter.
This distinction matters in reserve communication. A symmetric interval can conceal how much risk lies in an upper tail, while an interval with a nonnegative lower endpoint can still have an upper endpoint that understates large outcomes. Presenting simulation-based quantiles alongside a normal approximation, where useful, makes the shape assumption visible. Such comparisons are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
Check Your Understanding
For each problem, use the stated predictive distribution or simulation convention. Monetary amounts are in thousands of dollars; assume no interest adjustment or mortality component unless stated otherwise.
- A hypothetical gross reserve has \(P(R=0)=0.60\), \(P(R=10)=0.30\), and \(P(R=50)=0.10\). Find its central 95% equal-tailed interval. (Answer: \([0,50]\) thousand dollars.)
- A simulation has 200 outcomes: 20 are 0, 150 are 4, 25 are 20, and 5 are 80. Using \(Q(p)=\inf\{r:F(r)\geq p\}\), find the central 95% empirical interval. (Answer: \([0,20]\) thousand dollars.)
- A normal approximation for a nonnegative reserve has mean 12 and standard deviation 8. Calculate its approximate central 95% interval using 1.96. What feature makes it a warning sign? (Answer: \([-3.68,27.68]\) thousand dollars; its negative lower endpoint is incompatible with a nonnegative gross reserve.)
- A simulation of 100 reserve outcomes contains 70 values of 2, 25 values of 5, and 5 values of 25. Calculate its empirical mean and standard deviation using divisor 100 for the empirical variance, then find its central 95% interval using the inverse empirical CDF convention. (Answer: mean \(3.9\) thousand dollars, standard deviation approximately \(5.009\) thousand dollars, interval \([2,25]\) thousand dollars.)
- Suppose a nonnegative reserve’s central equal-tailed interval from simulation is \([0,60]\), while its normal approximation is \([-15,35]\). Give two distinct reasons not to report the clipped normal interval \([0,35]\) as a calibrated central 95% interval. (Answer: clipping does not preserve the normal interval’s nominal coverage or equal-tail construction, and the normal upper endpoint may omit material right-tail outcomes represented in the simulated distribution.)