Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Aggregate Loss and Compound Risk Models

Abstract. Aggregate insurance loss is a random total generated by the joint operation of the number of claims, their monetary amounts, the insurance contract, and the dependence structure of the portfolio. This monograph develops the mathematical foundations of aggregate loss distributions, beginning with the collective random-sum model and its individual-risk counterpart. It derives probability-generating and transform identities, expectations, variances and higher-order cumulants under stated independence and integrability conditions. Exact convolution and recursive computation are compared with Fourier methods, simulation, analytical approximation and heavy-tail asymptotics. Contractual transformations are incorporated at the claim and annual aggregate levels, including deductibles, per-occurrence caps, aggregate attachment, exhaustion, and stop-loss premiums. Reproducible numerical experiments illustrate the practical consequences of severity granularity, count overdispersion, aggregation dependence, and coverage order. The analysis distinguishes a computationally exact distribution conditional on a specified model from an empirically credible prediction, and shows how the same expected aggregate loss can coexist with materially different no-loss probabilities, dispersion, tail quantiles and ceded payments. The principal contribution is an integrated critical synthesis of established actuarial results, transparent derivations, and independently recomputed synthetic tests. No new actuarial theory, doctrine, or statistical estimator is claimed.

Keywords: Aggregate loss; collective risk model; individual risk model; compound Poisson distribution; compound negative binomial distribution; claim frequency; loss severity; probability-generating function; Panjer recursion; convolution; Fourier inversion; Tweedie model; stop-loss; aggregate limit; dependence; heavy tails; numerical validation.

Introduction

An insurer's annual financial result depends on an aggregate amount, while most observable insurance data consist of separate claims, insured objects, exposure periods and contractual payment rules. The research problem is the transformation from those microscopic observations to a probability law for a portfolio total. A conditional expectation is only one property of that law: pricing a stop-loss layer, identifying an annual-cap exhaustion probability, or assessing a remote stress scenario requires substantially more distributional information. The established actuarial synthesis in Loss Models: From Data to Decisions develops frequency, severity and aggregate models as connected decision problems. The mathematical and computational investigation here focuses on their aggregate-loss boundary.

The foundational collective-risk research by Panjer (1981) treats the distribution of a random sum of claim amounts and provides an important computational simplification for a special class of claim counts. Sundt and Jewell (1981) characterize and extend the recursive class. In a different portfolio representation, De Pril (1986) computes the aggregate distribution for individual life risks. Their results illuminate two distinct descriptions of insurance uncertainty: a randomly sized collection of identically modeled claims and a fixed roster of heterogeneous insurance contracts.

The object of the present analysis is a nonnegative monetary aggregate over a stated observation period, with extensions to marked events, grouped portfolios and layered contractual payments. Monetary units are arbitrary in the synthetic experiments. The observation horizon, exposure base, valuation convention and expense treatment must be fixed before a fitted frequency–severity pair becomes a meaningful insured-payment distribution. A reported claim, an occurrence, a claimant and a paid transaction can have different counting laws; the model's event unit is therefore part of its substantive scientific specification.

A rigorous aggregate analysis proceeds in a defined order. First, identify the random objects and their independence or dependence relations. Second, apply any contract-specific transformation to the correct level of loss. Third, derive moments and a distributional representation. Fourth, select an algorithm with explicit approximation error or a finite exact benchmark. Fifth, assess whether the observed data support the chosen tail, dependency and temporal assumptions. Heckman and Meyers (1983) already recognized parameter uncertainty as a major practical obstacle in collective-risk calculation; that concern remains central whenever numerical precision is mistaken for statistical certainty.

Scope, notation, and inferential discipline

Let (N) be a nonnegative integer-valued claim count and (X_1,X_2,\ldots) be nonnegative severities. Write (S=\sum_{i=1}^{N}X_i), with (S=0) when (N=0). Under the standard collective assumptions, the severities are identically distributed, mutually independent and independent of (N). These assumptions are declared wherever used; their relaxation changes the distribution. A random variable (Y) will denote an insurer's payment after specified contract terms, which can differ sharply from ground-up (S). Expectations, probabilities and quantiles refer to an explicitly stated model rather than to undisclosed observed insurance records.

The research uses derivations from probability theory, published actuarial scholarship, a comparison of exact methods with approximations, and calculations on synthetic distributions. Publication metadata and specific claims are checked against journal, scholarly-publisher and professional-society records. No private insurer data, prospective climate or loss scenarios, jurisdiction-specific coverage rulings, or external refereeing are represented as part of the evidence.

Aggregate Loss Processes

The random sum and its conditional structure

The classical aggregate model has two levels of randomness. Conditioning on (N=n) gives an (n)-fold convolution of the severity law, and averaging over the count distribution gives the aggregate law. If (p_n=P(N=n)) and (F_X^{*n}) is the (n)-fold convolution of (F_X), including a unit mass at zero for (n=0), then

[ F_S(s)=\sum_{n=0}^{\infty}p_nF_X^{*n}(s),\qquad s\geq 0. ]

This identity is a direct application of the law of total probability. It remains valid for arbitrary severity laws under the iid-and-independence assumptions, with no requirement of an existing density. Its monetary interpretation depends on whether (X) is a ground-up event loss, a reported amount or a contract-adjusted payment. A pure count model with several latent occurrences per paid claim cannot be inserted into a paid-severity distribution without accounting for the reporting and payment mechanisms.

For a nonnegative integer-valued severity, define (f_j=P(X=j)), (g_s=P(S=s)), and (G_X(z)=E[z^X]) for (|z|\leq1). Conditioning yields the compact identity

[ G_S(z)=G_N(G_X(z)). ]

For continuous severity, the Laplace transform (L_X(t)=E[e^{-tX}]) exists for every (t\geq0), and (L_S(t)=G_N(L_X(t))). By contrast, a positive-exponent moment-generating function may fail for heavy-tailed severities. This distinction matters: numerical transform methods may remain available even when some theoretical moments do not exist.

Poisson aggregation and random event processes

When (N\sim\mathrm{Poisson}(\lambda)), for (\lambda\geq0), the generating function is (G_N(z)=\exp{\lambda(z-1)}), so

[ G_S(z)=\exp{\lambda(G_X(z)-1)}, \qquad L_S(t)=\exp{\lambda(L_X(t)-1)}. ]

The point probability of zero aggregate loss is (P(S=0)=\exp{\lambda(f_0-1)}). The familiar expression (e^{-\lambda}) requires strictly positive severity, meaning (f_0=0). If the count includes zero-payment events, a Poisson count of positive-payment claims results from thinning, and the two count intensities should be explicitly distinguished. The integral (\lambda) can represent accumulated exposure-dependent intensity over an observation interval when the arrival process is nonhomogeneous and the marked claims satisfy the necessary independent-mark conditions.

A compound Poisson distribution is infinitely divisible: independent Poisson components with compatible mark structures may be combined by adding their exponents. The construction clarifies how common event shocks must be represented. If one event produces several correlated claims or affects several portfolios, decomposing the event into supposedly independent claim counts can erase covariance and underestimate the uncertainty of the aggregate. A compound model may remain mathematically convenient while the simple one-event/one-loss mark interpretation becomes inadequate.

Time aggregation and its limits

A stationary compound Poisson process has independent increments and constant frequency intensity. Those properties give a consistent way to move between monthly and annual horizons when severity laws and claim definitions are unchanged. In insurance experience, exposures, thresholds, inflation, claims development and catastrophe regimes can change within a year. The annual process should then be expressed through interval-specific or stochastic intensity and mark laws. Simply multiplying a quarterly fitted mean by four preserves a first moment under restrictive conditions but need not preserve an annual tail distribution or a risk-adjusted contract payment.

Temporal escalation is especially important where incurred claim amounts are settled in later monetary units. A homogeneous model calibrated on unadjusted nominal losses can confound frequency, severity inflation and development. This is a scientific input-design problem rather than an error in convolution mathematics. The aggregate distribution must be indexed by the measurement basis and valuation date, and the relevant loss definition must remain consistent across frequency estimation, severity fitting and payment rules.

Heavy tails and the claim-size mechanism

The tail of (S) depends fundamentally on the right tail of (X). The treatment in Modelling Extremal Events demonstrates the relevance of extreme-value and heavy-tail methods in insurance. When severity is subexponential and (N) is independent with suitable light-tail conditions and finite positive mean, a standard large-threshold relation is (P(S>x)\sim E[N]P(X>x)). The required conditions are material; there are compound models with dependent marks, heavy-tailed counts or different asymptotic regimes in which this expression cannot simply be asserted. Work on large claims and ruin probabilities illustrates the financial consequences of heavy-claim behavior under more elaborate surplus models.

The qualitative conclusion is that rare severe losses may control the aggregate right tail even when routine small claims control the observed frequency distribution. Aggregate caps can temper a contractual payment distribution, but they do not erase ground-up severity uncertainty or the need to model the attachment probability. Distributional adequacy is therefore evaluated at the loss layers relevant to the actual research question.

Collective and Individual Models

The collective representation

A collective model represents a period's total as a random count of exchangeable severity observations. It is parsimonious because one count law and one mark law can describe a sizable book, sometimes after grouping by exposure. The model can naturally accommodate the possibility of more than one claim per insured unit. It is often suitable for property and casualty portfolios measured at an event or reported-claim level, conditional on carefully selected rating classes.

The count can be Poisson, negative binomial, binomial, a mixture of Poisson laws, or a more general renewal or clustered process. These are substantive hypotheses. A Poisson model has equal mean and variance for the count. An overdispersed count can arise from unobserved heterogeneity in intensity, common environmental conditions, or clustering. A negative binomial may fit observed variance yet still fail to represent the causal dependence structure required for multivariate capital aggregation. Estimation of an aggregate distribution includes the task of identifying which uncertainty is explained by covariates and which remains in residual dependence or mixing.

The individual representation

An individual risk model begins with identified contracts. Let (I_j) be an indicator of a covered event for policy (j), with probability (q_j), and let (B_j) denote a contractual benefit amount. For fixed benefits and independent indicators,

[ S_{\mathrm{ind}}=\sum_{j=1}^m I_jB_j,\qquad G_{S_{\mathrm{ind}}}(z)=\prod_{j=1}^m(1-q_j+q_jz^{B_j}) ]

when benefits are nonnegative integers in a consistent monetary unit. The expectation is (\sum_jq_jB_j) and the variance is (\sum_jq_j(1-q_j)B_j^2). The product representation explicitly retains heterogeneity. It permits exact finite polynomial multiplication and risk-by-risk sensitivity calculations, but it depends on independent policy indicators and the assumption of at most one modeled benefit occurrence for each indicator. De Pril's individual-life recursion addresses exact aggregate calculation for the individual life framework and provides a historical counterpart to collective recursions.

For independent random benefits (Z_j) payable only when (I_j=1), the factor generalizes to (1-q_j+q_jG_{Z_j}(z)) where the independence assumptions justify it. Common mortality shocks, group catastrophes, fraud networks or legal changes can break the product structure. Dependence requires a joint generating object or a conditional model, not an unqualified multiplication of marginal risk probabilities.

When a collective approximation is defensible

An individual portfolio with small event probabilities and comparable single-event payouts can sometimes be approximated by a Poisson count. For heterogeneous event probabilities (q_j), the expected number is (\lambda=\sum_jq_j); a classical small-probability approximation is supported by a bound controlled by (\sum_jq_j^2) for the count law. The approximation can be poor when a few contracts carry substantial probabilities or unusually large benefits. Even an accurate count approximation does not automatically imply a good approximation for every high-threshold monetary tail, because severity weights may concentrate on those exceptional contracts.

A large number of independent identically distributed policies does not itself establish the appropriateness of a single iid severity law if policy amounts, deductibles, policy periods and reporting thresholds differ. A meaningful collective approximation groups the risk units according to statistically and contractually supported commonality. The practical comparison is therefore between how much heterogeneity is preserved and how much statistical estimability or numerical simplicity is gained.

Exact individual example

Consider three independent Bernoulli contracts with event probabilities (0.1,0.2,0.3) and benefits (1,2,4). Their generating function is ((0.9+0.1z)(0.8+0.2z^2)(0.7+0.3z^4)). The mean payment is (0.1+0.4+1.2=1.7) and the variance is (0.09+0.64+3.36=4.09). The no-loss probability is (0.9\times0.8\times0.7=0.504). A total strictly exceeding four occurs exactly when the third event occurs jointly with at least one of the first two, giving (P(S_{\mathrm{ind}}>4)=0.3[1-(0.9)(0.8)]=0.084). These identities are exact and independently enumerated in the numeric QA ledger.

Matching this model to a Poisson distribution with mean number of claims (0.6) and a convenient average severity may preserve the aggregate mean, but it need not reproduce the mass at zero or the tail at four. The heterogeneity of the fourth-unit benefit makes the approximation's relevant threshold explicit. The example also warns against equating a portfolio's number of insurance contracts with a stochastic count of claims.

Shared-shock and multivariate aggregation

Let (K_0,K_1,K_2) be independent Poisson variables with means (0.5,1,1). Define two portfolio counts by (N_A=K_0+K_1) and (N_B=K_0+K_2). Both marginals have mean (1.5), while (\operatorname{Cov}(N_A,N_B)=\operatorname{Var}(K_0)=0.5). If each A event costs one monetary unit and each B event costs two, the portfolio losses have covariance one. Treating the marginals as independent would omit the positive cross term (2\operatorname{Cov}(S_A,S_B)=2) in the variance of their sum. The multivariate recursive direction explored by Sundt (1999) addresses models in which claim events generate loss vectors rather than isolated scalar severities.

The example distinguishes aggregation within a portfolio from aggregation across correlated portfolios. A univariate compound distribution for each line is insufficient for a joint distribution of their sum unless a cross-line dependence assumption is supplied and tested.

Moments of Compound Sums

Conditioning and the first two moments

Assume (E[N]<\infty), (E[X]<\infty), iid severities and independence of (N). The conditional expectation (E[S\mid N]=N\mu_X) gives the central identity

[ E[S]=E[N]\mu_X. ]

When the relevant second moments exist, conditional variance gives (\operatorname{Var}(S\mid N)=N\sigma_X^2), so the law of total variance yields

[ \operatorname{Var}(S)=E[N]\sigma_X^2+\operatorname{Var}(N)\mu_X^2. ]

This decomposition separates severity variation conditional on claim count from count variation in the number of severity terms. It is computationally simple and highly diagnostic. If a fitted model understates count variance, even perfectly estimated severity moments cannot recover the correct aggregate variance. The common shortcut (\operatorname{Var}(S)=\lambda E[X^2]) holds when (N) is Poisson with mean (\lambda), since (\operatorname{Var}(N)=\lambda).

For heterogeneous independent individual risks, the correct variance instead adds risk-specific contributions, or adds covariance terms when independence fails. The two variance representations describe different probability constructions. Their superficial similarity is not evidence that a particular portfolio satisfies both sets of assumptions.

Higher derivatives and cumulants

If suitable moments exist, derivatives of (G_N(G_X(z))) provide factorial moments for integer variables. For a continuous compound Poisson model with a moment-generating function near zero, the cumulant-generating function is

[ K_S(t)=\lambda(M_X(t)-1),\qquad\kappa_r(S)=\lambda E[X^r] ]

for each finite raw severity moment of order (r) in the domain where differentiation is justified. Hence the third cumulant is (\lambda E[X^3]) and the standardized skewness is (\lambda E[X^3]/(\lambda E[X^2])^{3/2}). The fourth cumulant involves (E[X^4]), and the fourth central moment additionally contains three times the squared variance. Claims that skewness or kurtosis are finite require explicit assumptions: a heavy-tailed severity distribution may have an infinite third or fourth moment even when aggregate expectation and variance exist.

At a formal level, the first three cumulants of a general independent random sum can be written using the cumulants of (N) and of (X):

[ \kappa_1(S)=\kappa_1(N)\kappa_1(X), ] [ \kappa_2(S)=\kappa_1(N)\kappa_2(X)+\kappa_2(N)\kappa_1(X)^2, ] [ \kappa_3(S)=\kappa_1(N)\kappa_3(X)+3\kappa_2(N)\kappa_1(X)\kappa_2(X)+\kappa_3(N)\kappa_1(X)^3. ]

These identities follow from composing cumulant generating functions and differentiating with the chain rule. They also clarify that matching mean and variance of frequency and severity distributions does not match aggregate skewness. Tail-sensitive treaty prices can therefore differ even between models having apparently identical two-moment summaries.

Mixing intensity and count overdispersion

Consider a doubly stochastic model in which (N\mid\Lambda\sim\mathrm{Poisson}(\Lambda)), with random nonnegative intensity (\Lambda), and the severity sequence independent of both (N) and (\Lambda). Then (E[N]=E[\Lambda]) and (\operatorname{Var}(N)=E[\Lambda]+\operatorname{Var}(\Lambda)). The variance of aggregate loss becomes

[ \operatorname{Var}(S)=E[\Lambda]E[X^2]+\operatorname{Var}(\Lambda)\mu_X^2. ]

The second term is the additional uncertainty due to mixing. If (\Lambda\sim\mathrm{Gamma}(r,\theta)) with shape (r) and scale (\theta), the unconditional count is negative binomial. The count PGF is (G_N(z)=(1+\theta(1-z))^{-r}), and (P(N=0)=(1+\theta)^{-r}). This is one route from latent heterogeneity to count overdispersion; the same observed negative-binomial margins may also reflect clustered processes with different multivariate implications.

Choose (r=2), (\theta=1), and severity (X) taking values one with probability (0.75) and three with probability (0.25). Then (E[N]=2), (\operatorname{Var}(N)=4), (E[X]=1.5), (\operatorname{Var}(X)=0.75), (E[S]=3), and (\operatorname{Var}(S)=2(0.75)+4(2.25)=10.5). With a nonmixed Poisson count of the same mean, the aggregate variance is (2E[X^2]=6). Both models have expected aggregate loss three; the gamma-mixed model has (P(S=0)=P(N=0)=0.25), versus (e^{-2}\approx0.135335) under Poisson. Greater variance does not, by itself, imply every upper quantile is larger, so a tail-comparison conclusion requires a separate calculation.

Dependence between frequency and severity

The usual moment formulas cease to apply unchanged if the severity distribution depends on the count. Let (\mu_n=E[S\mid N=n]). Then (E[S]=\sum_np_n\mu_n). If individual severities are conditionally iid given (N=n) with conditional mean (m_n), this becomes (E[S]=E[Nm_N]); it equals (E[N]E[X]) only under further compatibility conditions. A common catastrophe environment may increase both claim frequency and average severity, producing an upward dependence term that is absent under unconditional independence.

An empirically responsible compound model should test whether conditional severity parameters vary with event multiplicity, geographic concentration, catastrophe status or claims development. Smyth and Jørgensen (2002) demonstrate that modeling dispersion as well as mean is relevant when combining claim counts and insurance costs. This is a fitting problem within familiar compound families; it does not restore independence merely by expressing results in generalized-linear-model notation.

Convolutions and Recursion

Exact distributional convolution

For integer severities, set (f^{0}_0=1) and recursively define (f^{(n+1)}s=\sum{j=0}^{s}f_jf^{*n}{s-j}). Direct mixture convolution then computes (g_s=\sum{n\geq0}p_nf^{*n}_s). For finite-support claim counts and severity, this is a fully finite exact algorithm in exact arithmetic. For infinite count support, truncation at some maximum (n) introduces an error bounded above by (P(N>n)) in a probability calculation, before considering machine rounding. Severity discretization and truncation introduce additional errors if the original monetary model is continuous.

The direct method is conceptually useful as a benchmark but can repeat expensive convolutions. Gerber (1982) and Heckman and Meyers (1983) give foundational actuarial discussions of calculating aggregate distributions and stop-loss quantities. Their shared computational question is how to obtain enough of the distribution to evaluate the contractually relevant tail without surrendering the underlying count–severity specification.

The Panjer counting class

For a discrete count distribution satisfying

[ p_n=\left(a+\frac{b}{n}\right)p_{n-1},\qquad n\geq1, ]

and a nonnegative integer-valued severity law (f_j), the aggregate distribution can be obtained by the Panjer recursion:

[ g_0=G_N(f_0),\qquad g_s=\frac{1}{1-af_0}\sum_{j=1}^{s}\left(a+\frac{bj}{s}\right)f_jg_{s-j},\qquad s\geq1. ]

Provided the relevant denominator and count-law conditions are satisfied, this recursion computes one aggregate probability at a time. In the familiar Poisson case, (a=0), (b=\lambda); in the negative-binomial parametrization (G_N(z)=[(1-q)/(1-qz)]^r), (a=q), (b=(r-1)q). A binomial count with fixed maximum (m) and success probability (q) has (a=-q/(1-q)) and (b=(m+1)q/(1-q)). The generating functions and parameter domains establish the legitimacy of these values; inserting an arbitrary empirical count law into this recurrence would be a model error.

The initial value (g_0=G_N(f_0)) is often overlooked when some severity mass is at zero. Setting (g_0=p_0) is justified only if every counted severity is strictly positive. The recursion is usually applied after severity discretization onto a nonnegative monetary grid; its output is exact for that discretized model when arithmetic is exact. It is not automatically an exact numerical result for an underlying continuous loss distribution.

The Sundt–Jewell characterization shows why the count recurrence is a structural restriction and develops generalizations. Computational method selection must follow the count model, rather than allowing the existence of a convenient algorithm to dictate a scientifically unsupported frequency distribution.

Derivation in the Poisson case

With (f_0=0), the compound Poisson PGF satisfies (G'_S(z)=\lambda G'_X(z)G_S(z)). Equating coefficients of (z^{s-1}) gives

[ sg_s=\lambda\sum_{j=1}^{s}j f_jg_{s-j}, ]

hence (g_s=(\lambda/s)\sum_{j=1}^{s}j f_jg_{s-j}). If (f_0>0), the (j=0) term in the derivative vanishes, while the initial probability changes to (e^{\lambda(f_0-1)}). For this elementary derivation, the probability generating function is differentiable inside its radius of convergence, and coefficient comparison can be justified directly on the open unit disk.

Fourier inversion and numerical aliasing

For lattice severities, the characteristic function of the aggregate is (\phi_S(t)=G_N(\phi_X(t))). A discrete Fourier transform can approximate the sequence of probabilities by evaluating the transform on a finite frequency grid and performing an inverse transform. The attractive computational speed comes with a conceptual hazard: a finite discrete Fourier inversion represents circular rather than linear convolution unless adequate padding, tilting or other controls are used. Probability from losses above the chosen range may wrap into lower bins. If a result appears to assign positive probability to impossible small losses, check grid design, wraparound and transform conventions before interpreting it as a feature of the actuarial distribution.

Discrete Fourier inversion and Panjer recursion serve different roles. Recursion is often transparent for a suitable count class and a moderate target lattice. Fourier methods can be efficient for large grids and general count PGFs, but their discretization, truncation, aliasing and floating-point errors require explicit control. For both methods, reporting only an apparently precise percentile without the grid size and error-control choices fails numerical reproducibility.

A method-selection comparison

Direct convolution is valuable as a transparent finite benchmark and for individual risks with few possible benefits. Panjer recursion is a leading option for a lattice severity and a supported ((a,b))-class count. Fourier inversion is useful for broad numerical grids with transform evaluation and verified wraparound control. Monte Carlo supports models with complicated contractual rules or dependence, provided uncertainty in a rare tail estimate is explicitly quantified. Conditional analytical formulas, gamma mixtures and saddlepoint approximations can accelerate particular parameter families. Each algorithm computes or approximates a probability law only after the underlying risk model is fixed.

The criterion for selecting a computational method is the target quantity and required error tolerance, together with the type of count law, loss support, threshold location and dependence information. The distinction between actuarial modeling error and numerical error is primary: extra numerical digits cannot repair missing covariates, wrong attachment definitions or dependence structures.

Coverage and Aggregate Payments

The order of contractual transformations

An insurance payment is constructed from raw losses through contract terms. For a ground-up claim (X), an ordinary per-occurrence deductible (d) and payment limit (\ell) may yield (Z=\min((X-d)+,\ell)). In this notation ((u)+=\max(u,0)). If the frequency process counts all occurrences but only a subset produces positive payments, the derived severity distribution has a mass at zero. One may alternatively thin the count to positive-payment occurrences, provided both the frequency and conditional severity distributions are transformed consistently.

For a period total (T=\sum_{i=1}^N Z_i), an annual aggregate deductible (D) and an annual payment limit (L) generate (Y=\min((T-D)_+,L)). The order matters. In general,

[ \min\left(\left(\sum_i X_i-D\right)+,L\right) \neq\sum_i\min((X_i-d)+,\ell). ]

The right-hand expression transforms occurrences before aggregation, whereas the left-hand expression transforms the aggregate once. Contract analysis must establish exactly which limits and retentions apply to which loss unit, whether the coverage includes allocated adjustment expenses, whether reinstatements exist and whether losses can be reopened or adjusted retrospectively. These are mathematically distinguishable payment functions even when their ordinary-language labels sound similar.

Stop-loss functionals

For a nonnegative aggregate (T) and attachment (D), the expected stop-loss payment is

[ \pi(D)=E[(T-D)_+]=\int_D^\infty P(T>t),dt. ]

For nonnegative integer-valued (T) and an integer deductible (D), (\pi(D)=\sum_{k=D+1}^{\infty}P(T\geq k)). The integral representation follows from Tonelli's theorem and holds even when the expectation is infinite. If (T) has finite mean, the fundamental decomposition (E[T]=E[\min(T,D)]+\pi(D)) follows pointwise from (T=\min(T,D)+(T-D)_+).

For an aggregate layer paying at most (L),

[ E[\min((T-D)_+,L)]=\pi(D)-\pi(D+L). ]

This identity is an exact pricing relationship for the expected payment under the specified loss law. It does not add expense, risk margins, cost of capital, brokerage, taxes, credit risk, reinstatement pricing or contract interpretation. The numerical treatment of aggregate losses and excess pure premiums is a central theme in Heckman and Meyers (1983).

Aggregate deductibles and nonlinearity

The map (T\mapsto(T-D)+) is convex. Therefore uncertainty in a distribution may affect expected stop-loss payment even when aggregate means agree. Convex-order comparisons, if their assumptions are met, give a precise way to discuss this influence; a mere comparison of variances is insufficient. A second nonlinearity arises at a limit: (T\mapsto\min((T-D)+,L)) ceases to be globally convex once the layer is exhausted. A more variable loss may increase attachment probability and also increase exhaustion, with competing consequences for the layer's expected payment.

Because aggregate covers pay according to tails of period totals, the fitting target cannot be only a claim-level mean or a single overall loss ratio. The relevant fitted model must be adequate around the attachment and exhaustion regions. Gerber's stop-loss analysis shows the longstanding centrality of these functionals in aggregate distribution computation.

Frequency–severity example with a finite layer

Let (N\sim\mathrm{Poisson}(2)), and let independent severities equal one unit with probability (0.75) and three units with probability (0.25). The severity PGF is (0.75z+0.25z^3); therefore (G_S(z)=\exp[2(0.75z+0.25z^3-1)]). The aggregate mean is three and variance is six. Exact lattice recursion gives (g_0=0.135335283237), (g_1=0.203002924855), (g_2=0.152252193641), (g_3=0.143793738439), and (g_4=0.130048748735). These values can also be obtained from independent Poisson thinning: numbers of unit-one and unit-three claims are independent Poisson variables with means (1.5) and (0.5).

The no-loss probability is about 13.53 percent. The probability of a total exceeding three is (1-(g_0+g_1+g_2+g_3)=0.365615859828). For annual aggregate attachment (D=3), the stop-loss expectation equals (3g_0+2g_1+g_2=0.964263893061), because (E[S]=3) and the retained amount (\min(S,3)) can be expressed using the probabilities of totals zero, one and two. With an annual aggregate payment cap (L=2), the expected payment is (P(S\geq4)+P(S\geq5)=0.601182970922). This last value provides an independent cross-check on (\pi(3)-\pi(5)).

The example isolates a recurring economic fact: a claim-level severity mean of 1.5 and a frequency mean of two establish expected ground-up loss three, but the insurer's expected payment is determined by the nonlinear coverage function and the full aggregate distribution. The loss amount and covered amount are different stochastic objects.

Tail metrics and their interpretation

For a continuous distribution, a high loss quantile and a conditional mean beyond that quantile may summarize capital-sensitive tail outcomes. For discrete aggregates the distinction between strict exceedance, inclusive exceedance and a quantile atom must be stated. Define the lower (\alpha)-quantile (q_\alpha=\inf{s:F_S(s)\geq\alpha}). Under this definition, the Poisson example has (q_{0.95}=8) monetary units by lattice summation. The claim does not apply to all alternative quantile definitions and is a model-based result rather than an observed insurance 95th percentile.

Risk-management treatment of tail loss, dependence and capital aggregation is developed in Quantitative Risk Management. Expected shortfall and stop-loss payments answer different questions: one describes a tail risk functional at a probability threshold, whereas the other values an attachment layer at a monetary threshold. A dossier about aggregate loss must retain that conceptual separation.

Approximation and Numerical Tests

Exact finite calculations and an auditable test design

A numerical test becomes scientifically interpretable when its sample space, formula, numerical method and comparison standard are specified. The present tests use integer currency units, simple count laws and deterministic computations, so the principal values can be checked independently without insurer records. The Poisson example is evaluated by both the recursion and the independent compound-Poisson decomposition into (N_1\sim\mathrm{Poisson}(1.5)), (N_3\sim\mathrm{Poisson}(0.5)), with (S=N_1+3N_3). The individual-policy example is evaluated by eight Bernoulli states. The count-mixing example is checked analytically from the gamma-mixed generating function. The shared-shock example is checked by covariance identities.

For the Poisson model, (P(S=s)=e^{-2}\sum_{k=0}^{\lfloor s/3\rfloor}(1.5)^{s-3k}/(s-3k)!,(0.5)^k/k!). This formula supplies a computational reference independent of the Panjer recurrence. Reconciliation of the first few probabilities, moment identities, layer expectation and 95th quantile is a more credible QA record than a single unqualified Monte Carlo point estimate. For the individual model, multiplying three two-term polynomials reproduces every state probability and preserves normalization exactly in rational arithmetic.

Normal and gamma approximations

A moment-matched normal approximation for aggregate loss has mean (m=E[S]) and variance (v=\operatorname{Var}(S)). It can be useful in portfolios with high claim frequency, finite variance and sufficiently regular severity behavior, but it assigns positive probability to negative amounts and may distort the point mass at zero. In the Poisson example, (m=3) and (\sqrt v=\sqrt6\approx2.44949); the normal approximation yields (P(S\leq0)\approx\Phi(-1.224745)), a clearly inappropriate negative-or-zero mass for an atom at zero of 0.135335. Normal tail extrapolation is particularly unreliable in low-frequency, skewed or heavy-tailed portfolios.

A moment-matched gamma approximation avoids negative loss support but has no positive point mass at zero and therefore cannot exactly reproduce compound-Poisson zero inflation. For (m,v>0), its shape and scale are (m^2/v) and (v/m), respectively. In the illustrative example these are 1.5 and 2. The gamma law has a continuous density over positive amounts, while the actual compound model is supported on nonnegative integers. Moment matching alone cannot establish adequacy for deductible attachment or low-count portfolio probabilities.

Compound Poisson–gamma and the Tweedie representation

If the claim count is Poisson and positive claim severities are gamma distributed, the aggregate has a point mass at zero and a continuous positive component. This compound Poisson–gamma construction belongs to the Tweedie exponential dispersion family for power parameter (1<p<2). With gamma severity shape (\alpha>0), its Tweedie power is (p=(\alpha+2)/(\alpha+1)). The statistical treatment by Smyth and Jørgensen (2002) examines mean and dispersion in insurance claim costs; Dunn and Smyth (2005) examine numerical evaluation of Tweedie density series; Delong, Lindholm, and Wüthrich (2021) compare parametrizations and estimation costs.

The distinction between a compound Poisson–gamma probability model and a Tweedie regression fit is essential. One is a stochastic construction; the other specifies a conditional mean–variance relationship and parameter estimation scheme across exposures. Covariate effects may enter count intensity, severity shape or scale, and dispersion separately. A fitted conditional total-loss model without observed counts may provide useful premium estimates while leaving the underlying frequency–severity decomposition weakly identified.

Simulation and rare-event error

A basic Monte Carlo algorithm samples a count (N) and then independently samples (N) severity values for every replicate. It aggregates the simulated losses and applies the coverage function to each replicate. An estimated exceedance probability (\hat p=k/M) from (M) independent replicates has standard error approximately (\sqrt{p(1-p)/M}), and its relative error grows when (p) is small. In a tail with probability (10^{-4}), one million independent trials produce only about one hundred expected exceedances; the approximate relative standard error is near ten percent. This is adequate for some illustrative tests and inadequate for high-precision rare-event calibration.

Importance sampling, conditional Monte Carlo and stratification can improve tail estimation, but each method introduces an estimator and requires its own validity and variance analysis. Simulation of a common-shock process must sample the shared shock once per period and propagate it to all affected lines. Independently simulating already-correlated marginals would be an incorrect model, even if the simulation code were numerically flawless.

Heavy-tail approximations and finite-portfolio restraint

For large thresholds under suitable subexponential severity and light-tail count conditions, the one-big-claim approximation (P(S>x)\approx E[N]P(X>x)) can be informative. It is an asymptotic statement as (x\to\infty), not a finite-threshold accuracy guarantee. At the actual attachment points used for treaty pricing, multiple moderate losses may still contribute substantially. Embrechts, Klüppelberg, and Mikosch (1997) supply the mathematical framework for such tail analysis. The analysis should test asymptotic approximations against exact recursion or simulation over a relevant threshold range rather than substitute asymptotics for validation.

A finite cap changes the distribution of the insurer's payment even when the underlying ground-up loss is heavy-tailed. The loss beyond the cap remains relevant to reinsurers, policyholders, retrocessionaires or the insurer's uninsured economic exposures depending on contract structure. At each stage, the risk variable and its legal and economic owner must be identified before a quantitative tail conclusion is translated into a solvency or pricing statement.

Model uncertainty and validation hierarchy

Parameter uncertainty, frequency–severity misspecification, exposure drift, claims development and latent dependence can dominate an apparently small numerical error. One may represent parameter uncertainty by (\Theta) and consider the posterior or scenario mixture (P(S\leq s)=E_\Theta[P(S\leq s\mid\Theta)]) where justified by the inferential framework. In general, (E_\Theta[q_\alpha(S\mid\Theta)]) is not the unconditional quantile of that mixture, and an average of conditional risk measures cannot be silently substituted for a risk measure of the predictive distribution.

An effective validation program has several layers. It checks counts per exposure and their dispersion, claim-size support and truncation, zero reporting, conditional relationships, aggregate totals, claim-size thresholds, geographic and temporal concentration, and simulated or computed payments under the actual contract. An independent implementation or exact benchmark tests code logic. Out-of-sample or forward validation tests predictive calibration where data support it. Stress tests investigate model fragility outside the estimation region; they are identified as scenarios and do not acquire empirical probabilities by being placed in a simulation engine.

The problem of aggregating dependence-sensitive and extreme outcomes is analyzed in McNeil, Frey, and Embrechts (2015). The distinction between risk modeling and operational decision-making is deliberate: a distributional result can inform but does not alone establish an insurer's contractual liability, statutory capital, reserves or reinsurance price. Those decisions require further jurisdiction-specific, accounting and governance premises outside this mathematical dossier.

Reproducibility table

Synthetic case Quantity Independently verified result
Poisson frequency (\lambda=2), severity (P(X=1)=0.75), (P(X=3)=0.25) (E[S]) 3
Same (\operatorname{Var}(S)) 6
Same (P(S=0)) 0.135335283237
Same (P(S>3)) 0.365615859828
Same (E[(S-3)_+]) 0.964263893061
Same (E[\min((S-3)_+,2)]) 0.601182970922
Same Lower (0.95)-quantile 8
Three independent Bernoulli policies with (q=(0.1,0.2,0.3)), (B=(1,2,4)) (P(S=0)) 0.504
Same (P(S>4)) 0.084
Same (E[S],\operatorname{Var}(S)) 1.7; 4.09
Gamma-mixed Poisson count with shape 2, scale 1; same severity (E[S],\operatorname{Var}(S),P(S=0)) 3; 10.5; 0.25
Shared Poisson count with common intensity 0.5 and loss multipliers 1 and 2 Cross-portfolio covariance 1

All figures in the table come from specified synthetic models. They do not estimate any identifiable insurer, jurisdiction or industry portfolio. A result labelled exact refers to a mathematically specified discrete model; its numerical decimal display is rounded. The formulas can be reproduced with elementary probability calculations, a polynomial expansion and a finite lattice recursion to the threshold of interest.

Analytical synthesis

The mathematical object (S), its empirical construction and the contractual payment (Y) form a necessary chain of reasoning. At the first stage, a frequency law and severity law describe occurrences and marks. At the second stage, exact transforms and moments reveal the law implied by the assumptions; dependence can alter this result without changing the observed single-line marginals. At the third stage, coverage acts as a nonlinear transformation that selects the portions of the loss distribution having economic relevance for a specified contract. At the fourth stage, numerical and statistical validation examine distinct errors: arithmetic, discretization and inference.

Three independently derived comparisons make the central point. First, equal expected aggregate losses coexist with very different aggregate variances and no-loss probabilities under Poisson and mixed-Poisson count laws. Second, heterogeneous individual risks have a finite polynomial law whose economically relevant large-benefit tail can be misrepresented by a collective approximation calibrated only to the mean. Third, a compound Poisson ground-up loss of expected value three produces expected payment approximately 0.60118 after an aggregate attachment of three and limit of two. Every comparison follows from existing probabilistic models and transparent calculations. Their integration is the analytical contribution of this dossier.

Conclusion

Aggregate loss models translate a specified frequency mechanism and severity mechanism into a distribution over the total monetary consequences of insured events. The collective representation uses a random sum and the individual representation retains policy-specific contingencies; both become scientifically useful when their counting units, exposure bases, dependence relations and payment definitions are explicit. Their core mathematical results are conditional laws, transform compositions, moment decompositions and exact or approximate distributional computations.

The distribution of an aggregate contains information not carried by its expectation. Count mixing changes dispersion and the atom at zero; heterogeneous benefit amounts concentrate risk at economically important thresholds; common shocks create covariance across lines; heavy tails make high-loss outcomes sensitive to severity assumptions. Contractual deductibles and limits then transform the aggregate through nonlinear payment functions. These mechanisms explain why aggregate claims, insurer payments, stop-loss expectations, tail quantiles and capital-sensitive risk measures require separate definitions.

Panjer recursion, individual-risk convolution, numerical Fourier methods, analytical approximation and Monte Carlo simulation are complementary tools. Their mathematical validity derives from the assumptions of the chosen probability model; their practical credibility derives from correct data definitions, statistical identification, sensitivity analysis, numerical error control and source-supported interpretation. The reproducible calculations in this monograph verify the principal formula chains on explicit examples and identify where convenient approximations lose crucial distributional information.

The resulting methodological conclusion is direct: rigorous insurance aggregation begins by identifying exactly what is being counted and paid, proceeds through an explicit joint probability structure, and ends with contract-specific, error-aware evaluation. The monograph consolidates established actuarial research into an auditable framework for studying aggregate losses; it introduces no new theory or doctrine, and it carries no claim of external peer review or insured-portfolio empirical calibration.

References