Use this guide to connect actuarial formulas with the assumptions and decisions they support. Each concept includes a worked example and a specific error to avoid. The Society of Actuaries Advanced Short-Term Actuarial Mathematics Examination (SOA Exam ASTAM) assumes knowledge from the Fundamentals of Actuarial Mathematics (FAM) Exam. Begin with severity and frequency models, then apply those foundations to credibility, coverage, ratemaking, reserving and aggregate risk.
Severity Models and Statistical Estimation
1. Survival functions and expected loss
For a nonnegative loss X, the survival function S(x) = P(X > x) describes the probability of exceeding a loss amount. Its integral over nonnegative amounts equals E[X], provided the expectation exists. This connects tail probabilities with expected cost and works for continuous, discrete and mixed loss distributions.
Worked example: A loss is zero with probability 0.4 and 200 with probability 0.6. The survival function is 0.6 between zero and 200, so E[X] = 0.6 × 200 = 120.
Mistake to avoid: Integrating the cumulative distribution function instead of the survival function to obtain an expected nonnegative loss.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
2. Heavy tails and moment existence
A heavy-tailed model can have a finite mean but an infinite variance. For a Lomax distribution with S(x) = (1 + x/θ)^(-α), θ > 0 and α > 0, the moment E[X^r] exists only when r < α. Check this condition before applying formulas that require second moments.
Worked example: With θ = 300 and α = 2.5, the mean is 300/(2.5 − 1) = 200. The variance exists, but the third moment does not.
Mistake to avoid: Assuming that a finite fitted mean guarantees a finite variance or supports every moment-based approximation.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
3. Inflation as a scale transformation
If every ground-up loss increases by a constant factor c, the transformed loss is Y = cX. Its mean becomes cE[X], its variance becomes c²Var(X), and its quantiles multiply by c. These relationships concern ground-up losses; fixed deductibles and limits can change how inflation affects insurer payments.
Worked example: A loss distribution has mean 500 and variance 40,000. Under 8% uniform inflation, the new mean is 540 and variance is 1.08² × 40,000 = 46,656.
Mistake to avoid: Multiplying variance by the inflation factor instead of its square, or automatically applying ground-up inflation to limited payments.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
4. Mixtures and between-group variation
A mixture distribution represents losses drawn from different underlying populations. Its mean is the weighted average of component means. Its variance includes both average within-component variance and variation between component means, by the law of total variance. Pooling groups therefore can produce substantial variability even when each group is internally stable.
Worked example: Two equally likely groups have means 100 and 300, each with variance 400. The mixture mean is 200 and variance is 400 + 10,000 = 10,400.
Mistake to avoid: Averaging component variances while ignoring variation between their means.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
5. Left truncation and conditional likelihood
Left truncation means losses below a threshold are absent from the dataset. For observations above d, the relevant density is f(x)/S(d), rather than f(x) alone. The likelihood must reflect selection into the sample. For exponential ground-up losses, the excess above d remains exponential with the same mean parameter.
Worked example: Only losses exceeding 100 are recorded, and the observed amounts are 150 and 250. Under an exponential model, the mean estimate is the average excess: (50 + 150)/2 = 100.
Mistake to avoid: Using the average recorded ground-up amount, 200, as the exponential mean without adjusting for truncation.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
6. Right censoring and survival contributions
A right-censored observation establishes that a loss exceeds a known amount without revealing its full value. Its likelihood contribution is a survival probability. Exact observations contribute densities. For exponential losses with mean θ, the maximum likelihood estimate is total observed exposure to loss amounts divided by the number of exact observations.
Worked example: Exact losses are 100 and 200, and another loss is censored at 300. The exponential mean estimate is (100 + 200 + 300)/2 = 300.
Mistake to avoid: Treating the censored amount as a fully observed loss and dividing by all three records.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
7. Maximum likelihood and parameterization
Maximum likelihood selects parameters that maximize the probability density or probability of the observed data under the stated model. The estimator depends on parameterization: an exponential mean θ is estimated by the sample mean, while the rate λ = 1/θ is estimated by its reciprocal. Units help distinguish these parameters.
Worked example: Exact losses of 80, 120 and 160 give θ̂ = 120. The corresponding rate estimate is λ̂ = 1/120 per monetary unit.
Mistake to avoid: Substituting an estimated mean into a formula that expects a rate, especially when software uses different parameter names.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
8. The delta method for transformed estimates
The delta method approximates uncertainty after a smooth parameter transformation. If an estimate T has small variance, Var(g(T)) is approximately [g′(T)]²Var(T). It is useful for converting uncertainty in rates, logarithms or fitted parameters into uncertainty in quantities used for pricing or prediction.
Worked example: An estimated exponential rate is 0.01 with standard error 0.001. For mean g(λ) = 1/λ, the approximate standard error is 0.001/0.01² = 10.
Mistake to avoid: Transforming a standard error by taking its reciprocal rather than using the derivative of the transformation.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
9. Empirical distribution diagnostics
The empirical cumulative distribution function gives the proportion of observed losses at or below each amount. Comparing it with a fitted cumulative distribution reveals local discrepancies. A discrepancy at one amount is a diagnostic observation; a supremum-based goodness-of-fit statistic requires checking the relevant range, and its calibration depends on parameter estimation.
Worked example: For losses 100, 200, 400 and 800, the empirical cumulative probability at 300 is 0.5. A fitted value of 0.7 gives a local discrepancy of 0.2.
Mistake to avoid: Reporting one local discrepancy as the overall maximum discrepancy without examining the rest of the distribution.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
10. Likelihood-based model comparison
The Akaike information criterion balances fit against parameter count: AIC = 2k − 2ℓ, where k is the number of estimated parameters and ℓ is the maximized log-likelihood. Lower AIC is preferred among comparable models fitted to the same observations using compatible likelihoods. It does not establish that the selected model captures the tail adequately.
Worked example: Model A has k = 2 and ℓ = −120, giving AIC 244. Model B has k = 3 and ℓ = −118, giving AIC 242; B is preferred.
Mistake to avoid: Comparing AIC values from different datasets or interpreting a lower AIC as proof of accurate extreme-loss predictions.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
Frequency and Aggregate Loss Models
11. Exposure-adjusted claim frequency
Claim counts must be interpreted relative to exposure. Under a homogeneous Poisson model, expected claims equal rate per exposure unit multiplied by exposure. For independent observations with a common rate, the maximum likelihood rate estimate is total claims divided by total exposure. The exposure definition must remain consistent across observations.
Worked example: Two portfolios produce 18 claims over 120 policy-years and 12 over 80 policy-years. The estimated rate is 30/200 = 0.15 claims per policy-year.
Mistake to avoid: Averaging raw claim counts from portfolios with different exposure volumes or mixing policy-months with policy-years.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
12. Overdispersion and the negative binomial
A Poisson count has equal mean and variance. Heterogeneity in claim rates can create overdispersion, with variance exceeding the mean. A gamma mixture of Poisson rates yields a negative binomial distribution. In a common parameterization with mean μ and shape r, its variance is μ + μ²/r.
Worked example: A negative binomial model has μ = 4 and r = 2. Its variance is 4 + 16/2 = 12, substantially above the Poisson variance of 4.
Mistake to avoid: Assuming every overdispersed dataset proves a gamma mixing distribution; clustering or omitted predictors can also explain excess variation.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
13. Poisson thinning under a deductible
If ground-up claims arrive as a Poisson process and each independently exceeds a deductible with probability q, payment-producing claims have a Poisson distribution with mean λq. Independence of claim severities from the count process is essential. Frequency reduction and the conditional distribution of paid severity are separate calculations.
Worked example: Ground-up annual frequency is Poisson with mean 10. If half of losses exceed the deductible, the annual number of positive payments is Poisson with mean 5.
Mistake to avoid: Reducing claim frequency while leaving the severity distribution unchanged when modeling payments conditional on exceeding the deductible.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
14. Shared latent rates and count dependence
Counts can be independent conditional on a latent risk parameter yet dependent after averaging over that parameter. If N₁ and N₂ are conditionally independent Poisson counts with common random mean Λ, their unconditional covariance is Var(Λ). This distinguishes independent process variation from dependence created by persistent risk heterogeneity.
Worked example: A shared annual risk parameter has mean 3 and variance 2. Each year's count has variance 5, while the covariance between two years is 2.
Mistake to avoid: Treating conditional independence as unconditional independence when the same unobserved risk parameter influences multiple periods.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
15. Expected aggregate loss
For aggregate loss S = X₁ + ⋯ + X_N, assume severities are identically distributed and independent of N, with finite means. Conditioning on the count gives E[S] = E[N]E[X]. This separates frequency and severity contributions, allowing changes in either component to be translated into expected portfolio cost.
Worked example: Expected annual count is 8 and expected severity is 250. Expected aggregate loss is 8 × 250 = 2,000.
Mistake to avoid: Using this product without checking independence when years with more claims also systematically have larger claim amounts.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
16. Aggregate loss variance
For independent, identically distributed severities independent of N, Var(S) = E[N]Var(X) + Var(N)(E[X])². The first term measures severity variation within a given count; the second measures variation in claim volume. With Poisson frequency, this simplifies to Var(S) = λE[X²], rather than λVar(X).
Worked example: Let E[N] = Var(N) = 4, E[X] = 100 and Var(X) = 900. Aggregate variance is 4 × 900 + 4 × 10,000 = 43,600.
Mistake to avoid: Omitting the frequency contribution and calculating aggregate variance as only expected count times severity variance.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
17. Probability-generating functions
For nonnegative integer-valued severities, the aggregate probability-generating function satisfies G_S(z) = G_N(G_X(z)) under the usual independence assumptions. For Poisson frequency, G_S(z) = exp(λ(G_X(z) − 1)). Its value at zero gives the probability of zero aggregate loss, including the possibility of zero-valued individual losses.
Worked example: With Poisson mean 3 and strictly positive severities, G_X(0) = 0. Therefore P(S = 0) = exp(−3), approximately 0.0498.
Mistake to avoid: Equating zero aggregate loss with zero claims when the severity model permits claims of amount zero.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
18. Convolution of discrete severities
Convolution combines independent loss distributions by summing the probabilities of all combinations producing each total. For two losses, P(X₁ + X₂ = s) = Σ P(X₁ = x)P(X₂ = s − x). Repeated convolution handles fixed claim counts and provides the conditional distributions needed in compound models.
Worked example: Each of two independent claims equals 1 or 3 with equal probability. Their total equals 2, 4 or 6 with probabilities 0.25, 0.50 and 0.25.
Mistake to avoid: Counting only one ordering of unequal claim amounts; both 1 + 3 and 3 + 1 contribute to total 4.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
19. Compound Poisson recursion
For Poisson frequency with mean λ and positive integer severities with probabilities f_j, aggregate probabilities satisfy g₀ = exp(−λ) and g_k = (λ/k)Σ j f_j g_(k−j), summing j from 1 to k. The recurrence computes each probability from earlier values and requires a consistent severity lattice.
Worked example: With λ = 2, f₁ = 0.75 and f₂ = 0.25, g₁ = 1.5exp(−2). Then g₂ = 0.75g₁ + 0.5g₀ = 1.625exp(−2).
Mistake to avoid: Dropping the severity multiplier j or forgetting to initialize the zero-loss probability.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
20. Normal approximation to aggregate loss
A normal approximation uses an aggregate distribution's mean and standard deviation to estimate probabilities through standardized values. Its usefulness depends on skewness, concentration and tail behavior, rather than claim count alone. Heavy-tailed severity or sparse frequency can make extreme-tail estimates poor even when the mean and variance are correctly calculated.
Worked example: For approximate mean 2,000 and standard deviation 300, a threshold of 2,493 has z = 493/300 ≈ 1.643, giving an approximate exceedance probability of 5%.
Mistake to avoid: Treating a convenient normal approximation as reliable for extreme losses without checking the aggregate distribution's shape.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
Credibility and Bayesian Estimation
21. Within-risk and between-risk variation
Credibility separates random fluctuation within a risk from persistent differences between risks. Expected process variance, EPV, is E[Var(X|Θ)]. Variance of hypothetical means, VHM, is Var(E[X|Θ]). Their sum is the unconditional variance of one observation. Stronger between-risk differences make an individual risk's experience more informative.
Worked example: If EPV = 100 and VHM = 25, the variance of a single observation is 125. Of that total, 25 reflects differences in underlying risk means.
Mistake to avoid: Interpreting all observed variation as persistent risk differences rather than distinguishing process noise from heterogeneity.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam; Studying for Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
22. Limited-fluctuation full credibility
Limited-fluctuation credibility asks whether experience is sufficiently stable relative to a specified tolerance and probability. For a Poisson count with expected value n, a normal approximation gives n ≥ (z/k)² for relative tolerance k and a two-sided probability associated with z. This is an approximate statistical criterion, not a universal measure of reliability.
Worked example: For 10% relative tolerance and approximately 95% probability, z = 1.96 gives n ≥ 384.16, so an integer expected count of 385 satisfies the approximation.
Mistake to avoid: Using this Poisson count criterion for aggregate losses without accounting for severity variation.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
23. Partial credibility blending
A credibility estimate combines experience and a complementary estimate: Z × experience + (1 − Z) × complement. Under a common limited-fluctuation partial credibility rule, Z = min(1, √(n/n₀)), where n₀ is the full-credibility volume. Both estimates must concern the same exposure basis, coverage and cost level.
Worked example: With n = 96 and n₀ = 384, Z = 0.5. Blending an experience estimate of 140 with a complement of 100 produces 120.
Mistake to avoid: Blending an unadjusted historical cost with a current prospective cost simply because both are expressed in money.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
24. Bühlmann credibility
For equal-exposure observations under Bühlmann assumptions, Z = n/(n + K), where K = EPV/VHM. The estimate is Z times the risk's sample mean plus 1 − Z times the collective mean. Larger process noise reduces credibility; larger variation in underlying risk means increases it.
Worked example: EPV = 200 and VHM = 50 give K = 4. With six observations, Z = 0.6. A risk mean of 130 and collective mean of 100 yield 118.
Mistake to avoid: Substituting total unconditional variance for EPV when calculating K.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
25. Bühlmann–Straub exposure weighting
Bühlmann–Straub credibility accommodates unequal exposure when the conditional variance of an observed loss rate is inversely proportional to its exposure weight. The risk mean is exposure-weighted, and credibility is Z = w/(w + K), where w is total exposure. A large period should contribute more than a small period.
Worked example: Rates of 80 and 120 have weights 1 and 3, giving a weighted mean of 110. With K = 6, Z = 4/10; blending with 90 gives 98.
Mistake to avoid: Taking an unweighted average of loss rates before applying exposure-based credibility.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
26. Estimating credibility variance components
In a balanced credibility dataset, average within-risk sample variance estimates EPV. Variation between observed risk means includes both VHM and sampling noise, so the latter must be removed. With n observations per risk, subtract EPV/n from the sample variance of risk means to estimate VHM.
Worked example: Risk observations are 80, 120 and 160, 200. Each within-risk sample variance is 800. Risk means 100 and 180 have sample variance 3,200, giving VHM estimate 3,200 − 800/2 = 2,800.
Mistake to avoid: Treating the full variance of observed risk averages as VHM without subtracting within-risk sampling variation.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
27. Beta-binomial updating
A beta prior for a claim probability is conjugate to independent Bernoulli observations conditional on that probability. Starting with Beta(α, β), observing x claims among n trials gives Beta(α + x, β + n − x). The posterior mean is the predictive claim probability for one additional comparable trial.
Worked example: A Beta(2, 8) prior and three claims among ten trials produce Beta(5, 15). The posterior predictive claim probability is 5/20 = 0.25.
Mistake to avoid: Adding all ten trials to the claim parameter instead of separating claims from nonclaims.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
28. Gamma-Poisson updating with exposure
For Poisson counts conditional on rate λ, use exposure explicitly: N|λ has mean wλ. A gamma prior with shape α and rate β updates to shape α + N and rate β + w. The posterior mean rate is (α + N)/(β + w), which can then be multiplied by future exposure.
Worked example: A Gamma(shape 3, rate 2) prior, nine claims and exposure 4 give posterior mean rate 12/6 = 2. Expected claims over future exposure 2 are 4.
Mistake to avoid: Confusing gamma rate with scale or adding observed claims to the prior rate parameter.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
Insurance Coverage and Ratemaking
29. Payment per loss versus payment per payment
With ordinary deductible d, payment per ground-up loss is Y = max(X − d, 0), including zero payments. Payment per payment conditions on X > d. These have different means: E[Y] = P(X > d)E[X − d|X > d]. The distinction connects ground-up claim frequency with payment-producing frequency.
Worked example: Losses of 100 and 300 are equally likely, with deductible 150. Mean payment per loss is 75; mean payment conditional on a positive payment is 150.
Mistake to avoid: Multiplying ground-up claim frequency by mean payment per payment without adjusting for the probability of a positive payment.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
30. Limited expected value
A payment capped at u with no deductible is min(X, u). Its expectation is the limited expected value, equal to the integral of S(x) from zero to u for nonnegative losses. This measures expected cost after the cap and also provides building blocks for deductible and layer calculations.
Worked example: Losses are 100 or 400 with equal probability. A cap of 250 gives expected payment 0.5 × 100 + 0.5 × 250 = 175.
Mistake to avoid: Capping the expected loss instead of calculating the expectation of individually capped losses.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
31. Ordering deductibles, caps and coinsurance
Coverage terms form a payment function, and their order matters. If a contract first subtracts deductible d, caps the resulting covered amount at u, and then applies insurer share c, payment is c min(max(X − d, 0), u). Alternative contractual orders produce different payments, so define the function before calculating.
Worked example: For X = 500, d = 100, u = 300 and c = 0.8 under this ordering, insurer payment is 0.8 × 300 = 240.
Mistake to avoid: Applying coinsurance before the cap when the specified coverage caps the amount before applying the insurer's share.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
32. Aggregate deductibles
An aggregate deductible applies to the total loss over a defined period, giving payment max(S − D, 0). It differs from applying a deductible separately to each claim because claims jointly exhaust the aggregate retention. Expected payment requires the aggregate distribution, rather than just individual severity information or mean aggregate loss.
Worked example: Aggregate loss is 0, 100 or 400 with probabilities 0.2, 0.5 and 0.3. With aggregate deductible 150, expected payment is 0.3 × 250 = 75.
Mistake to avoid: Calculating max(E[S] − D, 0), which ignores the distribution of losses around the deductible.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
33. Attachment points and layer widths
A layer attaching at a with width l pays min(max(X − a, 0), l). Attachment determines when payment begins; width determines its maximum. Expected layer loss equals the difference between stop-loss expectations at a and a + l, linking layer pricing directly to the severity tail.
Worked example: Losses of 100 and 500 have probabilities 0.4 and 0.6. A layer attaching at 200 with width 200 pays zero or 200, giving expected payment 120.
Mistake to avoid: Treating the layer width as the exhaustion point; this layer exhausts at ground-up loss 400.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
34. Written and earned premium
Written premium records premium associated with policies written during a period. Earned premium allocates premium to elapsed coverage. Matching earned premium with losses arising during that coverage period supports meaningful loss ratios. Straight-line earning is appropriate only when coverage and risk earning are assumed uniform over the policy term.
Worked example: An annual policy begins October 1 with premium 1,200. Assuming uniform earning and no cancellation, three months of coverage earn 300 by December 31.
Mistake to avoid: Using all 1,200 as earned premium for the calendar year when only one quarter of coverage has elapsed.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
35. Pure premium and frequency-severity decomposition
Pure premium is expected loss cost per exposure unit, before expense and profit provisions. In experience data, total losses divided by exposure also equals claim frequency times average severity when the same claims and exposure basis are used. Prospective calculations additionally require suitable development, trend and coverage adjustments.
Worked example: Sixty claims total 120,000 over 600 exposure units. Frequency is 0.1, severity is 2,000 and pure premium is 200 per exposure.
Mistake to avoid: Including claim expenses in one component while excluding them from the other, breaking the consistency of the decomposition.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
36. Expense and profit provisions in a rate
Separate fixed expenses per exposure from expenses expressed as a proportion of premium. If expected loss cost is L, fixed expense is F, variable expense proportion is v and profit provision is p, the indicated premium is (L + F)/(1 − v − p), provided the denominator is positive.
Worked example: With L = 200, F = 20, v = 0.20 and p = 0.05, indicated premium is 220/0.75 = 293.33 per exposure.
Mistake to avoid: Adding 25% of loss cost for variable expenses and profit when those provisions are defined as percentages of premium.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
37. Loss-ratio rate indications
The loss-ratio approach compares the adjusted experience loss ratio with a permissible loss ratio. When expense and profit provisions are proportions of premium, the permissible ratio is one minus those provisions. The indicated rate factor is adjusted loss ratio divided by permissible loss ratio, assuming the experience has been made prospective and placed at current rate level.
Worked example: An adjusted loss ratio of 0.70 and permissible ratio of 0.65 give a rate factor of 1.0769, or a 7.69% increase.
Mistake to avoid: Using an undeveloped historical loss ratio directly as the prospective adjusted loss ratio.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
38. Premium on-level adjustment
On-leveling expresses historical premium at a common rate level so experience is comparable with current pricing. Divide the target rate-level index by the average historical index. The averaging weights must reflect how premium was earned and how rate changes affected policies; a simple midpoint assumption is not automatically appropriate.
Worked example: Suppose equal earned-exposure halves carry indices 1.00 and 1.10. The average index is 1.05, so premium 100 on-levels to 100 × 1.10/1.05 = 104.76.
Mistake to avoid: Applying the full 10% increase to premium that already partly reflects the higher rates.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
39. Combining frequency and severity trends
When frequency and severity trends apply independently on consistent exposure and coverage bases, pure premium trend is their product. Multiplicative trends compound over time. Choose the trend interval from the relevant experience period to the prospective period, rather than simply counting calendar labels or adding percentages.
Worked example: Annual frequency falls 1% while severity rises 4%. Annual pure premium factor is 0.99 × 1.04 = 1.0296; over two years, cost 200 becomes approximately 212.02.
Mistake to avoid: Adding the trend percentages and then applying a simple, noncompounding adjustment over multiple years.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
40. Class relativities and the base rate
Class relativities allocate a required average premium across risk classes. The base rate must account for the exposure-weighted average relativity: base rate equals total required premium divided by the sum of exposure times relativity. Changing class mix can therefore change the required base rate even if relativities remain fixed.
Worked example: For 100 exposures at relativity 1 and 50 at 1.5, required average premium 200 implies base rate 30,000/175 = 171.43 and higher-class rate 257.14.
Mistake to avoid: Setting the base rate equal to the overall average premium without adjusting for the portfolio's average relativity.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
Loss Reserving and Predictive Uncertainty
41. Origin, development and calendar periods
A loss triangle organizes observations by origin period and development age. Calendar periods run along diagonals. With development age starting at zero, calendar year equals origin year plus development age. Distinguishing these dimensions helps identify whether changes arise from the insured cohort, claim maturity or conditions affecting all activity in a calendar period.
Worked example: Origin year 2024 at development age 2 corresponds to calendar year 2026. Origin year 2026 at age 1 corresponds to 2027 and is future at year-end 2026.
Mistake to avoid: Reading an unobserved future cell as a zero loss or mixing development-age conventions.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
42. Incremental and cumulative losses
Incremental losses measure activity during one development period. Cumulative losses sum activity through that period. Convert cumulative amounts to increments by differencing adjacent observations, and increments to cumulative amounts by summing. A genuine zero increment differs from an unobserved cell; negative increments may occur in some incurred or recovery-adjusted datasets.
Worked example: Incremental paid amounts 100, 60 and 40 produce cumulative amounts 100, 160 and 200. Conversely, 200 − 160 recovers the final increment of 40.
Mistake to avoid: Adding cumulative observations together, thereby counting earlier payments repeatedly.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
43. Paid, case reserves and IBNR
In a common gross reserving convention, reported incurred loss equals paid loss plus outstanding case reserves. Total unpaid loss equals ultimate loss minus paid loss. IBNR, including development beyond current case estimates under this convention, equals ultimate loss minus reported incurred loss. Keep all amounts on the same recovery and expense basis.
Worked example: Ultimate loss is 1,000, paid loss 600 and case reserves 250. Reported incurred is 850, total unpaid is 400 and IBNR is 150.
Mistake to avoid: Calling ultimate minus paid IBNR, which includes case reserves as well.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
44. Volume-weighted development factors
A volume-weighted age-to-age factor divides the sum of cumulative losses at the later age by the sum at the earlier age, using only origin periods observed at both ages. It gives larger influence to larger earlier-age amounts. Factor selection still requires reviewing unusual claims, operational changes and consistency of the development pattern.
Worked example: Matched observations develop from 100 to 150 and from 200 to 260. The factor is 410/300 = 1.3667, rather than the unweighted ratio average of 1.4.
Mistake to avoid: Including an unmatched origin period in only one side of the factor calculation.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
45. Cumulative development to ultimate
A cumulative development factor multiplies successive selected age-to-age factors from the observed age to ultimate. Chain-ladder ultimate is latest cumulative loss times that factor. For a paid triangle, estimated unpaid loss is ultimate minus paid to date. These projections rely on the selected historical development pattern remaining appropriate.
Worked example: Selected remaining factors are 1.4 and 1.1, giving a cumulative factor of 1.54. Paid loss 200 projects to ultimate 308 and unpaid loss 108.
Mistake to avoid: Adding successive development factors instead of multiplying them.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
46. Tail development beyond observed maturity
A tail factor represents development beyond the final explicitly modeled age. It may reflect persistent late reporting, settlement or payment patterns. Its influence can be material because it applies across multiple origin periods. A tail assumption should be justified separately from observed age-to-age factors and subjected to sensitivity analysis.
Worked example: Ultimate loss before a tail adjustment is 800. A selected tail factor of 1.05 increases ultimate to 840, adding 40 to the estimated unpaid amount.
Mistake to avoid: Assuming the last observed development age is ultimate merely because the triangle ends there.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
47. Bornhuetter–Ferguson reserving
Bornhuetter–Ferguson combines observed emergence with an independently selected expected ultimate loss. If the cumulative development factor is F, the expected un-emerged proportion is 1 − 1/F. Estimated ultimate equals observed cumulative loss plus expected ultimate times this proportion. This reduces sensitivity to immature actual emergence while retaining dependence on the prior expectation.
Worked example: Paid loss is 400, F = 2 and independently expected ultimate is 1,000. Estimated unpaid loss is 500 and estimated ultimate is 900.
Mistake to avoid: Multiplying the observed cumulative amount by the un-emerged proportion instead of applying it to expected ultimate.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
48. Expected-loss-ratio reserving
The expected-loss-ratio method estimates ultimate loss as earned premium times a selected expected loss ratio, with premium and losses on consistent levels and coverage bases. It relies on the expectation rather than observed emergence. Unpaid loss is then estimated ultimate minus paid loss, so actual payments still affect the reserve.
Worked example: Earned premium 2,000 and expected loss ratio 0.60 imply ultimate loss 1,200. If 700 has been paid, the estimated unpaid loss is 500.
Mistake to avoid: Subtracting reported incurred loss when the requested result is total unpaid loss.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
49. Cape Cod estimation
A basic Cape Cod approach estimates a common expected loss ratio from aggregate emerged losses divided by premium weighted by expected emergence proportions. It then uses that ratio to estimate un-emerged losses for each origin period. Its interpretation depends on comparability across periods, including premium rate levels, loss trends and development assumptions.
Worked example: Two periods each have premium 1,000, emergence proportions 0.8 and 0.4, and paid losses 480 and 240. The estimated ratio is 720/1,200 = 0.60; second-period unpaid loss is 360.
Mistake to avoid: Dividing emerged losses by total premium without adjusting for how much loss is expected to have emerged.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
50. Mack model assumptions
The Mack framework specifies conditional means and variances without requiring a full loss distribution. In its standard form, expected next cumulative loss is f_j times current cumulative loss, and conditional variance is σ_j² times current cumulative loss. Independence across origin periods is another central assumption; fitted factors alone do not establish these conditions.
Worked example: With current cumulative loss 100, f_j = 1.3 and σ_j² = 9, next-period conditional mean is 130 and conditional variance is 900, giving standard deviation 30.
Mistake to avoid: Treating a distribution-free moment model as if it automatically supplied a complete predictive distribution.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
51. Process and parameter uncertainty
Predictive uncertainty includes future random loss development and uncertainty in estimated model parameters. The law of total variance separates expected conditional variance from variance of conditional expected outcomes. These components must be combined on the variance scale. Neither captures all possible model misspecification or structural change unless those are explicitly modeled.
Worked example: If estimated process variance is 25 and parameter variance is 9, combined predictive variance is 34 and predictive standard deviation is √34 ≈ 5.83.
Mistake to avoid: Adding standard deviations, 5 + 3, instead of adding variances.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
52. Overdispersed Poisson reserving models
An overdispersed Poisson reserving model commonly models incremental means through origin and development effects, with variance φ times the mean. A log link makes those mean effects multiplicative and keeps fitted means positive. The framework specifies a mean-variance relationship; it does not require monetary increments to be literal Poisson counts.
Worked example: An incremental cell has fitted mean 200 and dispersion φ = 3. Its modeled variance is 600 and standard deviation is approximately 24.49.
Mistake to avoid: Forcing a positive-mean multiplicative model onto materially negative increments without reconsidering the modeling approach.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
53. Bootstrap reserve distributions
A reserving bootstrap resamples suitably adjusted residuals, reconstructs data and refits the model to represent estimation uncertainty. Additional simulation of future development can represent process uncertainty. Residual scaling, centering and leverage adjustments depend on the fitted framework. The resulting distribution remains conditional on the chosen model and its assumptions.
Worked example: Three illustrative simulated reserves of 100, 150 and 200 have mean 150 and sample standard deviation 50; these describe the draws, not a validated confidence interval.
Mistake to avoid: Calling a refitting-only bootstrap fully predictive when future process variation has not been added.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
54. Calendar-period effects in development
Residual patterns along calendar diagonals can indicate effects such as inflation, settlement changes or reporting disruptions that an origin-development model misses. A persistent diagonal pattern challenges the assumption that historical development factors transfer unchanged. Diagnose the mechanism and its duration before deciding whether to adjust data, factors or model structure.
Worked example: Several cells on one calendar diagonal exceed fitted amounts; one shows actual 120 against fitted 100, a 20% excess. The shared pattern warrants investigation rather than automatic extrapolation.
Mistake to avoid: Turning one calendar-period discrepancy into a permanent development adjustment without checking other cells and underlying causes.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
Risk Measures and Aggregate Risk
55. Value at Risk for discrete losses
Value at Risk at probability p is the lower p-quantile: the smallest loss amount whose cumulative probability is at least p. It identifies a threshold, not the average loss beyond that threshold. Discrete distributions require careful attention to jumps in cumulative probability and the quantile convention being used.
Worked example: Loss is 0, 100 or 1,000 with probabilities 0.90, 0.08 and 0.02. Cumulative probability reaches 0.98 at 100, so 95% Value at Risk is 100.
Mistake to avoid: Interpreting Value at Risk as a maximum possible loss or as expected loss in the tail.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
56. Tail Value at Risk and probability atoms
Define Tail Value at Risk as the average quantile over probabilities from p to 1. For continuous distributions, this usually equals the conditional mean above Value at Risk. With discrete probability atoms, only the required fraction of the mass at the quantile enters the upper tail, so simple conditioning can give a different result.
Worked example: For probabilities 0.90 at 0, 0.08 at 100 and 0.02 at 1,000, 95% Tail Value at Risk is (0.03 × 100 + 0.02 × 1,000)/0.05 = 460.
Mistake to avoid: Using only losses strictly above Value at Risk, which would incorrectly produce 1,000 in this example.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
57. Risk capital above expected loss
A simplified economic capital convention measures unexpected loss as a selected risk measure minus expected loss. Under this convention, expected loss is funded separately and capital covers the additional uncertainty allowance. The definition must state the risk measure, probability level, horizon and whether the quantity concerns gross, net or discounted losses.
Worked example: If one-year Value at Risk is 500 and expected loss is 200, capital defined as Value at Risk minus expected loss is 300.
Mistake to avoid: Presenting this illustrative convention as a universal regulatory capital requirement or forgetting how expected loss is funded.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
58. Dependence and diversification
For two loss components, Var(X + Y) = Var(X) + Var(Y) + 2Cov(X, Y). Positive dependence reduces diversification benefits compared with independence; negative dependence increases them. Zero covariance does not generally imply independence, and variance-based diversification does not fully describe joint extreme-loss behavior.
Worked example: Two components each have variance 100 and covariance 50. Aggregate variance is 300 and standard deviation is 17.32, compared with 14.14 under independence.
Mistake to avoid: Adding component standard deviations or assuming independent losses merely because they belong to different business categories.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
59. Stop-loss transfer and retained loss
An unlimited aggregate stop-loss arrangement with retention a transfers max(S − a, 0), leaving retained loss min(S, a). Expected retained loss plus expected recovery equals expected gross loss when all quantities use the same basis. Actual contracts may include caps or other terms, which must be reflected in the payment function.
Worked example: Aggregate loss is 0, 100 or 500 with probabilities 0.80, 0.15 and 0.05. At retention 200, expected recovery is 15, expected retained loss 25 and expected gross loss 40.
Mistake to avoid: Treating expected recovery as the reduction in total cost without considering any premium or other costs of the transfer.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
60. Simulation error versus model uncertainty
For independent simulation draws, an estimated exceedance probability p̂ from n draws has approximate Monte Carlo standard error √(p̂(1 − p̂)/n). More draws reduce numerical sampling error but do not repair incorrect severity, frequency or dependence assumptions. Tail estimates also require checking whether enough simulated events represent the region of interest.
Worked example: Four hundred exceedances among 10,000 draws give p̂ = 0.04 and standard error √(0.04 × 0.96/10,000) ≈ 0.00196.
Mistake to avoid: Interpreting a small simulation standard error as evidence that the underlying actuarial model is accurate.
Source reference: Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
Sources
Exam identity verified:
- Advanced Short-Term Actuarial Mathematics (ASTAM) Exam
- Studying for Advanced Short-Term Actuarial Mathematics (ASTAM) Exam