Build fluency in severity tails, aggregate recursion, coverage adjustments, likelihood inference, credibility, reserving, and pricing. This October 2026 ASTAM guide pairs twenty original response tasks with sixty worked concepts, emphasizing assumptions, units, calculation structure, and interpretation. Written tasks develop explainable actuarial solutions; spreadsheet tasks require formulas and reconciliations. The actual assessment has six questions totaling 60 points over three hours, including one graded Excel question; this package is a representative preview. Familiarity with probability, financial mathematics, foundational actuarial mathematics, and mathematical statistics is assumed. Syllabus citations establish topic relevance; the worked mathematics illustrates those competencies rather than quoting technical rules from the outline.
What this guide covers
Twenty original ASTAM written and Excel tasks, with worked answers and 60 concepts for October 2026.
Severity Models
1. Scale and power transformations
For a positive severity X, multiplying by c > 0 gives FY(y) = FX(y/c), so quantiles and the mean multiply by c while variance multiplies by c². Raising X to a positive power r gives FY(y) = FX(y^(1/r)); a density calculation also requires the derivative of the inverse transformation. Moments transfer through E[Y^k] = E[X^(rk)]. These rules distinguish changes in monetary scale from changes in tail shape. A power transformation can destroy moment existence even when ordinary inflation scaling cannot.
Worked example: Suppose X has a Pareto tail with shape 3. Multiplying every loss by two doubles each finite quantile without changing the tail exponent. Squaring losses instead produces a tail exponent of 1.5, so the transformed mean exists but its variance does not.
Common mistake: Applying the scale rule to a power transformation overlooks that powers alter moment requirements and can change the tail exponent.
Official syllabus reference: [2]
2. Exponentiation and lognormal severity
If Z is normal with mean μ and variance σ², then X = exp(Z) is lognormal. Its median is exp(μ), its mean is exp(μ + σ²/2), and its pth quantile is exp(μ + σzp). Increasing σ while holding μ fixed preserves the median but raises the mean and spreads the distribution multiplicatively. All positive moments exist, although the moment generating function is infinite for positive arguments. This distinction matters when a proposed aggregate approximation requires an MGF rather than merely finite means and variances.
Worked example: An analyst sets μ = ln(100) and σ = 0.5 for claim amounts in dollars. The median is $100, but the expected claim is about $113.31. Replacing the expected claim with the median would understate the portfolio's expected aggregate loss.
Common mistake: A finite moment of every positive order does not guarantee that a distribution has a finite moment generating function near positive arguments.
Official syllabus reference: [2]
3. Mixture weights and between-component variance
A mixture selects a component first and then draws a loss from that component. Its distribution is F(x) = ΣpjFj(x), and its mean is the same probability-weighted combination of component means. Variance requires both expected within-component variance and variance of the component means. Mixtures can represent heterogeneous risks or latent claim types without imposing a sharp loss threshold between components. Tail behavior may be dominated by a small heavy-tailed component, so a low mixture weight does not necessarily make that component negligible for extreme quantiles.
Worked example: Eighty percent of claims have mean $100 and variance 2,500; the remainder have mean $500 and variance 10,000. The mixture mean is $180. Its variance is 4,000 of within-component variation plus 25,600 of between-component variation, totaling 29,600 dollars squared.
Common mistake: Averaging component variances alone omits differences between component means and therefore understates the variance of a heterogeneous mixture.
Official syllabus reference: [2]
4. Splicing normalized body and tail models
A splice models disjoint loss regions with separately normalized conditional distributions. With threshold u and body weight p, the body CDF below u is pFB(x)/FB(u), while survival above u is (1 − p)ST(x)/ST(u), when the component models use a common ground-up scale. These normalizations make the total probability one and match the CDF at the threshold. Density continuity is an additional constraint, not an automatic consequence. Splicing is useful when ordinary claims and large claims require different families but share one coherent severity distribution.
Worked example: A model assigns 90% probability below $10,000 and a Pareto conditional tail above that threshold. If the tail shape is 2, the unconditional probability above $20,000 is 0.10 times one quarter, or 0.025, regardless of the chosen bounded body distribution.
Common mistake: Joining two unnormalized densities at a threshold does not automatically create a probability distribution or preserve the intended tail weight.
Official syllabus reference: [2]
5. Tail comparison through moments and survival ratios
For Pareto-type survival proportional to x^(-α), the positive kth raw moment exists only when k < α. Comparing moment existence therefore distinguishes tails even when sample means look similar. A stronger asymptotic comparison uses the survival ratio: if S1(x)/S2(x) tends to infinity, distribution 1 eventually has the heavier tail under that criterion. This is an eventual statement and need not order moderate quantiles. Coefficients of variation or skewness are useful only when their required moments exist; numerical estimates cannot repair a theoretically infinite population moment.
Worked example: Two calibrated models have similar central claim amounts, but their Pareto shapes are 1.8 and 3.5. Both means exist. Only the second has finite variance and a finite third raw moment, so a variance-based comparison is mathematically unavailable for the first model.
Common mistake: A stable-looking sample variance does not establish finite population variance when the fitted tail model implies that the second moment diverges.
Official syllabus reference: [2]
6. Hazard rates and conditional loss intervals
For a continuous severity X, the hazard h(x) = f(x)/S(x) is the limiting conditional probability per dollar that X falls in (x, x + Δ], given X > x, as Δ approaches zero. Thus h(x)Δ approximates this probability over a short interval. For a finite interval, the exact probability is 1 − exp(−∫ from x to x + Δ h(t)dt). An increasing hazard means that, among losses already exceeding x, the local rate of ending rises with x. Hazard comparisons therefore describe conditional behavior, while survival probabilities describe unconditional exceedance.
Worked example: Suppose S(x) = exp[−(x/2000)²] for x ≥ 0. Then h(1000) = 0.0005 per dollar. Given X > $1,000, the probability X ends within the next $200 is 1 − exp(−0.11) ≈ 0.10417; the local hazard approximation gives 0.10.
Common mistake: Hazard measures the local conditional chance that a loss ends within the next interval; it is not the probability of exceeding that interval.
Official syllabus reference: [2]
7. Mean excess functions and deductible severity
The mean excess function is e(d) = E[X − d | X > d] = ∫d∞S(x)dx/S(d), provided the conditional mean exists. It links tail diagnostics directly to expected payment per positive claim under an ordinary deductible. Exponential severity has constant mean excess, whereas Lomax severity with α > 1 has e(d) = (θ + d)/(α − 1), increasing linearly. An empirical mean excess plot can suggest threshold behavior, but high-threshold estimates are unstable because few observations remain. The unconditional payment mean additionally multiplies by S(d).
Worked example: A Lomax severity has θ = $600 and α = 3. Its mean ground-up loss is $300, while its mean excess above a $400 deductible is $500. The larger conditional payment mean reflects selection of unusually large losses, rather than an arithmetic contradiction.
Common mistake: The mean excess is conditional on a positive payment; treating it as expected payment per original loss omits the survival probability.
Official syllabus reference: [2]
8. Generalized extreme value block maxima
The generalized extreme value family describes possible nondegenerate limits of normalized block maxima. Its CDF is G(x) = exp{−[1 + ξ(x − μ)/σ]^(-1/ξ)}, with σ > 0 and positive bracket. The ξ = 0 limit is Gumbel; positive ξ gives a heavy upper tail, while negative ξ gives a finite upper endpoint. Parameters describe maxima over the chosen block, not individual claim severities. For p between zero and one, its quantile is μ + σ{(−ln p)^(-ξ) − 1}/ξ, using the continuous limit at zero.
Worked example: For annual maxima modeled by a Gumbel distribution with μ = $20,000 and σ = $5,000, the 99th percentile is μ − σ ln(−ln 0.99), approximately $43,001. This is a percentile of annual maximum loss, not of an individual randomly selected claim.
Common mistake: A return-level calculation for annual maxima cannot be interpreted as an individual-claim percentile without a separate frequency and severity relationship.
Official syllabus reference: [2]
9. Generalized Pareto threshold stability and TVaR
A generalized Pareto excess Z has survival (1 + ξz/β)^(-1/ξ) on its valid support, with the exponential limit at ξ = 0. It arises as a limiting model for high-threshold exceedances under suitable tail conditions. Raising the threshold by v preserves ξ and changes scale to β + ξv. Unconditional tail probabilities also require the probability of exceeding the original threshold. For a continuous distribution and ξ < 1, the mean above a higher quantile q equals q plus the updated scale divided by 1 − ξ. For ξ ≥ 1, that tail mean is infinite.
Worked example: Above $5,000, excesses have β = $2,000 and ξ = 0.2. At a $10,000 threshold, the updated scale is $3,000. Expected loss conditional on exceeding $10,000 is therefore $10,000 plus $3,000/0.8, giving $13,750 under this fitted tail model.
Common mistake: Using the original excess scale at every higher threshold understates tail means when the generalized Pareto shape is positive.
Official syllabus reference: [2]
Aggregate Models
10. Convolution of discrete claim amounts
Convolution constructs the distribution of a sum of independent variables. For two discrete severities, P(X1 + X2 = s) = ΣxP(X1 = x)P(X2 = s − x). Repeated convolution gives the aggregate distribution conditional on a fixed claim count. Mixing these conditional distributions with frequency probabilities then gives the collective aggregate model, including its count-zero mass. Independence justifies multiplying probabilities; identical distributions are convenient but unnecessary for convolution itself. A consistent monetary lattice is essential because indices must correspond to compatible loss amounts throughout the calculation.
Worked example: Each of two independent claims is $100 with probability 0.7 or $200 with probability 0.3. Their total is $200, $300, or $400 with probabilities 0.49, 0.42, and 0.09. The middle probability includes both possible orders of the claim amounts.
Common mistake: Counting only one ordering of unequal claim amounts omits valid combinations and understates the probability of their sum.
Official syllabus reference: [2]
11. Probability generating functions for random sums
For nonnegative integer-valued lattice amounts, the probability generating function is PX(z) = E[z^X]. If claim severities are independent and identically distributed and independent of N, the aggregate PGF is PS(z) = PN(PX(z)). This composition encodes the mixture over claim counts. For Poisson frequency it becomes exp{λ(PX(z) − 1)}. Coefficients recover aggregate probabilities, while derivatives at one recover factorial moments. Amounts must first be expressed in lattice units; PGF arguments track counts of those units rather than the original currency values.
Worked example: Suppose each claim is one or two $100 units with equal probability, and the count is Poisson with mean 3. Then PS(z) = exp{3(0.5z + 0.5z² − 1)}. Its constant coefficient is exp(-3), the probability of no claims in this positive-severity model.
Common mistake: Substituting expected severity into a frequency PGF loses the severity distribution and cannot reproduce the aggregate probability mass function.
Official syllabus reference: [2]
12. Panjer recursion for the (a,b,0) class
A frequency distribution belongs to the (a,b,0) class when pn = (a + b/n)p(n−1) for n ≥ 1. For lattice severity masses fj, Panjer recursion gives gk = [Σj=1…k(a + bj/k)fjg(k−j)]/(1 − af0), with g0 = PN(f0). Poisson, binomial, and negative binomial frequencies have suitable parameter representations. The denominator and starting probability account for zero severities. Recursion is efficient because each new aggregate probability uses previously calculated values, but it requires the correct frequency parameterization and a nonnegative severity lattice.
Worked example: For Poisson mean 2, a = 0 and b = 2. If severities are always one unit, g0 = exp(-2), g1 = 2g0, and g2 = g1. The resulting aggregate is simply the Poisson count expressed in monetary units.
Common mistake: Using the zero-claim probability as g0 is incorrect when individual severities can be zero and positive counts can produce zero aggregate loss.
Official syllabus reference: [2]
13. The (a,b,1) first-probability correction
In the (a,b,1) class, pn = (a + b/n)p(n−1) is required only for n ≥ 2, so p1 need not equal (a + b)p0. The aggregate recursion therefore adds [p1 − (a + b)p0]fk to the usual Panjer numerator before division by 1 − af0. The initialization remains g0 = PN(f0). This correction accommodates zero-modified or zero-truncated frequency distributions without forcing their first probabilities into an incompatible relation. Omitting it silently substitutes a different count model, even if all later frequency ratios are correct.
Worked example: A Poisson distribution with mean 2 is conditioned on a positive count. Its new p0 is zero and p1 = 2exp(-2)/(1 − exp(-2)). With strictly positive severities, the correction supplies the nonzero probability of exactly one claim that an uncorrected recursion would miss.
Common mistake: A zero-truncated count distribution cannot be handled by merely setting the original Panjer starting probability to zero and leaving every recursion term unchanged.
Official syllabus reference: [2]
14. Nearest-point rounding of continuous severity
Rounding replaces a continuous nonnegative severity by its nearest grid point kh. For k ≥ 1, fk = F((k + 1/2)h) − F((k − 1/2)h), while f0 includes the interval from zero to h/2. This creates a discrete distribution suitable for convolution or recursive aggregation. Smaller grid widths usually improve resolution but increase computation. Rounding does not generally preserve the original mean, and a finite upper truncation introduces a separate tail approximation. Endpoint conventions matter for distributions with atoms but not for a fully continuous severity.
Worked example: For exponential severity with mean $1,000 and grid width $500, the zero mass is P(X < $250) = 1 − exp(-0.25). The mass at $500 covers losses from $250 to $750, so it equals exp(-0.25) − exp(-0.75).
Common mistake: Placing all losses from zero to one full grid width at zero implements a different discretization and can materially distort small-loss probabilities.
Official syllabus reference: [2]
15. Local linear moment matching
Local linear moment matching allocates a continuous loss between its two neighboring grid points. For x between kh and (k + 1)h, the lower-point weight is ((k + 1)h − x)/h and the upper-point weight is (x − kh)/h. These weights sum to one and reproduce x as their weighted average. Integrating the weights against the severity distribution yields grid probabilities that preserve the global first moment on an unbounded grid. The resulting second moment generally increases because replacing a fixed value by a two-point lottery adds conditional variance.
Worked example: A loss of $1,300 on a $1,000 grid allocates 70% weight to $1,000 and 30% to $2,000. Its allocated mean remains $1,300. Applying this rule across the distribution preserves expected severity while creating probabilities usable in aggregate recursion.
Common mistake: Preserving the first moment does not imply preserving variance; the endpoint allocation introduces additional dispersion within each grid interval.
Official syllabus reference: [2]
16. Compound Poisson superposition
Independent compound Poisson portfolios combine into another compound Poisson portfolio. The new intensity is λ = Σλi, and its severity distribution is the mixture with weights λi/λ. This result follows by multiplying aggregate transforms or by labeling events in the superposed Poisson process. The mixture weights reflect expected claim counts rather than premium volume, exposure shares, or equal portfolio weights. The construction requires independence of the component processes; common shocks can create dependence that invalidates this simple representation even when each marginal portfolio is compound Poisson.
Worked example: One business generates four expected claims of mean $200, while another generates one expected claim of mean $1,000. The combined severity mixture weights are 0.8 and 0.2. Its mean is $360, and intensity five gives total expected loss of $1,800.
Common mistake: Weighting severity components by their shares of expected total loss overweights high-severity portfolios and does not describe a randomly selected claim.
Official syllabus reference: [2]
17. Compound Poisson moments and cumulants
For compound Poisson aggregate S with intensity λ and severity X, E[S] = λE[X] and Var(S) = λE[X²]. The variance contains both random severity and random count effects, so it is not merely λVar(X). More generally, when the required moments exist, the rth aggregate cumulant is λE[X^r]. These identities provide useful checks on recursive probability calculations and guide aggregate approximations. They also expose a boundary: if severity has infinite second moment, compound Poisson aggregation does not create a finite aggregate variance.
Worked example: Claims arrive with Poisson mean 10. Individual severities have mean $100 and variance 2,500 dollars squared. Aggregate mean is $1,000, while variance is 10(2500 + 100²) = 125,000 dollars squared. Omitting frequency variation would produce only 25,000.
Common mistake: Using λ times severity variance treats the claim count as fixed and omits the variation caused by random numbers of claims.
Official syllabus reference: [2]
Coverage Modifications
18. Ordinary deductibles and limited expectations
An ordinary deductible d produces payment Y = (X − d)+ per original loss. Its expected value is ∫d∞S(x)dx = E[X] − E[min(X,d)], whenever the mean is finite. The payment distribution has an atom at zero equal to F(d), while positive payments retain information about the original upper tail. This survival-integral representation avoids deriving a separate payment density and extends naturally to layers. A deductible removes the first d dollars of each loss; it does not merely discard small losses while paying large losses in full.
Worked example: A loss is $300 or $1,000 with equal probability, and the deductible is $400. Payments are zero and $600, so the expected payment per loss is $300. Paying $1,000 on the second outcome would instead describe a different coverage modification.
Common mistake: Confusing an ordinary deductible with a franchise deductible overstates payments on losses that exceed the deductible threshold.
Official syllabus reference: [2]
19. Payment per loss versus payment per payment
Payment per loss includes every original claim, including zero insurer payments. Payment per payment conditions on the payment being positive. Under an ordinary deductible, the positive-payment probability is S(d), so E[Y | Y > 0] = E[Y]/S(d), assuming a positive denominator. The distinction also changes frequency: independent Poisson ground-up claims thin to Poisson paid claims with intensity λS(d). Aggregate expected loss is unchanged by choosing either representation when the matching severity denominator is used. Mixing paid-claim frequency with payment-per-loss severity counts the deductible reduction twice.
Worked example: A portfolio expects 100 ground-up claims, with 40% producing payment. Expected payment per ground-up claim is $120, so expected aggregate payment is $12,000. Equivalently, there are 40 expected positive payments with conditional mean $300, producing the same aggregate expectation.
Common mistake: Multiplying the reduced paid-claim count by the unconditional payment mean applies the zero-payment adjustment twice and understates expected aggregate payments.
Official syllabus reference: [2]
20. Maximum covered loss versus maximum payment
A maximum covered ground-up loss u with deductible d gives payment (min(X,u) − d)+, equivalent to min((X − d)+, u − d) when u > d. A maximum payment L instead gives min((X − d)+,L), whose ground-up exhaustion point is d + L. These contracts coincide only when L = u − d. Both create payment atoms at their caps, but at different thresholds if the numerical limits are confused. Writing the payment function before integrating or transforming the distribution makes the contract's actual layer explicit.
Worked example: With a $1,000 deductible, a maximum covered loss of $6,000 permits at most $5,000 payment. A separate contract with a $6,000 payment limit permits payment up to $6,000 and exhausts at a $7,000 ground-up loss. The same quoted limit number describes different benefits.
Common mistake: Subtracting a deductible from a limit explicitly defined as maximum insurer payment changes the contract and understates the covered layer.
Official syllabus reference: [2]
21. Coinsurance and the order of operations
Coinsurance multiplies a specified covered amount by an insurer participation fraction c. If a payment cap L applies afterward, Y = min(c(X − d)+,L), which equals c min((X − d)+,L/c). If instead a covered excess limit applies before coinsurance, the cap is multiplied by c. These are different payment functions. Once the correct function is established, positive scaling multiplies the mean by c and variance by c². Coinsurance percentages therefore cannot be inserted into a calculation without identifying the amount to which they apply.
Worked example: An insurer pays 80% above a $500 deductible, subject to a final payment cap of $4,000. The covered excess layer has width $5,000, and the payment reaches its cap at $5,500 ground-up loss. Capping the excess at $4,000 first would cap payment at $3,200.
Common mistake: Interchanging coinsurance and a final payment cap changes the exhaustion threshold because multiplication and capping do not generally commute.
Official syllabus reference: [2]
22. Loss elimination ratios
For an ordinary deductible d, the loss elimination ratio is LER(d) = E[min(X,d)]/E[X] = 1 − E[(X − d)+]/E[X], assuming a positive finite mean. It measures the proportion of expected ground-up dollars removed, rather than the proportion of claims eliminated. Frequency reduction is F(d), which generally differs because large claims contribute more dollars. LER is useful for translating ground-up loss costs into deductible-adjusted costs when other coverage terms are unchanged. Comparing LERs across models also highlights differences in how much expected loss lies near the origin.
Worked example: Suppose 70% of claims are below a deductible, but those claims and deductible portions of larger claims account for only 25% of expected loss dollars. The LER is 0.25, and expected insurer loss is 75% of ground-up expected loss, despite a much larger claim-count reduction.
Common mistake: Equating the percentage of claims eliminated with the percentage of expected loss eliminated ignores differences in claim sizes.
Official syllabus reference: [2]
23. Increased limits and deductible factors
An increased limits factor compares expected limited losses at two limits, commonly E[min(X,L2)]/E[min(X,L1)] under a stated common coverage basis. A deductible factor similarly compares expected payments under selected deductibles or against ground-up expected loss. Ratios must hold other terms constant, including coinsurance, inflation basis, and whether the denominator counts losses or payments. Survival integrals provide the needed limited expectations. Increased limits factors are not generally proportional to the limit ratio because only losses reaching the added layer generate additional payments.
Worked example: For exponential losses with mean $1,000, expected payments under limits $1,000 and $2,000 are $632.12 and $864.66. The increased limits factor is about 1.3679, rather than two, because most claims do not fully use the additional thousand-dollar layer.
Common mistake: Using the ratio of policy limits as an increased limits factor assumes every loss exhausts both limits and overstates expected added-layer cost.
Official syllabus reference: [2]
24. Inflation with fixed deductibles and limits
If ground-up losses inflate to cX while a deductible d and payment limit L remain fixed, expected payment becomes E[min((cX − d)+,L)] = cE[min((X − d/c)+,L/c)]. This identity converts inflation into scaled loss thresholds. It shows why insurer payments need not increase by c: fixed deductibles become smaller relative to losses, while fixed limits are exhausted more often. The net effect depends on the severity distribution and contract. If every monetary contract term also scales by c, payment instead scales exactly by c.
Worked example: A 10% inflation factor turns a fixed $550 deductible into an effective $500 deductible on the original loss scale. A fixed $2,200 payment limit similarly becomes $2,000 before multiplying the resulting payment by 1.10. Both transformed thresholds must enter the calculation.
Common mistake: Multiplying last year's limited loss cost by the inflation factor ignores the changing real value of fixed deductibles and limits.
Official syllabus reference: [2]
25. Aggregate stop-loss layers
An aggregate stop-loss treaty pays (S − r)+ above annual retention r. With a layer limit L, ceded loss is min((S − r)+,L), whose expected value is ∫r^(r+L)SS(s)ds. Retained loss is S minus this payment. Above exhaustion r + L, additional aggregate losses return entirely to the insurer, so retained loss is not capped at r. Aggregate coverage depends on the full distribution of S, including frequency and dependence, rather than only expected ground-up losses. Variance calculations require transforming each outcome or using appropriate joint moments.
Worked example: At aggregate loss $900, a treaty attaching at $500 with a $300 limit pays $300 and leaves $600 retained. An unlimited stop-loss treaty at the same attachment would pay $400 and leave $500 retained. Exhaustion explains the difference between the two contracts.
Common mistake: Treating retained loss as min(S,r) under a limited stop-loss treaty ignores the insurer's renewed exposure above treaty exhaustion.
Official syllabus reference: [2]
Construction and Selection of Parametric Models
26. Likelihood contributions for incomplete observations
Likelihood uses the probability of what was actually observed. An exact continuous loss contributes a density, right censoring at c contributes S(c), and an interval observation contributes F(b) − F(a). Left truncation at d additionally conditions each included record on X > d, dividing its contribution by S(d). Truncation changes which losses enter the sample; censoring changes how precisely an included loss is measured. These mechanisms can occur together. Correct contributions must be assembled before differentiation, since treating a bound as an exact amount changes both estimation and uncertainty.
Worked example: A study includes losses above $100, and one included record is known only to exceed $600. Its contribution is S(600)/S(100). An exact included loss of $250 contributes f(250)/S(100). Both require inclusion conditioning, but only the exact loss contributes a density.
Common mistake: Using a density at the censoring threshold treats an unknown larger loss as an exact observation and distorts the fitted severity distribution.
Official syllabus reference: [2]
27. Frequency maximum likelihood with exposure
For independent counts Ni conditional on rate λ, with Ni Poisson(wiλ), the log likelihood contains (ΣNi)log λ − λΣwi plus terms not involving λ. The MLE is total claims divided by total exposure. This rate estimate differs from the unweighted average of observed rates when exposures vary. Under the Poisson model, its variance is λ/Σwi, estimated by substituting the MLE. Counts, exposure definitions, and observation intervals must be compatible. Extra-Poisson heterogeneity or dependence affects uncertainty even if the aggregate count-to-exposure ratio remains a useful descriptive rate.
Worked example: One segment has 10 claims over 100 exposure units, and another has 12 over 300. The Poisson rate estimate is 22/400 = 0.055 per unit. Averaging the segment rates, 0.10 and 0.04, would yield 0.07 and incorrectly give equal weight to unequal exposure.
Common mistake: An arithmetic mean of segment claim rates generally differs from the exposure-based Poisson MLE and overweights small segments.
Official syllabus reference: [2]
28. Severity MLEs and parameter reparameterization
For complete independent exponential losses with mean θ, the log likelihood is −n log θ − Σxi/θ, yielding θ̂ = X̄. Under the rate parameterization λ = 1/θ, invariance of maximum likelihood gives λ̂ = 1/X̄. Invariance concerns maximizers, not unbiasedness: a transformed MLE can be biased even when the original estimator is unbiased. For more complex severity families, score equations may require numerical optimization. Parameter support, boundary solutions, and whether the stationary point is actually a maximum remain part of the estimation problem.
Worked example: Four complete exponential losses total $2,000. The estimated mean is $500 and the estimated rate is 0.002 per dollar. Taking the average of the four reciprocal loss amounts would not maximize the exponential likelihood and would estimate a different quantity.
Common mistake: Maximum likelihood invariance permits transforming the fitted parameter, but it does not permit replacing a transformed sample mean with a mean of transformations.
Official syllabus reference: [2]
29. Information matrices and normal confidence intervals
Observed information is the negative Hessian of the log likelihood at the fitted parameter. Expected information averages that curvature under the model. Under regularity conditions, the inverse information approximates the MLE covariance matrix, supporting normal confidence intervals. Off-diagonal entries describe joint estimation dependence and matter for functions of several parameters. Care is needed when information is expressed per observation rather than for the full sample. Boundary parameters, weak identification, and small samples can make a symmetric normal approximation unreliable even when numerical optimization returns a definite maximum.
Worked example: For 100 complete exponential losses with fitted mean $500, the estimated variance of the sample mean is 500²/100 = 2,500 dollars squared. The standard error is $50, giving a normal 95% interval of approximately $402 to $598 using 1.96.
Common mistake: Using the inverse of per-observation information without dividing by sample size overstates estimator variance by a factor equal to that sample size.
Official syllabus reference: [2]
30. Exact pivots and non-normal intervals
A pivotal quantity has a distribution that does not depend on the unknown parameter. For n complete exponential observations with mean θ, 2ΣXi/θ follows χ² with 2n degrees of freedom. Inverting its central probability gives an exact interval with endpoints 2Σxi/χ²upper and 2Σxi/χ²lower. This interval is asymmetric and remains positive, reflecting the sampling distribution more faithfully than a symmetric normal interval in small samples. Exactness depends on the specified sampling model; censoring or truncation generally changes the pivotal argument and cannot be ignored.
Worked example: Three complete exponential losses total $900. Using supplied χ² quantiles 1.237 and 14.449 for six degrees of freedom, the central 95% interval for θ is approximately ($124.58, $1,455.13). The upper endpoint uses the smaller quantile because θ appears in the denominator.
Common mistake: Matching the lower parameter endpoint with the lower chi-square quantile reverses the inequalities and produces an incorrectly ordered interval.
Official syllabus reference: [2]
31. Delta-method uncertainty for transformed parameters
The delta method linearizes a smooth function g near an estimated parameter. For a scalar, Var(g(θ̂)) is approximately [g′(θ)]²Var(θ̂). For a vector, the approximation is ∇gᵀΣ∇g, evaluated at the fitted parameter. This propagates estimation uncertainty to quantities such as means, tail probabilities, and limited expectations. Covariance terms are essential when several parameters jointly determine the target. The approximation can be poor near parameter boundaries, singularities, or regions of strong curvature, where a local linear description does not capture the estimator's transformed distribution.
Worked example: A mean estimate θ̂ is $100 with variance 400. For g(θ) = ln θ, the derivative at the estimate is 0.01, giving approximate variance 0.04 for the log estimate. Its standard error is 0.2, which can support an interval on the log scale.
Common mistake: Applying the function directly to an estimator's variance does not propagate uncertainty; the derivative and parameter covariance determine the delta-method calculation.
Official syllabus reference: [2]
32. Poisson–gamma Bayesian frequency estimation
For Poisson counts with rate λ and total exposure W, a gamma prior with shape a and rate b updates to gamma(a + total count, b + W). Under squared-error loss, the Bayesian estimate of λ is the posterior mean, (a + total count)/(b + W). The rate parameterization makes b comparable to exposure and a/b the prior mean. Future counts require integrating over this posterior, producing more variance than a Poisson model at the posterior mean. Confusing gamma rate with scale changes every update and predictive calculation.
Worked example: A gamma prior has shape 4 and rate 2, giving prior mean 2 claims per exposure unit. After 8 claims over 6 units, the posterior is gamma(12,8), and the estimated rate is 1.5. Expected claims over two future units are therefore three.
Common mistake: Adding exposure to a gamma scale parameter uses the wrong conjugate update; additive exposure updating applies to the rate parameter.
Official syllabus reference: [2]
33. Bayesian severity and aggregate prediction
For exponential severities parameterized by rate λ, a gamma(a,b) prior in shape-rate form updates to gamma(a + n,b + Σxi). The posterior predictive mean severity is E[1/λ | data] = (b + Σxi)/(a + n − 1), when the updated shape exceeds one. It differs from the reciprocal posterior mean rate. Aggregate prediction then combines frequency and severity while respecting any posterior dependence between their parameters. Expected aggregate loss factors into expected frequency times expected severity only under the required independence or conditioning structure; parameter uncertainty also contributes to predictive variance.
Worked example: An exponential rate prior has shape 3 and rate $1,000. Five observed losses total $2,000, giving posterior shape 8 and rate $3,000. Predictive mean severity is $3,000/7, or $428.57. An independent future count with mean ten gives expected aggregate loss of $4,285.71.
Common mistake: Taking the reciprocal of the posterior mean rate generally understates predictive mean severity because expectation and reciprocal transformation do not commute.
Official syllabus reference: [2]
34. Q–Q and survival plots for model diagnosis
A quantile–quantile plot compares empirical quantiles with fitted theoretical quantiles; approximate alignment supports agreement across the displayed range. Systematic upper-tail departures reveal discrepancies that a central fit can conceal. Survival plots on suitable transformed axes can distinguish tail forms: an exponential tail is linear on a semilog survival plot, while a Pareto-type tail is asymptotically linear on log–log axes. Graphical evidence is descriptive and sensitive to sampling variability, especially in extreme order statistics. It should identify the nature and location of misfit rather than substitute visual neatness for a calibrated test.
Worked example: A fitted lognormal Q–Q plot tracks ordinary claims closely, but its largest empirical quantiles lie above the reference line. The model may understate upper-tail losses even though its estimated mean is reasonable. A tail-focused comparison is therefore more informative than relying on central alignment alone.
Common mistake: A straight-looking plot over central observations does not establish adequate extreme-tail fit when the tail contains little data and different behavior.
Official syllabus reference: [2]
35. Kolmogorov–Smirnov testing with fitted parameters
The Kolmogorov–Smirnov statistic is the supremum of the absolute difference between an empirical CDF and a proposed continuous CDF. For ordered observations, both sides of each empirical jump must be checked. Standard distribution-free critical values assume a fully specified continuous null distribution. Estimating parameters from the same observations changes the statistic's null distribution, requiring an appropriate adjustment or calibrated simulation. The test measures the largest vertical CDF discrepancy and can be relatively insensitive to sparse extreme-tail differences. Failure to reject therefore does not establish accurate tail-risk estimation.
Worked example: An analyst fits an exponential mean using the same 50 losses used for a KS statistic. Applying a critical value intended for a known mean is not automatically valid. A parametric bootstrap can refit the mean in each simulated sample and calibrate the actual fitting-and-testing procedure.
Common mistake: The usual fully specified KS reference distribution does not remain automatically valid when model parameters are estimated from the tested sample.
Official syllabus reference: [2]
36. Chi-square goodness-of-fit and category design
A chi-square goodness-of-fit statistic sums (Oj − Ej)²/Ej over mutually exclusive, exhaustive categories. Expected counts come from fitted model probabilities times sample size. In regular settings with m categories and r independently estimated parameters, the reference degrees of freedom are typically m − 1 − r. Sparse expected cells can undermine the approximation, so categories may need sensible pooling. Category boundaries and parameter-estimation procedures affect calibration. Grouping also discards within-cell information, which means an acceptable result can coexist with important distributional differences hidden inside broad bins.
Worked example: Five loss bands contain 200 observations, and a one-parameter model is estimated by a method compatible with the chi-square approximation. The reference degrees of freedom are five minus one for the total count minus one fitted parameter, giving three. Pooling two bands would change that count.
Common mistake: Using categories minus one degrees of freedom while ignoring estimated parameters makes the reference distribution inconsistent with the fitted-model testing procedure.
Official syllabus reference: [2]
37. Likelihood ratio tests, AIC, and BIC
For regular nested models, the likelihood ratio statistic is twice the increase in maximized log likelihood, with asymptotic chi-square degrees of freedom equal to the parameter-count difference. Boundary or unidentified parameters can invalidate that reference. AIC minimizes −2ℓ + 2k, while BIC, also called SBC, minimizes −2ℓ + k ln n. Scores require comparable likelihoods fitted to the same data. Unlike a nested likelihood ratio test, they can compare suitable nonnested models. Their complexity penalties express different selection objectives, so disagreement is possible without any arithmetic error.
Worked example: Adding two parameters improves maximized log likelihood by three in a sample of 100. The likelihood ratio statistic is six. AIC improves by two, favoring the larger model, while BIC worsens by about 3.21 because its extra penalty exceeds the six-point fit improvement.
Common mistake: A lower information criterion identifies a preferred candidate under that criterion; it does not prove the selected model is the true data-generating distribution.
Official syllabus reference: [2]
Credibility
38. Bayesian greatest-accuracy prediction
Under squared-error loss, the predictor minimizing expected prediction error is the conditional expectation of the future quantity given observed information. In Bayesian credibility, a prior distribution describes risk heterogeneity, and likelihood information updates its probabilities or density. The posterior mean of the hypothetical mean becomes the prediction for a comparable future observation. For discrete risk types, posterior weights are proportional to prior weights times likelihoods. This greatest-accuracy result depends on the stated loss function and model; posterior medians or other summaries solve different decision problems and should not be substituted automatically.
Worked example: Two risk types have equal prior probabilities and annual Poisson rates one and three. After zero claims, their posterior weights are proportional to exp(-1) and exp(-3). The predicted next count is the posterior-weighted average of one and three, rather than the unweighted collective mean of two.
Common mistake: Choosing the single most probable risk type discards posterior uncertainty and generally differs from the squared-error-optimal posterior mean prediction.
Official syllabus reference: [2]
39. EPV and VHM as structural variance components
Let μ(Θ) be a risk's conditional mean and v(Θ) its conditional observation variance. Expected process variance is E[v(Θ)], while variance of hypothetical means is Var[μ(Θ)]. The law of total variance gives Var(X) = EPV + VHM. EPV measures noise within risks; VHM measures persistent differences between risks. Their ratio K = EPV/VHM determines the volume of comparable experience needed for credibility. A zero VHM means no modeled heterogeneity to learn, while negligible EPV makes individual experience highly informative about each risk's mean.
Worked example: Annual loss variance is 1,300 dollars squared, comprising EPV 1,000 and VHM 300. Two observations from the same risk share covariance 300 under conditional independence. That common component represents persistent risk quality, while their remaining annual fluctuations arise from process variation.
Common mistake: Treating total observation variance as EPV includes between-risk heterogeneity in the noise term and produces an excessively large credibility denominator.
Official syllabus reference: [2]
40. Bühlmann credibility as a linear projection
Bühlmann prediction is the best linear function of equally weighted experience under its structural assumptions. For n observations, it takes the form Z X̄ + (1 − Z)μ, where Z = n/(n + K) and K = EPV/VHM. The factor arises from covariance-based least-squares projection, rather than a claim-count threshold for full credibility. It equals the unrestricted Bayesian posterior mean in certain conjugate settings, but otherwise approximates it within a linear class. Increasing n reduces the process variance of X̄ while leaving persistent between-risk variation available to learn.
Worked example: With EPV 800, VHM 200, and two observations, K is four and Z is one third. A risk average of $160 against a collective mean of $100 produces a $120 prediction. The $60 experience deviation receives one third weight because two observations remain noisy.
Common mistake: Bühlmann credibility is not universally identical to Bayesian credibility; equality requires a model whose posterior mean has the appropriate linear form.
Official syllabus reference: [2]
41. Bühlmann–Straub precision weighting
Bühlmann–Straub extends credibility to observations with unequal precision. Conditional means are common within a risk, while Var(Xj | Θ) = v(Θ)/wj for known positive weights. The experience statistic is X̄w = ΣwjXj/W, with W = Σwj, and credibility is Z = W/(W + K). Weights often represent exposure but must match the inverse-variance assumption. The model predicts a mean rate, which can then be multiplied by future exposure. Merely having different policy counts does not justify weighting unless the corresponding conditional mean and variance structure is appropriate.
Worked example: A risk has loss rates $80 over one exposure unit and $140 over three. Its weighted rate is $125, rather than the arithmetic average $110. With K = four, total weight four gives credibility one half, so a collective mean of $100 yields a $112.50 predicted rate.
Common mistake: Weighting total losses again by exposure double-counts exposure; the weighted statistic should combine rates with their corresponding precision weights.
Official syllabus reference: [2]
42. Balanced nonparametric empirical Bayes estimates
In balanced nonparametric empirical Bayes credibility, several risks each provide the same number n of observations. Average within-risk sample variance estimates EPV. The sample variance of risk means estimates VHM + EPV/n, so subtracting estimated EPV/n estimates VHM. This separates genuine heterogeneity from the sampling noise in observed averages without specifying a full conditional distribution. A negative raw VHM estimate can occur through sampling fluctuation and commonly requires a stated nonnegative convention. The credibility model still assumes comparable risks and a suitable conditional independence structure across observations.
Worked example: Across balanced risks with four years each, average within-risk variance is 400 and sample variance of risk means is 180. Estimated VHM is 180 − 400/4 = 80, giving K = five. Treating 180 directly as VHM would overstate heterogeneity and assign too much credibility.
Common mistake: The variance of observed risk averages includes process noise, so using it directly as VHM exaggerates persistent differences between risks.
Official syllabus reference: [2]
43. Weighted nonparametric empirical Bayes estimation
For unequal Bühlmann–Straub weights, pooled weighted within-risk residuals estimate EPV: sum w_ij(X_ij − X̄i)² divided by the total within-risk degrees of freedom. Let Wi be each risk's total weight, W their sum, and X̄ the overall weighted mean. The weighted between-risk sum B = ΣWi(X̄i − X̄)² has expectation (r − 1)EPV + [W − ΣWi²/W]VHM for r risks. Subtracting the noise term and dividing by the heterogeneity coefficient estimates VHM. This adjustment prevents high-volume risks and unequal sampling precision from distorting structural variance estimation.
Worked example: Two risks have total weights two and six, so W is eight and the heterogeneity coefficient is 8 − (4 + 36)/8 = three. If their weighted between-risk sum is 500 and estimated EPV is 200, estimated VHM is (500 − 200)/3 = 100.
Common mistake: Applying the balanced EPV/n correction to unequal exposure weights ignores different precisions and does not correctly remove sampling noise from between-risk variation.
Official syllabus reference: [2]
44. Semiparametric Poisson empirical Bayes credibility
A semiparametric credibility model specifies a conditional distribution but estimates population structure empirically rather than fully prescribing a prior family. If counts are conditionally Poisson with mean wλ, observed rates have conditional variance λ/w. Thus EPV for the rate model is the collective mean rate, while VHM measures variation in latent rates. Estimating the collective mean and between-risk variance supplies K and Bühlmann–Straub factors without requiring a gamma prior. This uses more distributional structure than nonparametric empirical Bayes but less than a fully specified conjugate Bayesian model.
Worked example: Across a portfolio, the estimated collective annual claim rate is 0.20 and latent-rate variance is 0.01. Under conditional Poisson sampling, EPV is 0.20 and K is 20 exposure units. A risk with ten exposure units receives credibility 10/(10 + 20), or one third.
Common mistake: Calling the method semiparametric does not eliminate assumptions; the Poisson conditional variance relationship is central to its estimate of process variation.
Official syllabus reference: [2]
45. Posterior predictive variance versus uncertainty in a mean
Prediction uncertainty includes both future process variation and uncertainty about a risk's conditional mean. The conditional variance decomposition is Var(Xfuture | data) = E[v(Θ) | data] + Var[μ(Θ) | data]. For Poisson risks, this becomes posterior mean λ plus posterior variance λ. Credibility shrinkage estimates the future mean, but it does not eliminate random future outcomes or make the fitted mean's uncertainty equal to predictive variance. Keeping these components separate is necessary for risk intervals, aggregate simulations, and comparisons of experience-based predictions across risks with different data volumes.
Worked example: A posterior rate distribution has mean two and variance 0.5. One future Poisson year has predictive mean two and variance 2.5. A Poisson distribution fixed at the posterior mean would have variance two and would omit the remaining uncertainty about the risk's underlying rate.
Common mistake: An interval for the latent mean is not a predictive interval for next year's outcome, which must also reflect future process variation.
Official syllabus reference: [2]
Reserving and Pricing for Short-Term Insurance Coverages
46. Expected loss ratio reserving
The expected loss ratio method estimates ultimate claims as earned premium on a compatible basis multiplied by a selected expected ultimate loss ratio. Outstanding claims are estimated ultimate claims minus cumulative paid claims when the target is total unpaid liability. If incurred claims are used instead, the subtraction estimates development beyond current incurred amounts, a different quantity. The method gives little direct weight to emerging loss experience, making it useful for immature periods but sensitive to the selected loss ratio. Premium rate changes and exposure mix can require adjustments before historical ratios inform the selection.
Worked example: An accident year has earned premium $2 million and selected expected ultimate loss ratio 65%. Expected ultimate claims are $1.3 million. With $400,000 paid, total unpaid claims are $900,000; subtracting $700,000 incurred instead would estimate only $600,000 of development beyond current incurred claims.
Common mistake: Subtracting incurred claims while labeling the result total unpaid claims omits case reserves already included in incurred amounts.
Official syllabus reference: [2]
47. Volume-weighted chain-ladder development
Chain-ladder estimates age-to-age factors by dividing summed cumulative claims at the later age by summed corresponding claims at the earlier age. Multiplying future selected factors gives a cumulative development factor, and latest cumulative claims times that factor estimates ultimate claims. Outstanding paid reserves subtract latest paid claims. The method assumes the historical development pattern remains relevant to the projected cohorts. Changes in settlement speed, case reserving, claim mix, or inflation can violate that stability. A terminal development assumption or tail factor is needed when the last observed age is not truly ultimate.
Worked example: Two cohorts develop from 100 to 150 and from 300 to 420. The volume-weighted factor is 570/400 = 1.425, rather than the simple average 1.45 of their individual ratios. A new cohort at 200 projects to 285 at the next age under that selection.
Common mistake: A simple average of individual development ratios is not the same estimator as the volume-weighted ratio of corresponding cumulative totals.
Official syllabus reference: [2]
48. Bornhuetter–Ferguson emergence weighting
Bornhuetter–Ferguson combines observed emergence with an independent expected ultimate loss. If the estimated emerged proportion is p, ultimate loss is observed cumulative claims plus (1 − p) times expected ultimate loss. On a paid basis, the second term is the outstanding reserve. This differs from multiplying actual claims by a development factor: unexpectedly high early payments do not proportionately inflate the entire reserve. The emergence pattern and expected ultimate selection remain critical assumptions. Paid and reported versions must use matching cumulative observations and matching emergence proportions to retain their intended interpretation.
Worked example: Expected ultimate loss is $1 million, estimated paid proportion is 30%, and actual paid claims are $400,000. Bornhuetter–Ferguson outstanding claims are $700,000, producing ultimate loss $1.1 million. Chain-ladder would instead project $400,000/0.30, or about $1.333 million, from actual emergence.
Common mistake: Multiplying the expected ultimate by the development factor confuses the unpaid proportion with the factor used to project observed cumulative claims.
Official syllabus reference: [2]
49. Bayesian updating of immature ultimate counts
A simple Bayesian reserving model treats reported count R as Poisson(pλ), where λ is expected ultimate count and p is the known reported proportion. With gamma(a,b) prior for λ in shape-rate form, the posterior is gamma(a + R,b + p). Conditional on λ, unreported count U is Poisson((1 − p)λ), independent of reported count. Posterior expected outstanding count is therefore (1 − p)(a + R)/(b + p). This combines prior ultimate expectations with actual emergence and explicitly quantifies uncertainty, while relying on the assumed reporting proportion and conditional count model.
Worked example: Take prior shape 20 and rate 0.1, giving expected ultimate count 200. At 25% reported, 60 claims have appeared. The posterior mean ultimate intensity is 80/0.35 = 228.57, and expected unreported count is 0.75 times that value, or 171.43 claims.
Common mistake: Using the posterior ultimate count directly as outstanding count fails to remove the portion expected to have already emerged.
Official syllabus reference: [2]
50. Frequency–severity decomposition of outstanding claims
Frequency–severity reserving models the number of outstanding claims and their expected future payments separately. A useful decomposition distinguishes claims not yet reported from reported claims with remaining payments, since their payment distributions may differ. Expected outstanding loss sums expected counts times corresponding expected payment severities under suitable independence or conditional assumptions. Severity must reflect future payment amounts, not necessarily full ultimate claim size. Inflation, settlement timing, and claim mix enter the severity component. This decomposition can reveal whether reserve changes arise from claim counts, average costs, or their interaction.
Worked example: An insurer expects 40 unreported claims averaging $3,000 ultimate payment and 60 reported open claims averaging $800 of remaining payment. Estimated outstanding claims are $120,000 plus $48,000, totaling $168,000. Using full ultimate severity for the already reported claims would duplicate amounts previously paid.
Common mistake: Applying ultimate claim severity to every open reported claim overstates reserves when part of each claim has already been paid.
Official syllabus reference: [2]
51. Mack model process and estimation uncertainty
Mack's chain-ladder framework specifies conditional cumulative means E[C(i,j+1) | history] = fjC(i,j) and conditional variances σj²C(i,j), together with independence across accident years. It supports prediction-error estimation without fully specifying a claim amount distribution. Reserve uncertainty includes both future process variation and estimation error in selected development factors. Different accident years share factor estimates, creating prediction-error covariance even when their processes are independent. The familiar chain-ladder point estimate therefore comes with a structured uncertainty calculation whose assumptions can fail under changing development patterns or unsuitable cumulative data.
Worked example: Two years each require the same estimated final development factor. A higher true factor would raise both ultimate losses together. Consequently, the variance of their combined reserve includes a positive shared-factor covariance, rather than merely the sum of the separate accident-year prediction variances.
Common mistake: Independence of accident-year claim processes does not imply independence of their fitted predictions when those predictions use common estimated development factors.
Official syllabus reference: [2]
52. Poisson development models for incremental claims
A Poisson development model represents incremental claim counts with means μij = exp(αi + βj), subject to an identification constraint. Accident-year effects describe volume, while development effects describe emergence. Independent Poisson observations have variance equal to their mean. Fitting the observed triangle and projecting missing cells estimates outstanding counts. Likelihood estimates in the standard multiplicative triangular setting connect closely to chain-ladder projections. Raw currency amounts are not literal Poisson counts, so any use for monetary claims requires an appropriate modeling interpretation rather than assuming dollars are individual Poisson events.
Worked example: A fitted count model gives expected unobserved increments of 20, 12, and eight claims. Expected outstanding count is 40, and under independent Poisson increments its process variance is also 40. Parameter estimation adds uncertainty beyond that conditional process calculation.
Common mistake: Equating expected outstanding counts with their complete prediction variance ignores uncertainty in the accident-year and development parameters estimated from the triangle.
Official syllabus reference: [2]
53. Overdispersed Poisson quasi-likelihood reserving
The overdispersed Poisson development model retains a multiplicative mean structure but specifies Var(Yij) = φμij, allowing variance to differ from the Poisson mean. A common dispersion parameter can be estimated using appropriately scaled residual sums and residual degrees of freedom. Quasi-likelihood fitting can preserve familiar mean projections while enlarging uncertainty when φ > 1. The mean-variance specification alone does not uniquely define a full predictive distribution. Negative incremental payments also require care because an ordinary nonnegative Poisson interpretation cannot directly accommodate them without additional modeling choices.
Worked example: Three future incremental cells have fitted means $100, $200, and $300 in a stated monetary scale. With dispersion φ = four and conditional independence, their process variances sum to 4(100 + 200 + 300) = 2,400 in compatible squared units, before adding parameter uncertainty.
Common mistake: Multiplying the fitted mean by a dispersion parameter changes the point forecast unnecessarily; dispersion modifies the variance relationship rather than the selected mean.
Official syllabus reference: [2]
54. Calendar-diagonal effects in development triangles
An accident-year row tracks a cohort, a development column tracks maturity, and a calendar diagonal tracks observations occurring in the same calendar period. Economic inflation, operational changes, or settlement initiatives can therefore appear across a diagonal rather than within a single cohort. A development-only model may misinterpret such effects as stable age-to-age behavior. Comparing residual patterns and paid versus incurred emergence helps distinguish possible causes. Accident, development, and calendar effects also have an identification relationship, so adding unrestricted effects for all three does not automatically produce uniquely interpretable parameters.
Worked example: Payments increase unusually across several accident years during the same calendar year after a settlement initiative. The common diagonal pattern suggests a calendar effect rather than simultaneous independent deterioration in each cohort. Extrapolating the resulting faster payment factors indefinitely could misstate future reserve emergence.
Common mistake: Treating every unusual development ratio as a permanent maturity effect can turn a temporary calendar event into a distorted long-term projection.
Official syllabus reference: [2]
55. Compound loss trends and midpoint timing
Loss trend projects historical costs to the expected level during the future coverage period. Under constant annual compound trend t and a time gap h years, the factor is (1 + t)^h, with fractional years permitted. The appropriate gap depends on the historical and projected periods' average loss dates, not simply their labels. Frequency and severity trends can be modeled separately, combining multiplicatively when the exposure basis is consistent. Trend should not duplicate changes already reflected in development, benefit adjustments, or other projections. Different assumptions about calendar timing can materially change the indicated rate.
Worked example: Historical losses are centered at July 1, 2024, and projected losses at January 1, 2027. The gap is 2.5 years. A 4% annual compound loss trend gives factor 1.04^2.5, about 1.103. Using two years solely from the calendar labels would underproject expected costs.
Common mistake: Adding annual trend percentages or counting only whole calendar years misstates compound growth when the projection gap includes fractional years.
Official syllabus reference: [2]
56. Loss cost pricing with fixed and variable expenses
The loss cost method begins with projected losses per exposure unit and adds fixed expenses measured on the same basis. If variable expenses and target underwriting profit are fractions v and q of premium, the indicated rate is (loss cost + fixed expense)/(1 − v − q), provided the denominator is positive. The equation follows because premium must fund both dollar costs and premium-proportional costs. A profit provision expressed on another basis needs a different equation. Consistent exposure units and explicit expense definitions are essential when converting historical claim experience into a future rate.
Worked example: Projected claims are $700 per vehicle-year and fixed expenses are $60. Variable expenses consume 18% of premium and target profit 6%. The indicated rate is $760/0.76 = $1,000 per vehicle-year, which allocates $180 to variable expenses and $60 to the stated profit provision.
Common mistake: Adding variable expense percentages directly to loss dollars treats premium-based costs as loss-based markups and fails to solve the premium equation.
Official syllabus reference: [2]
57. Loss ratio indications and compatible bases
The loss ratio method compares projected ultimate losses with premium adjusted to the current rate level. With fixed expenses explicitly expressed as a ratio to current premium, the indicated rate factor is (projected loss ratio + fixed-expense ratio)/(1 − variable-expense ratio − profit ratio). Subtracting one gives the rate change. This agrees with loss cost pricing when exposures, projected loss levels, and premium bases match. An unadjusted historical premium denominator can confuse past rate changes with underlying loss adequacy. Expense allocations must also avoid counting the same cost in both numerator and denominator.
Worked example: Projected losses are 72% of current on-level premium, and fixed expenses are 8%. Variable expenses and profit total 25% of indicated premium. The rate factor is 0.80/0.75 = 1.0667, giving a 6.67% increase despite a projected loss ratio below 100%.
Common mistake: Comparing projected losses with premium at mixed historical rate levels can create an apparent indication driven by old rates rather than future adequacy.
Official syllabus reference: [2]
58. Premium on-level adjustment
On-leveling restates historical earned premium as though current rates had applied, holding the underlying exposure and relevant rating characteristics consistently defined. Under simplifying uniform writings and earning assumptions, geometric earning patterns can estimate the share of premium generated at each rate level. The current rate index divided by the average earned rate index supplies an on-level factor. This adjustment differs from premium trend due to changing insured values or exposure characteristics. Actual policy-level rerating can be more precise when available, especially with nonuniform writings, multiple changes, or shifting classification distributions.
Worked example: Suppose historical earned premium has average rate index 1.05, while the current index is 1.10. The on-level factor is 1.10/1.05 = 1.047619. Historical premium of $2.1 million becomes $2.2 million on-level, before any separate adjustment for changing exposures or insured values.
Common mistake: Applying the latest percentage rate change to all historical earned premium ignores the portion already earned at newer rate levels.
Official syllabus reference: [2]
59. Risk classification differential estimation
Classification differentials express a class's indicated cost relative to a chosen base class under a common pricing basis. A simple univariate estimate divides projected loss cost for each class by that of the base, but exposure differences, credibility, and interactions with other rating variables can affect interpretation. A raw observed loss ratio differential also reflects existing rate differences and must be translated carefully into proposed relativities. Selecting new differentials changes the portfolio's average relativity. Estimating relative risk and deciding the overall revenue level are connected calculations but represent different pricing decisions.
Worked example: Class A has projected loss cost $400 and class B $600, giving a B-to-A loss cost relativity of 1.5. If the current relativity is already 1.5, equal loss ratios can indicate adequate relative pricing rather than equal underlying risk. Setting both relativities to one would discard that distinction.
Common mistake: Equal class loss ratios do not imply equal expected loss costs when the classes already pay different rates per exposure.
Official syllabus reference: [2]
60. Balancing back classification relativities
For fixed exposures wi and relativities ri, average rate equals the base rate times Σwiri/Σwi. Changing relativities therefore changes revenue even before changing the base. A revenue-neutral proposed base is the current base multiplied by the ratio of old to new exposure-weighted average relativities. To achieve a separate overall rate factor F, multiply that neutral base by F. This balancing-back calculation preserves the intended portfolio indication for the specified mix. If exposure mix changes, the reconciliation must use the intended projection weights rather than assuming the old mix still applies.
Worked example: Current average relativity is 1.20 and proposed average is 1.32, with a $500 base. The neutral base is $454.55. Applying a 10% overall increase brings the base back to $500, because the differential change itself already raises the weighted average by 10%.
Common mistake: Increasing the original base by the full overall indication after raising average relativities produces more revenue than the intended portfolio rate change.
Official syllabus reference: [2]
Official syllabus sources
- Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
- October 2026 Syllabus with Learning Objectives/Outcomes and Readings