Selecting Direct Calculation, Recursion, or Approximation

Why Method Selection Matters

A compound aggregate model can be written as $L=\sum_{i=1}^{N}Y_i$, where $N$ is the number of claims and the $Y_i$ are the corresponding payments. Finding a particular probability, such as $P(L\leq x)$, may be straightforward or computationally expensive depending on the frequency distribution, the possible payment amounts, and how much of the aggregate distribution is needed.

Four useful approaches are conditioning on $N$, direct discrete convolution, Panjer recursion, and normal approximation. They are not interchangeable shortcuts. Each has conditions that make it suitable, and each has outputs or limitations to check. The goal is to obtain the requested result with a method whose assumptions match the model and whose computational effort is proportionate to the task.

Definition: In a compound aggregate model, $N$ is a nonnegative integer-valued frequency, and $Y_1,Y_2,\ldots$ are identically distributed claim payments independent of $N$ and mutually independent. The aggregate payment is $L=\sum_{i=1}^{N}Y_i$, with $L=0$ when $N=0$. A claim payment may be zero if the model includes claims that produce no payment; if the frequency counts only positive payments, the payment distribution should instead describe those positive payments.

A Decision Framework

First identify the requested output. A single probability at a small value, a whole probability distribution up to a large threshold, and an approximate high quantile do not necessarily call for the same method. Next inspect both the frequency model and the support of the payment distribution. The following guide is a starting point, not a substitute for checking the formulas and assumptions.

Method	Often suitable when	Check before using
Condition on $N$	Only a few claim counts contribute to the requested result, or the frequency has small bounded support	Include all possible counts that can affect the target; calculate the conditional sum distribution correctly
Discrete convolution	Payments lie on a small lattice and the needed aggregate range is short	Convolve the payment distribution the correct number of times and retain the frequency probabilities
Panjer recursion	The frequency is in a supported Panjer class and payments have been represented on a lattice	Use the appropriate class parameters, severity probabilities, starting probability, and lattice scale
Normal approximation	The aggregate is sufficiently spread out for an approximation and a rough or approximate distributional result is acceptable	Verify finite moments, calculate the compound variance correctly, and assess skewness, discreteness, and tail sensitivity

Conditioning is conceptually direct: sum the conditional answer over the possible values of $N$. Discrete convolution explicitly constructs sums of payments. Panjer recursion can be more efficient when the frequency and lattice meet its requirements and many aggregate probabilities are needed. A normal approximation can be economical for a large aggregate calculation, but it does not produce exact probabilities and may be poor in a skewed or sparse distribution.

Specify the target.
Decide whether the task asks for one atom, a cumulative probability, many probabilities, or an approximation to a tail or quantile.

Inspect frequency and support.
Check whether the claim count is bounded or small, belongs to a Panjer class, or is large enough to make an approximation worth considering. Determine whether payments lie on a manageable lattice.

Match method to output.
Use an exact method when feasible and required. Avoid computing an entire distribution if conditioning gives the requested probability in a few terms; avoid a lengthy direct calculation when a suitable recursion is available.

Check the result.
Verify probability mass, units, moment assumptions, and any approximation correction. If practical, compare two methods on a small part of the calculation.

Moment Checks for Aggregate Calculations

A normal approximation requires an aggregate mean and variance. These should be derived under explicit independence and moment assumptions rather than inferred from the severity moments alone.

Theorem: Suppose the $Y_i$ are identically distributed, mutually independent, independent of $N$, and nonnegative. If $E[N]$ and $\operatorname{Var}(N)$ are finite and $E[Y_1^2]$ is finite, then $$ E[L]=E[N]E[Y_1] $$ and $$ \operatorname{Var}(L)=E[N]\operatorname{Var}(Y_1)+\operatorname{Var}(N)\big(E[Y_1]\big)^2. $$ Equivalently, $\operatorname{Var}(L)=E[N]E[Y_1^2]+(\operatorname{Var}(N)-E[N])\big(E[Y_1]\big)^2$.

To see why, condition on the claim count. Given $N=n$, the aggregate is a sum of $n$ independent payments, so $E[L\mid N=n]=nE[Y_1]$ and $\operatorname{Var}(L\mid N=n)=n\operatorname{Var}(Y_1)$. The law of total expectation gives $E[L]=E[N]E[Y_1]$. The law of total variance then gives

\begin{aligned} \operatorname{Var}(L) &=E[\operatorname{Var}(L\mid N)]+\operatorname{Var}(E[L\mid N])\\ &=E[N]\operatorname{Var}(Y_1)+\operatorname{Var}(N)\big(E[Y_1]\big)^2. \end{aligned}

The stated moment conditions matter: finite severity moments alone do not ensure a finite aggregate variance. For a Poisson frequency with parameter $\lambda$, $E[N]=\operatorname{Var}(N)=\lambda$, so the formulas simplify to $E[L]=\lambda E[Y_1]$ and $\operatorname{Var}(L)=\lambda E[Y_1^2]$. This Poisson simplification should not be used for a different frequency model without deriving its moments.

Exact Calculation: Conditioning and Panjer Recursion

For discrete payments on a lattice, conditioning and recursion can both give exact results. For a frequency distribution with probability function $p_n=P(N=n)$, conditioning gives

P(L=k)=\sum_{n=0}^{\infty}P(N=n)P(Y_1+\cdots+Y_n=k\mid N=n),

where the conditional sum for $n=0$ is zero. If only a few terms can contribute to the target, this expression is often the clearest calculation. If the frequency probabilities satisfy $p_n=(a+b/n)p_{n-1}$ for $n\geq1$, the frequency is in a Panjer class. With lattice payment probabilities $f_j=P(Y_1=j)$, aggregate probabilities $g_k=P(L=k)$, and a valid starting value $g_0$, the recursion for $k\geq1$ is

g_k=\frac{1}{1-af_0}\sum_{j=1}^{k}\left(a+\frac{bj}{k}\right)f_jg_{k-j},

provided $1-af_0\ne0$ and the model satisfies the assumptions of the recursion. When payments are strictly positive, $f_0=0$. If amounts are multiples of a monetary unit $h$, apply the recursion to the lattice index, not to dollar values as though each dollar were a possible payment. The method is attractive when several probabilities are required; for a single probability with only a few relevant claim counts, conditioning may involve less work.

Worked Example: Choose Recursion for a Poisson Aggregate Atom

A hypothetical portfolio has $N\sim\operatorname{Poisson}(2)$. Given a claim, its payment $Y$ is $1 or $2, each with probability $0.5$, independently across claims and of $N$. Find $P(L=\$2)$. Use dollars as the lattice unit.

Method. The frequency is Poisson, a Panjer class with $a=0$ and $b=2$, and the positive severity has a small integer lattice. Use recursion for the aggregate probabilities needed, then verify by conditioning.

The severity probabilities are $f_1=f_2=0.5$, with all other $f_j=0$. Since the severity is positive, $g_0=P(N=0)=e^{-2}=0.1353353$. For $k=1$,

g_1=\frac{2}{1}\left(1\cdot 0.5\,g_0\right)=g_0=0.1353353.

For $k=2$, both severity amounts can contribute:

\begin{aligned} g_2 &=\frac{2}{2}\left(1\cdot0.5\,g_1+2\cdot0.5\,g_0\right)\\ &=0.5(0.1353353)+1(0.1353353)\\ &=1.5e^{-2}=0.2030029. \end{aligned}

Thus $P(L=\$2)=0.2030$, to four decimal places. To check, conditioning on $N$, the only contributions are $N=1$ with a $2 payment and $N=2$ with two $1 payments:

P(L=2)=\left(2e^{-2}\right)(0.5)+\left(2e^{-2}\right)(0.25) =e^{-2}+0.5e^{-2}=1.5e^{-2}=0.2030.

The check agrees with the recursion. A common error is to include only the one-claim route to $2 and omit the two-claim route.

When Direct Convolution Is Preferable

For a small finite support and a small number of claims, direct convolution makes the possible sums visible and is easy to audit. For random $N$, the conditional distribution given each count must still be weighted by $P(N=n)$. This is a natural choice when the count has small bounded support or only a short range of counts contributes to the requested result.

Worked Example: Convolve a Small Payment Distribution

A hypothetical group has $N\sim\operatorname{Binomial}(2,0.5)$. Each payment is independently $1 or $3, each with probability $0.5$, and payments are independent of $N$. Find $P(L\leq\$3)$.

Method. Condition on the small, bounded claim count and convolve the two-point payment distribution when $N=2$. For $N=0$, $L=0$, so the contribution is $P(N=0)=0.25$. For $N=1$, the payment is at most $3, so the contribution is $P(N=1)=0.5$.

For $N=2$, the possible sums are $2, $4, and $6. The sum is $2 only when both payments are $1, with conditional probability $0.5^2=0.25$. It is therefore the only two-claim outcome at most $3. Since $P(N=2)=0.25$, that count contributes $0.25(0.25)=0.0625$. Combining the count cases,

P(L\leq3)=0.25+0.5+0.0625=0.8125.

The answer is $0.8125$. As a mass check, the aggregate probabilities at $0, $1, $2, $3, $4, and $6 are respectively $0.25, 0.25, 0.0625, 0.25, 0.125,$ and $0.0625$; they sum to 1. The $3 mass comes from one claim with payment $3. Treating $N=2$ as certain would discard the substantial zero- and one-claim contributions.

When a Normal Approximation Is Reasonable

For a compound aggregate with finite mean and variance, a normal approximation uses the distribution $N(E[L],\operatorname{Var}(L))$. It is an approximation, not an exact method. Its usefulness depends on the aggregate’s shape as well as its moments: a large expected claim count often helps, but substantial skewness, a large atom at zero, or a tail probability driven by rare large payments can undermine accuracy.

When aggregate payments lie on a lattice with spacing $h$, a continuity correction can align a normal probability with the discrete event. For example, if $x$ is a lattice value, approximate $P(L\geq x)$ using a normal upper tail beginning at $x-h/2$. This adjustment does not make the normal model exact. It is not a general correction for non-lattice payments.

Worked Example: Approximate a Large Poisson Aggregate Tail

A hypothetical portfolio has $N\sim\operatorname{Poisson}(25)$. Each independent payment is $2 or $4 with probability $0.5$ each, independently of $N$. Approximate $P(L\geq\$105)$ using a normal approximation with a continuity correction. Give the probability to four decimal places.

Method. Calculate the aggregate moments, then use the spacing of the aggregate lattice. Here every payment is an even number of dollars, so $L$ lies on a lattice with $h=\$2$. The event $L\geq105$ is equivalent to $L\geq106$; its continuity-corrected boundary is $105.

For one payment, $E[Y]=0.5(2)+0.5(4)=3$ dollars and $E[Y^2]=0.5(2^2)+0.5(4^2)=10$ dollars squared. The Poisson aggregate formulas give

E[L]=25(3)=75,\qquad \operatorname{Var}(L)=25(10)=250,\qquad \sigma_L=\sqrt{250}=15.8114.

The corrected standardized value is

z=\frac{105-75}{\sqrt{250}}=1.8974.

Thus the approximate probability is $1-\Phi(1.8974)\approx0.0289$, to four decimal places. This is explicitly an approximation, not an exact tail probability.

Common mistake and correction. Using $\operatorname{Var}(Y)=10-3^2=1$ in place of $E[Y^2]=10$ would give the incorrect Poisson aggregate variance $25$ and standard deviation $5$. For Poisson frequency, the aggregate variance is $25E[Y^2]=250$, not $25\operatorname{Var}(Y)$. The resulting incorrect standardized value would be $30/5=6$, producing a misleadingly tiny upper-tail estimate.

Common Method-Selection Traps

Choosing a method before identifying the output. A short conditioning sum may be simpler than generating an entire recursive distribution to answer one probability.
Applying Panjer recursion to an unsuitable input. Check the frequency class, lattice, and starting probability. A continuous severity must be represented on a lattice before a discrete recursion can be applied.
Confusing conditional and unconditional probabilities. A probability calculated given $N=n$ must be multiplied by $P(N=n)$ before it contributes to an unconditional aggregate probability.
Using severity variance alone for a Poisson aggregate. The Poisson aggregate variance is $\lambda E[Y^2]$, which includes variability in claim count as well as payment amounts.
Reporting an approximation as exact. A continuity correction improves the alignment of a normal approximation with a lattice event, but it does not establish that the aggregate distribution is close to normal.

Key takeaway: Choose a method by matching the frequency structure, payment support, requested output, and required accuracy. Exact conditioning or convolution is often clearest for small problems; Panjer recursion is useful for suitable discrete models requiring multiple probabilities; normal approximation is efficient only when its assumptions and diagnostics are acceptable.

Check Your Understanding

For each problem, identify the calculation method and check that the answer matches the stated model.

A hypothetical portfolio has $N\sim\operatorname{Poisson}(1)$; each payment is $1 or $2 with probability $0.5$ each. Find $P(L=\$2)$ exactly. (Answer: $0.625e^{-1}\approx0.2299$)
A group has exactly two independent payments, each $1 or $3 with probability $0.5$ each. Find the probability their total is at most $4. (Answer: $0.75$)
A hypothetical portfolio has $N\sim\operatorname{Binomial}(3,0.25)$, and each claim payment is a fixed $2. Find $P(L\geq\$4)$. (Answer: $0.15625$)
A hypothetical portfolio has $N\sim\operatorname{Poisson}(25)$ and independent payments of $2 or $4 with equal probability. Using a normal approximation and the lattice continuity correction, estimate $P(L\geq\$105)$. (Answer: approximately $0.0289$)
A frequency model has finite $E[N]$, but infinite $\operatorname{Var}(N)$; every payment is a fixed $1. Is the finite-variance aggregate formula justified? (Answer: No; $E[Y^2]$ is finite, but the frequency variance condition fails and $\operatorname{Var}(L)$ is infinite.)

Method	Often suitable when	Check before using
Condition on \(N\)	Only a few claim counts contribute to the requested result, or the frequency has small bounded support	Include all possible counts that can affect the target; calculate the conditional sum distribution correctly
Discrete convolution	Payments lie on a small lattice and the needed aggregate range is short	Convolve the payment distribution the correct number of times and retain the frequency probabilities
Panjer recursion	The frequency is in a supported Panjer class and payments have been represented on a lattice	Use the appropriate class parameters, severity probabilities, starting probability, and lattice scale
Normal approximation	The aggregate is sufficiently spread out for an approximation and a rough or approximate distributional result is acceptable	Verify finite moments, calculate the compound variance correctly, and assess skewness, discreteness, and tail sensitivity

What You'll Learn