Why Method Choice Matters
An aggregate risk estimate depends on more than the statistic being requested. It also depends on the information available and the assumptions used to represent future losses. A large historical data set may support an empirical estimate, while a small data set may require a parametric model. A discrete frequency-severity model may permit an exact recursion; simulation may be more flexible when the model is complicated. Each method has limits, especially in the tail.
The decision is not simply “exact” versus “approximate.” An exact calculation under a poorly supported model can be less useful than a transparent empirical estimate. Conversely, a short historical record may not contain enough extreme observations to estimate a high quantile reliably. A sound analysis states the target, the data or model used, the method, and the uncertainty or limitations that remain.
What Each Method Uses
| Method | Primary input | Best suited to | Main limitation |
|---|---|---|---|
| Empirical | Observed aggregate outcomes | Direct estimates when observations are numerous and representative | Little information about outcomes beyond the observed range |
| Parametric | A specified probability model and estimated parameters | Structured extrapolation and calculations from a fitted model | Results depend on model form and parameter estimates |
| Recursive | A discrete aggregate model with suitable frequency structure | Computing aggregate probabilities accurately on a grid | Requires usable discrete severity and frequency inputs |
| Normal approximation | Aggregate mean and variance | Rough estimates for sums with sufficiently regular, not-too-skewed behavior | Can be poor in skewed tails or for discrete distributions |
| Simulation | A model from which aggregate outcomes can be generated | Flexible estimates for complicated models and coverage structures | Monte Carlo error can be substantial for rare tail events |
An empirical estimate is most defensible when the observations correspond to the same exposure period, population, and coverage definition as the risk being assessed. It makes few distributional assumptions, but cannot reliably describe tail outcomes absent from the data. A parametric method specifies a distributional structure, such as a frequency-severity model, and uses estimated or assumed parameters. It can produce estimates beyond the observed range, but that extrapolation is conditional on the model being appropriate.
A recursive method is a computational approach within a model, not a substitute for choosing a model. For example, for a compound Poisson aggregate with discrete severities, recursion can calculate aggregate probabilities without repeatedly simulating outcomes. A normal approximation summarizes the aggregate distribution with its mean and variance. Simulation draws repeated outcomes from a specified model, and is often flexible enough for complex features, but its output remains subject to simulation error as well as model uncertainty.
Recursive and Normal Calculations
For a compound model, let \(N\) be the claim count and let \(X_1,X_2,\ldots\) be independent claim amounts, independent of \(N\). Aggregate loss is \(S=\sum_{j=1}^{N}X_j\), with \(S=0\) when \(N=0\). If \(N\) is Poisson with mean \(\lambda\), and each claim amount has mean \(E[X]\) and second moment \(E[X^2]\), then
The variance uses the second moment of claim amount, not just its variance. This distinction matters when checking a normal approximation.
For discrete severities, a compound Poisson recursion can calculate the probability mass function. Suppose a claim amount takes positive integer values \(j\) with probabilities \(q_j\), and let \(g_s=P(S=s)\). Starting with \(g_0=e^{-\lambda}\), the recursion is
To see why, write the probability generating function of \(S\) as \(G(z)=\exp\{\lambda(Q(z)-1)\}\), where \(Q(z)=\sum_{j\geq1}q_jz^j\). Differentiating gives \(G'(z)=\lambda Q'(z)G(z)\). The coefficient of \(z^{s-1}\) on the left is \(s g_s\); on the right it is \(\lambda\sum_{j=1}^s j q_jg_{s-j}\). Equating coefficients gives the recursion. This calculation assumes Poisson frequency and positive integer-valued severities on a common grid; the formula should not be applied unchanged to continuous severities or a different frequency model.
A normal approximation instead uses \(S\) approximately distributed as \(N(E[S],\operatorname{Var}(S))\), where the second argument is the variance. It is more credible when aggregate loss is built from many contributions and is not dominated by a few large claims. It can be inaccurate for small frequencies, strong skewness, heavy tails, or discrete outcomes. A continuity correction may improve a lattice approximation, but it does not repair a fundamentally poor normal fit.
Worked Example: Empirical Tail Estimates From a Short Record
A hypothetical insurer records ten annual aggregate losses, in thousands of dollars, under unchanged coverage: \(0,0,10,10,20,20,30,60,100,200\). Calculate the empirical \(80\%\) VaR and \(80\%\) TVaR, using the empirical distribution and the fractional-tail convention.
Method. Sort the observations, then use the empirical distribution, which assigns probability \(0.1\) to each recorded year. The \(80\%\) quantile is the smallest value with cumulative probability at least \(0.8\). The upper \(20\%\) tail contains two years.
The cumulative probability through \(60\) is \(8/10=0.8\), so the empirical \(80\%\) VaR is \(60\) thousand dollars. The two largest observations are \(100\) and \(200\), so the average of the upper \(20\%\) is
Check. The VaR is the eighth ordered observation, \(60\); the tail average exceeds it, as expected. This estimate is fully determined by the sample but does not establish that losses above \(200\) are impossible. The common mistake is to treat the observed maximum as a reliable estimate of the far tail; ten years provide little direct evidence about rarer outcomes.
Worked Example: Recursive VaR and a Normal Check
A hypothetical portfolio has Poisson claim frequency with \(\lambda=2\) per year. Each claim is independently \(1\) or \(2\) thousand dollars, each with probability \(0.5\). Calculate the exact recursive \(90\%\) VaR, then compare it with a normal approximation.
Method. Use the compound Poisson recursion on the thousand-dollar grid. Here \(q_1=q_2=0.5\), all other \(q_j=0\), and \(g_0=e^{-2}=0.135335\). Applying the recursion successively gives
Thus \(P(S\leq5)=0.863890\), while \(P(S\leq6)=0.926107\). The smallest aggregate value with cumulative probability at least \(0.90\) is \(6\), so the recursive \(90\%\) VaR is \(6\) thousand dollars.
For a normal check, \(E[X]=1.5\) and \(E[X^2]=(1^2+2^2)/2=2.5\), in thousands of dollars and squared thousands of dollars, respectively. Therefore \(E[S]=2(1.5)=3\) and \(\operatorname{Var}(S)=2(2.5)=5\). The uncorrected normal quantile is
This is close to the exact lattice VaR, but it is not itself a possible aggregate outcome. Rounding the approximate value upward to the next grid point gives \(6\) thousand dollars. As a further check, the continuity-corrected normal estimate of \(P(S\leq5)\) is \(\Phi((5.5-3)/\sqrt{5})\), approximately \(0.868\), below \(0.90\); through \(6\) it is approximately \(\Phi((6.5-3)/\sqrt{5})=0.941\), above \(0.90\).
Common mistake. Using \(\lambda\operatorname{Var}(X)=2(0.25)=0.5\) as the aggregate variance is wrong. For a compound Poisson sum, the variance is \(\lambda E[X^2]=5\), not \(\lambda\operatorname{Var}(X)\). The exact recursion and approximate normal quantile agree at the rounded grid point here, but the recursion is preferred when an accurate discrete distribution is required.
Simulation and Precision
Simulation generates repeated aggregate outcomes under a specified model. For an event such as \(S>u\), if \(K\) of \(M\) independent simulated outcomes exceed \(u\), the estimated exceedance probability is \(\hat p=K/M\). Its estimated Monte Carlo standard error is
This standard error describes simulation variability conditional on the model and its parameters. It does not account for uncertainty in the model choice or estimated parameters. Tail quantiles also require care: the estimate is based on the simulated order statistics, and a small number of simulated tail outcomes can make it unstable. Increasing the number of simulations reduces Monte Carlo error, but does not make a misspecified model correct.
Worked Example: Interpreting a Simulated Exceedance Estimate
A hypothetical fitted model is used to simulate \(2{,}000\) annual aggregate losses. The threshold is \(80\) thousand dollars, and \(190\) simulated outcomes exceed it. Estimate the exceedance probability and a rough \(95\%\) interval based on Monte Carlo error.
Method. The estimate is \(\hat p=190/2{,}000=0.095\). Its estimated standard error is
Using the approximate normal interval \(\hat p\pm1.96\operatorname{SE}(\hat p)\), the margin is \(1.96(0.00656)=0.01285\). The resulting interval is \(0.095\pm0.01285\), or approximately \(0.0822\) to \(0.1079\). The estimated exceedance probability is \(0.0950\), with a rough \(95\%\) Monte Carlo interval of \(0.0822\) to \(0.1079\).
Check. The interval contains \(0.10\), so these simulations do not sharply distinguish a \(9.5\%\) exceedance probability from \(10\%\). The calculation measures simulation error only. A confidence statement about the true future exceedance probability would also need to address whether the fitted model and its parameters are appropriate.
A Practical Selection Process
Specify the aggregate period, coverage basis, and target, such as a quantile, tail average, or exceedance probability. Different targets can favor different methods.
Ask whether aggregate observations are numerous and comparable, or whether frequency and severity data support a fitted model. Sparse data make empirical tail estimates fragile; a parametric model introduces assumptions that must be disclosed.
If the model has discrete severities and a compatible frequency distribution, recursion may provide accurate aggregate probabilities. If only moments are readily available, a normal approximation may provide a useful check, subject to shape and tail limitations.
For simulation, quantify Monte Carlo error and ensure enough simulated tail outcomes. Compare methods or assumptions where possible, and distinguish sampling or simulation uncertainty from model risk.
For a decision sensitive to extreme losses, empirical estimates alone may be inadequate when the observed record is short. A parametric tail estimate can extend beyond the data, but should be tested for sensitivity to plausible model choices. Simulation can handle intricate model features, while recursion can be preferable for a well-structured discrete model. A normal approximation is often a useful benchmark rather than a default tail method.
The central exam and practice discipline is to match the method to the question and then state what the result does—and does not—measure. A precise numerical answer should not conceal limited data, an unverified tail assumption, a rough approximation, or simulation error. This method-selection framework is aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.
Check Your Understanding
For each problem, select or apply the method indicated and check whether its assumptions fit the stated setting.
- A hypothetical record of ten annual aggregate losses, in thousands, is \(0,0,10,10,20,20,30,60,100,200\). Using the empirical distribution and fractional-tail convention, find the \(80\%\) VaR and TVaR. (Answer: VaR \(=60\) thousand dollars; TVaR \(=150\) thousand dollars.)
- Annual claim count is Poisson with mean \(1.5\), and each claim is exactly \(2\) thousand dollars. Find the exact \(80\%\) aggregate VaR using the count distribution. (Answer: \(4\) thousand dollars.)
- A compound Poisson model has \(\lambda=4\), claim severity mean \(10\) thousand dollars, and severity variance \(100\) thousand-dollars-squared. Use a normal approximation to estimate the \(95\%\) aggregate VaR, using \(z_{0.95}=1.6449\). (Answer: approximately \(86.5\) thousand dollars.)
- In \(800\) independent simulations, \(64\) aggregate outcomes exceed a specified threshold. Calculate the estimated exceedance probability and approximate \(95\%\) Monte Carlo interval using \(1.96\) standard errors. (Answer: estimate \(0.0800\); interval approximately \(0.0612\) to \(0.0988\).)
- A portfolio has only a few years of aggregate observations, but a validated discrete severity model and Poisson frequency assumption are available. The decision requires accurate probabilities on a monetary grid. Which method is the most direct fit, and what limitation should still be reported? (Answer: recursive calculation; its accuracy is conditional on the assumed frequency-severity model and input parameters.)