Advanced claims reserving connects data quality, development assumptions, stochastic prediction, and contractual cash flows. This guide provides twenty original numerical and essay tasks with worked answer guides and assessable rubrics, followed by sixty technical lessons. Practice covers credibility methods, growth curves, Mack prediction error, GLMs, bootstrap and Bayesian distributions, risk margins, retrospective premium assets, and reinsurance. All content sits within A. Estimation of Claims Liabilities, the Fall 2026 outline's sole domain at 100%. The allocation is a representative preview, not an official item count. Citations establish syllabus relevance; technical explanations use established actuarial knowledge rather than claiming the outline proves each formula.
What this guide covers
Twenty written practice tasks and sixty lessons on advanced claims reserving for CAS Exam 7, Fall 2026.
A. Estimation of Claims Liabilities
1. Accident, development, and calendar time
A claims triangle organizes experience by origin period and development age. Each diagonal corresponds to a calendar period, subject to the indexing convention used. Accident-year effects describe differences among cohorts, development effects describe emergence as cohorts mature, and calendar effects influence several cohorts simultaneously. This structure helps locate data problems and economic changes. It also creates a modeling boundary: calendar time is determined by origin time and development age, so unrestricted effects on all three axes are not independently identifiable without constraints or additional assumptions.
Worked example: Suppose paid severity rises across several accident years on the same valuation diagonal. A common calendar influence, such as repair-cost inflation, deserves investigation before assigning separate unfavorable accident-year adjustments to every affected cohort.
Common mistake: Treating each unfavorable diagonal cell as independent accident-year deterioration can hide one shared calendar mechanism and distort future projections.
Official syllabus reference: [2]
2. Triangle reconciliation and basis consistency
Reconciliation establishes that a triangle represents the intended claims population and valuation basis. Cumulative amounts should equal accumulated increments, and selected totals should reconcile to credible transaction or financial records after explained adjustments. Check valuation dates, currencies, expenses, recoveries, reopened claims, and gross versus net definitions. A correct arithmetic total does not establish consistent coverage across years. Preserve adjustment trails so changes in estimates can be separated from changes in data. Reconciliation precedes modeling because statistical sophistication cannot repair an unidentified shift in what observations measure.
Worked example: A paid triangle totals $42 million, while the ledger shows $45 million. Transaction review identifies $3 million of allocated expenses omitted from the extract. The analyst either adds those expenses consistently or explicitly estimates them separately.
Common mistake: Forcing a triangle to match a ledger through an unexplained balancing entry can preserve totals while corrupting development patterns.
Official syllabus reference: [2]
3. Case-reserve adequacy versus ultimate deterioration
Reported claims combine paid claims and case outstanding. A change in case-reserving practice can therefore move reported development without changing expected ultimate payments. Examine case outstanding per open claim, claim severity, closure rates, and matched claim histories to distinguish stronger case estimates from genuine cost deterioration. Adjusting historical reported data may improve comparability when the practice change is measurable. The adjustment should not mechanically remove adverse experience that reflects real claim information. Paid data provide a useful comparison, although payment timing can also change and obscure interpretation.
Worked example: A new review program increases case outstanding from $20 million to $26 million while paid severity and closures remain stable. Matched files attribute $5 million to revised reserving guidance, supporting a targeted comparability adjustment rather than eliminating the entire increase.
Common mistake: Assuming every reported-loss increase represents higher ultimate costs ignores changes in case recognition and can exaggerate the reserve response.
Official syllabus reference: [2]
4. Settlement-speed distortion in paid development
Paid chain ladder assumes a reasonably stable relationship between development age and the proportion of ultimate claims paid. Faster settlement increases early paid claims even when ultimate costs are unchanged. Applying historical factors to that higher early base can overstate ultimate; slower settlement can produce the opposite effect. Diagnose the change through closure rates, payment counts, open-claim inventories, and paid-to-reported comparisons. A timing adjustment should reflect the affected cohorts and intervals. Faster settlement can also change settlement amounts, so pure acceleration should be distinguished from an accompanying severity effect.
Worked example: Historically, 40% of ultimate was paid by 12 months. A new settlement process raises that proportion to 50%. For $50 million paid, the historical factor indicates $125 million ultimate, whereas the revised emergence assumption indicates $100 million.
Common mistake: Reading increased early payments as proportionate ultimate deterioration assumes that the payment process stayed stable despite evidence of operational acceleration.
Official syllabus reference: [2]
5. Segmentation and changing portfolio mix
A pooled development pattern is a weighted combination of its constituent risks. When their development speeds differ, a change in business mix can change aggregate emergence without changing either segment's behavior. Segment by characteristics that explain material differences, such as coverage, claim size, jurisdiction, or attachment. Balance homogeneity against credibility: excessive segmentation creates unstable estimates. Recombining segment projections should reproduce the portfolio's exposure and claim totals. Mix diagnostics should use relevant current and historical weights rather than assuming premium shares necessarily equal ultimate-loss shares.
Worked example: Short-tail business pays 80% by one year and long-tail business pays 30%. A shift toward long-tail exposures lowers aggregate early payment percentages even if both segments remain stable. Separate projections reveal that the pooled slowdown is compositional.
Common mistake: Applying a pooled historical factor after a major mix change can misinterpret composition as development deterioration within every segment.
Official syllabus reference: [2]
6. Inflation adjustment and double-counting control
Inflation adjustments put claim amounts on a comparable cost basis or project them to future payment periods. Distinguish origin-period severity changes from calendar-period inflation affecting unsettled claims. Historical development already contains the inflation experienced during its observation periods, so adding a separate future inflation assumption requires checking what the selected factors implicitly project. Frequency, claim mix, limits, and settlement timing can confound observed average severity. A useful adjustment states the cost index, exposure basis, timing convention, and whether it changes ground-up claims before contractual limits are applied.
Worked example: A severity model projects payments two years ahead from a current cost level of $10,000 per claim. With an assumed 5% annual increase, projected severity is $11,025. The analyst first removes any separately embedded future inflation allowance.
Common mistake: Adding an inflation load to development factors that already project the same cost increase can count one source of deterioration twice.
Official syllabus reference: [2]
7. Salvage and subrogation recovery timing
Salvage and subrogation can reduce net claim costs while following a different development pattern from gross payments. Recoveries may arrive after claim settlement, making net incremental payments small or negative in later periods. Choose a consistent basis: project gross payments and recoveries separately, or justify a stable net model. Separate modeling is especially useful when recovery practices or recoverable populations change. Avoid subtracting recoveries twice when financial data already present payments net of collections. The remaining recovery asset depends on both expected amounts and their relationship to settled or outstanding claims.
Worked example: Gross ultimate payments are $30 million, expected total recoveries are $4 million, gross paid is $20 million, and recoveries received are $1 million. Gross unpaid is $10 million, remaining recoveries are $3 million, and net unpaid is $7 million.
Common mistake: Projecting net payments and then subtracting a separate full recovery estimate can duplicate recoveries already embedded in the development data.
Official syllabus reference: [2]
8. Volume-weighted age-to-age factor selection
A volume-weighted age-to-age factor divides the sum of next-age cumulative claims by the sum of corresponding current-age cumulative claims. It equals an average of individual factors weighted by current cumulative volume. This selection emphasizes larger observed cohorts and aligns with common chain-ladder assumptions, but large anomalous years can dominate it. Compare the factor with individual ratios, alternative averaging periods, and underlying operational changes. Excluding observations requires a substantive reason and consistent treatment; selecting the lowest convenient average does not establish a defensible estimate.
Worked example: Two cumulative pairs are 100 to 140 and 20 to 40. The pooled factor is 180/120 = 1.50, while the simple average is 1.70. The larger cohort receives greater weight under the pooled calculation.
Common mistake: Calling a simple average of individual ratios volume weighted overlooks the different influence assigned to large and small cohorts.
Official syllabus reference: [2]
9. Chain-ladder ultimate and unpaid reconciliation
Chain ladder multiplies current cumulative claims by selected remaining development factors to estimate ultimate claims. The product must include every future interval and any selected tail. Unpaid claims equal estimated ultimate less paid claims at the valuation date. If reported claims are the projection basis, ultimate less reported measures future reported development rather than total unpaid liability. Distinguish these balances explicitly to avoid excluding case outstanding. The method responds strongly to current emergence, making stable development patterns and comparable historical cohorts central to its interpretation.
Worked example: Reported claims are $75 million, paid claims are $50 million, and the reported age-to-ultimate factor is 1.20. Ultimate is $90 million. Future reported development is $15 million, but total unpaid claims are $40 million, including existing case outstanding.
Common mistake: Subtracting reported claims from ultimate and labeling the result total unpaid claims leaves existing case reserves out of the liability.
Official syllabus reference: [2]
10. Bornhuetter–Ferguson emergence weighting
Bornhuetter–Ferguson estimates ultimate as observed cumulative claims plus the unreported or unpaid proportion of an independently selected expected ultimate. With emerged proportion p, its formula is C + (1 − p)U0. Equivalently, it blends chain-ladder ultimate C/p with U0 using weights p and 1 − p. This reduces the influence of immature emergence, but reliability still depends on both the expected ultimate and the development pattern. On a reported basis, subtract paid claims to obtain total unpaid liability; subtract reported claims only to isolate future reported development.
Worked example: Expected ultimate is $100 million, reported claims are $30 million, and 25% is expected reported. Bornhuetter–Ferguson ultimate is 30 + 0.75 × 100 = $105 million, compared with chain ladder's $120 million indication.
Common mistake: Treating Bornhuetter–Ferguson as independent of development assumptions ignores that its unreported proportion is itself derived from an emergence pattern.
Official syllabus reference: [2]
11. Benktander's iterative credibility mechanism
Benktander updating replaces the initial expected ultimate in a Bornhuetter–Ferguson calculation with an updated estimate. Under fixed emerged proportion p and recursion Un = C + (1 − p)Un−1, the weight on the original prior after n iterations is (1 − p)^n. The remaining weight falls on chain-ladder ultimate. This provides a transparent progression from the prior toward experience. State the indexing convention because names for successive iterations can differ. Repeating the calculation does not create new claim information or remove bias in the development proportion.
Worked example: With p = 0.50, initial ultimate $80 million, and chain-ladder ultimate $120 million, the first update is $100 million. A second update is $110 million, placing 75% weight on chain ladder and 25% on the initial estimate.
Common mistake: Interpreting each iteration as an independent new observation overstates credibility; the update repeatedly reuses the same emerged claim information.
Official syllabus reference: [2]
12. Brosius-style linear credibility prediction
A linear credibility approach predicts ultimate claims from observed immature claims using an intercept and slope rather than requiring a proportional projection through the origin. Under squared-error loss, the best linear predictor has slope Cov(U,C)/Var(C) and intercept E[U] minus that slope times E[C]. The intercept reflects prior expected volume, while the slope measures how informative current emergence is about ultimate. Exposure normalization and comparable cohorts matter when estimating these relationships. The predictor is linear; nonlinear dependence, changing portfolios, or extrapolation beyond historical experience may require another specification.
Worked example: Suppose E[U] = 100, E[C] = 40, Cov(U,C) = 120, and Var(C) = 80, all on a consistent million-dollar scale. The slope is 1.5 and intercept 40. Observed C = 50 predicts ultimate 115.
Common mistake: Forcing the regression through zero removes the prior-volume component and assumes proportionality that the credibility relationship may not support.
Official syllabus reference: [2]
13. Credible loss-ratio reserve weighting
Credible loss-ratio reserving combines a prior expected loss ratio with experience-based information, using a weight justified by relative uncertainty. This is the central decision behind the credible-reserve methods compared in Hürlimann's reading. The experience signal and its variance depend on maturity, exposure, and the adopted development model; maturity alone does not determine every optimal weight. When combining two unbiased, independent ultimate estimates with variances vE and vP, inverse-variance weighting assigns experience weight vP/(vE + vP). Dependence or bias changes that result and must be considered explicitly.
Worked example: An experience ultimate has variance 400 and a prior-based ultimate has variance 100, in matching squared units. Under independent unbiased-estimate assumptions, experience receives weight 100/500 = 20%. Estimates of 140 and 100 combine to produce 108.
Common mistake: Using an arbitrary credibility weight because a cohort is young bypasses the model assumptions that determine how informative its experience actually is.
Official syllabus reference: [2]
14. Cape Cod expected-loss-ratio estimation
Cape Cod estimates a common expected loss ratio from observed claims relative to emerged exposure. In a simple formulation, divide total cumulative claims by the sum of premium times the expected emerged proportion, then apply the estimated ratio to un-emerged premium. This uses the portfolio's own experience to estimate the prior level rather than treating it as wholly external. Premiums must be comparable for rate level, mix, and coverage. A shared loss ratio can be misleading when cohorts have materially different profitability, so trend or other justified adjustments may be needed.
Worked example: Two cohorts each have premium 100, emerged proportions 0.80 and 0.40, and reported claims 56 and 32. The estimated ratio is 88/120 = 73.33%. Applying it to 80 of un-emerged premium gives future development of 58.67.
Common mistake: Dividing observed claims by all premium instead of emerged premium systematically confuses incomplete development with a lower expected loss ratio.
Official syllabus reference: [2]
15. Clark growth curves: Weibull and loglogistic
A growth curve G(t) represents the expected proportion of ultimate claims emerged by age t. Common Clark formulations use Weibull or loglogistic curves, with scale and shape parameters controlling timing and curvature. Expected increments are ultimate times differences in G across interval boundaries. The selected family influences extrapolation beyond observed ages, especially the tail. Compare residuals and reserve sensitivity across plausible forms rather than choosing only by in-sample fit. Development time must be measured consistently, and interval definitions matter when converting a continuous curve into expected observed cells.
Worked example: For G(t) = t²/(t² + 24²), emergence is 50% at 24 months and approximately 69.23% at 36 months. An ultimate of $130 million therefore produces expected payments of $25 million during that interval.
Common mistake: Assuming two growth curves with similar observed-age fit must imply similar tails overlooks their different extrapolation behavior beyond the data.
Official syllabus reference: [2]
16. Clark incremental likelihood and parameter risk
Clark-style stochastic fitting uses expected incremental claims derived from a growth curve and estimates parameters under an explicit likelihood or quasi-likelihood structure. The mean for an interval is proportional to G(t2) − G(t1), with ultimate scale determined by the selected model. Modeling increments avoids pretending overlapping cumulative totals are independent. Estimated curve parameters create reserve uncertainty, and weakly observed tails can produce strong parameter correlations. A converged optimizer is insufficient evidence of a stable reserve: inspect parameter plausibility, residuals, sensitivity, and the effect of alternative starting values or curve families.
Worked example: A fitted curve predicts 10%, 15%, and 20% emergence in three observed intervals. Multiplying these differences by an ultimate of $200 million gives means of $20 million, $30 million, and $40 million for the incremental likelihood.
Common mistake: Treating a successful numerical optimization as proof of adequate tail information confuses computational convergence with statistical identification and predictive reliability.
Official syllabus reference: [2]
17. Tail-factor extrapolation and sensitivity
A tail factor projects claims beyond the last development age represented in the triangle. It can be selected from external experience, fitted development decay, growth curves, or a blend supported by the portfolio. Tail uncertainty affects both the reserve level and predictive variability, especially when remaining claims are long tailed. State whether the factor applies to paid or reported claims and whether expenses or recoveries are included. A small change in an apparently modest tail factor can materially affect reserves because it multiplies a large cumulative amount.
Worked example: Cumulative claims at the final modeled age are $500 million. Tail factors of 1.02 and 1.05 imply additional development of $10 million and $25 million. The three-point factor difference changes the remaining estimate by $15 million.
Common mistake: Omitting tail uncertainty because the selected factor is close to one can materially understate risk on a large cumulative base.
Official syllabus reference: [2]
18. Compound frequency–severity reserve moments
For aggregate claims S formed by summing N independent, identically distributed severities, with N independent of severity, E[S] = E[N]E[X]. Variance is E[N]Var(X) + Var(N)E[X]². The first term measures severity variation conditional on count; the second measures count variation. For Poisson counts, this simplifies to intensity times the severity second moment. These relationships support simulation checks and reserve decompositions. They require a clearly defined remaining-claim population: outstanding known claims may have different severity information from unreported claims, so one compound model may not suit both.
Worked example: Expected unpaid count is 100 with variance 150. Mean severity is $10,000 and severity standard deviation is $20,000. Expected payments are $1 million, while variance is 100 × 20,000² + 150 × 10,000² = $55 billion squared-dollar units.
Common mistake: Multiplying expected count by severity variance alone omits variability in the number of claims and understates aggregate uncertainty.
Official syllabus reference: [2]
19. Calendar effects and development-factor stability
Historical age-to-age factors combine development behavior with the calendar conditions under which claims emerged. Persistent calendar inflation or processing changes can therefore make a single development factor unstable across accident years. Plot individual factors by origin period and inspect residuals by calendar diagonal. Compare paid, reported, frequency, and severity measures to identify a plausible mechanism. A calendar model requires assumptions about future conditions; adding historical calendar effects does not automatically forecast them. This analysis connects factor diagnostics to economic interpretation rather than treating every factor fluctuation as sampling noise.
Worked example: Successive 24-to-36-month ratios rise from 1.10 to 1.15 to 1.22 while several other development ages worsen on matching calendar diagonals. A shared calendar driver is more plausible than unrelated age-specific changes and warrants a targeted model comparison.
Common mistake: Averaging an emerging factor trend over many years may reduce apparent noise while preserving a systematic mismatch with current conditions.
Official syllabus reference: [2]
20. Venter-style proportionality and intercept tests
A proportional development model assumes expected next cumulative claims equal a factor times current cumulative claims. Residuals should not retain systematic relationships with current size, origin period, or other explanatory information. A regression allowing an intercept can test whether forcing the relationship through zero is defensible. Curvature or changing variance may indicate a different functional form or segmentation need. Statistical significance alone does not establish economic materiality, especially with few observations. Compare implied reserve changes, diagnostics, and out-of-sample behavior before replacing a familiar factor model.
Worked example: Small cohorts consistently develop by an additional fixed amount, while large cohorts show similar percentage growth. Fitting next cumulative claims as a + bC may reveal a positive intercept that a pure factor model systematically misses for small cohorts.
Common mistake: A high correlation between successive cumulative values does not verify proportionality, zero intercept, or the variance assumptions required by the reserve model.
Official syllabus reference: [2]
21. Limited expected values for excess layers
For nonnegative severity X, the limited expected value LEV(d) is E[min(X,d)]. Expected payment for a layer of limit L excess of attachment A is LEV(A + L) − LEV(A). This identity follows from expressing the layer payment as the difference of two limited ground-up amounts. Multiplying by expected claim count produces expected aggregate layer losses when count and severity assumptions support it. Use consistent inflation, expense, and policy-limit bases. A layer expected value alone does not determine its variance or development pattern.
Worked example: A severity model gives LEV($1 million) = $120,000 and LEV($3 million) = $180,000. Expected payment for $2 million excess of $1 million is $60,000 per ground-up claim. With 50 expected claims, expected layer losses are $3 million.
Common mistake: Subtracting attachment from average ground-up severity does not reproduce expected excess loss because attachment applies separately to each claim.
Official syllabus reference: [2]
22. Trend with fixed attachments and limits
Severity trend changes which claims enter a layer and which exhaust it. For trend multiplier k, apply the layer function to kX when attachment and limit remain fixed. Generally, E[layer(kX)] differs from kE[layer(X)]. Equivalently, the adjusted ground-up distribution can be evaluated against attachment A/k and limit L/k before rescaling. The effect depends on severity density near contractual boundaries. This distinction is essential when transferring historical layer experience to current costs or comparing development across attachment levels.
Worked example: A $900,000 claim lies below a $1 million attachment. After 20% severity trend it becomes $1.08 million and produces an $80,000 layer payment, assuming sufficient limit. Trending its original zero layer payment would still produce zero.
Common mistake: Using ground-up inflation directly as a layer inflation factor ignores claims entering coverage and claims whose recoveries are already capped.
Official syllabus reference: [2]
23. Layer-specific development and attachment penetration
Excess claims develop as underlying claims grow through attachments, making layer emergence depend on claim-size development as well as reporting delay. A claim can be known early yet produce no excess reported amount until later case revisions. Different layers can therefore have materially different development patterns even on the same ground-up portfolio. Sahasrabuddhe's layer-development topic connects these patterns with trend and claim-size models. Analyze movement by size band and attachment, including exhaustion effects. Transferring a ground-up pattern requires an explicit mechanism linking immature and ultimate claim sizes.
Worked example: A reported claim grows from $700,000 at 12 months to $1.4 million at ultimate. A $1 million excess attachment sees no early amount but $400,000 ultimately. The layer appears late developing even though the underlying claim was reported promptly.
Common mistake: Assuming all reported underlying claims have already emerged into an excess layer confuses claim notification with penetration of the attachment.
Official syllabus reference: [2]
24. High-deductible retained-loss proportions
For per-claim deductible d and ground-up ultimate X, the insured's deductible obligation is min(X,d), while the excess amount is max(X − d,0), before other contract terms. High-deductible reserving must distinguish ultimate deductible losses, deductible reimbursements already received, and amounts still collectible. Development of the retained proportion can change as large claims grow beyond the deductible. Siewert's high-deductible topic addresses this relationship between size and development. A collection-risk adjustment is conceptually separate from estimating the underlying deductible loss obligation and should not be hidden inside claim development factors.
Worked example: Two claims have ultimate amounts of $80,000 and $400,000 under a $250,000 deductible. Ultimate deductible obligations total $330,000. If reimbursements received are $200,000, the remaining contractual reimbursement amount is $130,000 before any collection adjustment.
Common mistake: Applying one deductible to the portfolio total instead of each claim misstates both retained losses and the insurer's excess exposure.
Official syllabus reference: [2]
25. Per-occurrence versus aggregate deductibles
A per-occurrence deductible applies separately to each covered occurrence, while an aggregate deductible applies to a defined total over the contract period. Their order and interaction are governed by the contract, so economically similar labels can produce different recoveries. Model occurrence terms before aggregate terms when that is the specified structure. Remaining cover also depends on prior losses and payments that erode deductibles or limits. A total expected loss amount is insufficient to value nonlinear aggregate features without information about frequency, severity, dependence, or a stipulated aggregate distribution.
Worked example: Three losses of $400,000 each produce no recovery under a $500,000 per-occurrence attachment. Under a separate contract covering annual aggregate losses above $500,000 with adequate limit, the same losses produce $700,000 of recovery.
Common mistake: Treating aggregate and occurrence attachments as interchangeable overlooks how claim count and loss concentration change the contractual recovery calculation.
Official syllabus reference: [2]
26. Retrospective premium contract arithmetic
A retrospective premium formula links final premium to ratable losses through specified conversion factors and expense components, subject to contractual bounds. The precise formula must come from the scenario or contract; it is not universal. Apply minimum and maximum premium to the correct total, then subtract billed or otherwise recognized premium to identify the remaining asset or liability. Distinguish ratable losses from total claim costs when exclusions or loss limitations apply. This premium valuation complements claim reserving but should remain separately traceable because billing, collection, and loss development have different mechanics.
Worked example: Under P = min(max(100 + 1.2L, 250), 800), with amounts in thousands, L = 400 gives premium 580. If 500 has been billed, the remaining premium asset is 80, before discount or collection adjustments.
Common mistake: Applying the premium maximum only to the additional billing amount can exceed the contractual cap on final total premium.
Official syllabus reference: [2]
27. Expected premium under nonlinear loss bounds
When premium is a nonlinear function of ultimate losses, expected final premium generally differs from premium evaluated at expected losses. Minimums, maximums, and loss limitations change outcomes asymmetrically. Estimate expected premium by integrating the contractual formula over the loss distribution or by averaging scenario-level simulated premiums. Correlation matters when several loss components jointly determine premium. A central loss estimate can still be useful, but it does not contain enough information to price optionality created by bounds. Separate sensitivity to loss level from sensitivity to the distribution's spread.
Worked example: Premium equals the smaller of losses and 100, in thousands. Losses are 50 or 150 with equal probability. Expected premium is 75, whereas applying the formula to expected losses of 100 gives 100, overstating the expectation.
Common mistake: Applying a capped formula to mean losses assumes expectation commutes with a nonlinear transformation, which generally fails near contractual boundaries.
Official syllabus reference: [2]
28. Retrospective premium asset runoff
The retrospective premium asset is the expected remaining premium attributable to prior coverage, measured relative to amounts already recognized under the chosen accounting basis. Its runoff depends on ultimate ratable losses, adjustment schedules, contract closure, and collection. Teng and Perkins focus on estimating this premium asset rather than treating it as a fixed percentage of unpaid claims. Reconcile changes into loss-estimate effects, billings, collections, and assumption changes. Avoid counting the same receivable both as a billed premium balance and an unbilled retrospective asset. Discounting and credit adjustments require explicit, separate assumptions.
Worked example: Expected final premium is $12 million, recognized premium is $10 million, and $1 million of recognized premium remains uncollected. The unrecognized retrospective amount is $2 million; the separate billed receivable is $1 million, making the distinction operationally important.
Common mistake: Subtracting cash collected instead of the relevant recognized-premium balance can mix billed receivables with the unbilled retrospective premium asset.
Official syllabus reference: [2]
29. Mack chain-ladder conditional assumptions
The classical Mack framework specifies conditional means and variances for cumulative development without prescribing a full claim distribution. In its standard form, expected next cumulative claims equal fj times current cumulative claims, and conditional variance equals sigmaj² times current cumulative claims. Accident years are assumed independent at the process level. These assumptions support prediction-error estimation, not a unique predictive percentile distribution. Check proportionality, variance scaling, and cross-year dependence through diagnostics and business knowledge. Shared parameter estimates can still create covariance among reserve prediction errors despite independent underlying accident-year processes.
Worked example: If cumulative claims are 100 and the next-age factor is 1.15, conditional expected next cumulative claims are 115. With sigma² = 0.20 on the same monetary scale, conditional variance is 20, rather than a fixed variance independent of size.
Common mistake: Calling Mack distribution free does not mean assumption free, nor does it automatically provide a uniquely determined 99th-percentile reserve.
Official syllabus reference: [2]
30. Mack development-variance estimation
For a development interval, the standard Mack variance estimator averages weighted squared deviations of individual development ratios from the selected volume-weighted factor. With n eligible pairs, sigma-hat² = sum[Ci(Fi − f-hat)²]/(n − 1), under the classical specification. Larger cumulative observations receive more weight because their ratio variance is assumed smaller. Late ages may have too few pairs for direct estimation, requiring a transparent extrapolation or other selected treatment. The monetary units of sigma² depend on the scale of cumulative claims, so calculations must preserve scale consistency.
Worked example: Two current cumulative values are 100 and 200, with ratios 1.20 and 1.10. The pooled factor is 1.1333. Weighted squared deviations total approximately 0.6667, and division by one residual degree of freedom leaves sigma-hat² approximately 0.6667.
Common mistake: Using an unweighted variance of individual ratios assumes a different size–variance relationship from the classical Mack specification being applied.
Official syllabus reference: [2]
31. Mack process and estimation-error propagation
Reserve prediction error combines future process variation with uncertainty in estimated development parameters. In a one-step approximation, process variance is sigma²C and factor-estimation variance contributes C²Var(f-hat). Multi-step projections propagate uncertainty through products of factors and projected cumulative amounts. The two components answer different questions: future randomness remains even if parameters were known, while estimation uncertainty can shrink with informative data. A prediction standard error is not a confidence interval for the expected reserve and does not by itself specify tail shape or a percentile.
Worked example: For C = 50, sigma² = 0.10, and factor variance 0.002, process variance is 5 and parameter variance is 5. Total prediction variance is 10, giving standard error approximately 3.162 on the chosen monetary scale.
Common mistake: Reporting only uncertainty in the estimated mean omits future claim-process variation and understates the variability of actual reserve outcomes.
Official syllabus reference: [2]
32. Shared-factor covariance across accident years
Accident-year reserves often use common estimated development factors. An error in one factor moves several projections together, creating estimation covariance even when accident-year process errors are independent. For a shared one-step factor, the approximate covariance contribution is CiCjVar(f-hat). Aggregate prediction variance includes individual variances plus twice each distinct covariance. This is a central distinction between conditional process independence and independence of prediction errors. Ignoring it can exaggerate diversification. Additional shared inflation, tail, or exposure parameters can create further dependence beyond the common development factors.
Worked example: Two reserves have prediction variances 9 and 16 and covariance 4. Their aggregate variance is 9 + 16 + 2 × 4 = 33, giving standard error 5.745 rather than the independence-based value of 5.
Common mistake: Adding standalone standard errors, or adding only standalone variances without justified covariance treatment, does not correctly aggregate reserve prediction uncertainty.
Official syllabus reference: [2]
33. ODP GLM representation of chain ladder
An overdispersed Poisson incremental model commonly uses log(muij) = alphai + betaj and Var(Yij) = phi muij. On a suitable complete triangle with the conventional multiplicative structure, fitted projections reproduce familiar chain-ladder estimates. The ODP label specifies a mean–variance relationship rather than requiring integer-valued monetary observations. It does not uniquely determine a future simulation distribution. The GLM representation enables covariates, diagnostics, and stochastic extensions, but extra terms can change both projections and identifiability. Nonpositive increments require particular attention because the usual positive-mean model may not accommodate their mechanism.
Worked example: If exp(alpha) = 200 and development multipliers are 0.30, 0.20, and 0.10, expected increments are 60, 40, and 20. Their sum is 120 over those intervals, with further development represented by additional multipliers or a tail.
Common mistake: Assuming an ODP reserve model requires claim payments to be integer counts confuses the variance formulation with a literal ordinary Poisson data model.
Official syllabus reference: [2]
34. GLM parameter identifiability constraints
In an additive log-mean model with accident and development effects, adding a constant to every accident effect and subtracting it from every development effect leaves fitted means unchanged. Fix a reference effect or impose another valid constraint to obtain unique coefficients. Count only independently estimable parameters when calculating residual degrees of freedom. Adding unrestricted calendar effects creates further dependencies because calendar period follows from accident and development time. Identifiability constraints resolve algebraic ambiguity but do not guarantee stable extrapolation; sparse observations can still leave parameters weakly estimated.
Worked example: A model with five accident effects and five development effects has ten displayed coefficients but only nine independent mean parameters under one ordinary two-way constraint. With 15 observed cells, residual degrees of freedom are six, assuming full rank otherwise.
Common mistake: Counting every displayed coefficient as an independent parameter can miscalculate dispersion and obscure rank deficiencies in the fitted design matrix.
Official syllabus reference: [2]
35. Pearson dispersion and residual structure
For an ODP model, a Pearson residual before dispersion standardization is (Y − mu)/sqrt(mu). Summing its squares and dividing by residual degrees of freedom estimates phi. Inspect residuals against fitted values, development age, accident year, and calendar diagonal; a single scale estimate cannot correct systematic mean-model errors. A fan shape may suggest an unsuitable variance function, while diagonal structure may suggest calendar effects. Dispersion absorbs unexplained variation but should not be used to excuse a missing mechanism that materially changes projected means or dependencies.
Worked example: A triangle has 28 cells, 13 independent mean parameters, and Pearson sum of squares 45. Estimated dispersion is 45/15 = 3. A future cell with mean 20 then has modeled variance 60 on the same numerical scale.
Common mistake: Increasing dispersion until residuals appear tolerable cannot remove systematic bias caused by an incorrectly specified mean or missing calendar driver.
Official syllabus reference: [2]
36. ODP bootstrap's two simulation stages
An ODP bootstrap typically first simulates estimation uncertainty by resampling appropriately treated residuals, reconstructing historical pseudo-data, refitting, and projecting future means. A second stage simulates future process outcomes conditional on each projected mean and selected variance structure. Summing simulated future increments creates predictive reserve draws. Retaining only fitted future means produces a different, narrower distribution. The procedure assumes the residual pool is representative and the fitted model is adequate. Tail, dependence, or structural uncertainties absent from the model require explicit additional treatment rather than being presumed included.
Worked example: Across bootstrap refits, future mean reserves vary between 90 and 110. Drawing future increments around each mean broadens the resulting actual-outcome distribution. The first range describes variation in estimated means; it is not already the complete predictive range.
Common mistake: Labeling the distribution of refitted expected reserves a predictive distribution omits the randomness of future payments conditional on each fit.
Official syllabus reference: [2]
37. Leverage correction in residual resampling
Fitted residuals are generally less variable than unobserved model errors because parameter estimation pulls fitted values toward the observations. In GLM approximations, leverage hii measures that influence, and dividing a residual by sqrt(1 − hii) can correct its compressed variance. Some bootstrap implementations use broader degrees-of-freedom adjustments; apply a coherent method rather than combining corrections indiscriminately. High-leverage corner cells can make corrections unstable or undefined. Inspect their structure and use a defensible pool, aggregation, or parametric treatment rather than allowing one extreme adjusted residual to dominate simulation.
Worked example: A residual of 2 with leverage 0.75 becomes 2/sqrt(0.25) = 4 under the simple leverage correction. If leverage is effectively one, the same formula is unusable and requires explicit handling rather than blind numerical application.
Common mistake: Treating residuals from heavily fitted corner cells as ordinary exchangeable errors can understate variation or create unstable corrections during resampling.
Official syllabus reference: [2]
38. Negative increments and model support
Negative incremental claims can arise from recoveries, case releases, corrections, or other legitimate mechanisms. A standard positive-mean ODP workflow may not reproduce them appropriately, especially if simulation uses a positive-support distribution. Investigate the source before choosing a treatment. Alternatives include separate recovery modeling, revised segmentation, a distribution that supports the observations, or an explicitly justified data transformation with reversed adjustments in the output. Deleting negatives or setting them to zero changes both the reserve level and variability. The appropriate response depends on whether the negatives reflect errors or real claim behavior.
Worked example: A late development cell contains $2 million of payments and $3 million of subrogation collections, producing a net increment of negative $1 million. Separate gross-payment and recovery projections preserve the economic components that a positive-only net model cannot represent.
Common mistake: Replacing every negative increment with zero can increase estimated ultimate costs and remove genuine recovery variability rather than merely cleaning the data.
Official syllabus reference: [2]
39. Parameter percentiles versus predictive percentiles
A parameter distribution describes uncertainty about an unknown model quantity, such as a development factor or expected reserve. A predictive distribution describes future observable outcomes after integrating both parameter uncertainty and process variation. Their percentiles answer different questions. To simulate a predictive reserve, draw parameters from their uncertainty distribution and then draw future claims conditional on those parameters. A percentile of expected reserve draws cannot generally be substituted for a percentile of actual future claims. State the target explicitly when reporting confidence, credibility, or prediction intervals.
Worked example: Posterior expected reserves have a 95th percentile of $120 million. After adding future process variation, the predictive 95th percentile is $145 million. The first number concerns uncertainty in the mean; the second concerns actual unpaid-claim outcomes.
Common mistake: Using a parameter confidence interval as a future-payment prediction interval overlooks process risk and can materially understate the relevant tail.
Official syllabus reference: [2]
40. Total expectation and total variance
For future reserve R and uncertain parameters theta, E[R] = E[E[R | theta]]. The law of total variance gives Var(R) = E[Var(R | theta)] + Var(E[R | theta]). The first term is average conditional process variance, while the second captures variability in conditional means. These identities connect simulation, Bayesian prediction, and approximate stochastic reserve formulas. They also provide useful implementation checks. Additional uncertainty about model structure can be represented by extending the conditioning variables, but it is not included automatically when the simulation varies only fitted parameters.
Worked example: Two equally likely parameter states imply mean reserves of 80 and 120, each with conditional variance 100. Overall mean is 100. Between-state mean variance is 400, so predictive variance is 100 + 400 = 500.
Common mistake: Averaging conditional variances without adding variation in conditional means omits parameter uncertainty from the resulting predictive variance calculation.
Official syllabus reference: [2]
41. Moment-matched lognormal reserve approximation
For a positive lognormal reserve with mean m and variance v, set log-scale variance s² = ln(1 + v/m²) and log-scale mean a = ln(m) − s²/2. The pth percentile is exp(a + s zp), where zp is the standard normal percentile. This matches the first two moments but imposes a particular skew and tail shape. It may be unsuitable for reserves with meaningful negative outcomes, multimodality, or different extreme-tail behavior. Compare fitted percentiles with simulation or other plausible distributions rather than treating moment matching as validation.
Worked example: For mean 100 and standard deviation 50, s² = ln(1.25) ≈ 0.2231 and a ≈ 4.4936. Using z0.95 = 1.645 gives a 95th percentile near 194.5, on the original monetary scale.
Common mistake: Adding 1.645 standard deviations to a skewed reserve mean uses a normal approximation and does not reproduce the moment-matched lognormal percentile.
Official syllabus reference: [2]
42. Mixture-distribution quantiles
A predictive mixture combines conditional distributions using their probability weights. Its cumulative distribution is the weighted average of conditional cumulative distributions, but its percentile is generally not the weighted average of conditional percentiles. Calculate the mixture CDF or pool appropriately weighted simulation draws before finding the quantile. Mixtures can represent parameter states, economic scenarios, or alternative models when weights are justified. Quantile definitions matter for discrete outcomes. Report the convention used, especially when a probability jump spans the requested percentile and interpolation would invent an unattainable outcome.
Worked example: A reserve equals 50 with probability 0.80 and 200 with probability 0.20. Its 90th percentile under the smallest-value-reaching-probability convention is 200. Averaging the two state values gives 80, which is the mean rather than that percentile.
Common mistake: Averaging scenario percentiles can miss the percentile of the combined distribution because quantiles do not commute with mixture averaging.
Official syllabus reference: [2]
43. Poisson–gamma Bayesian updating
For N | theta ~ Poisson(e theta) and a gamma prior with shape a and rate b, the posterior is Gamma(a + N, b + e). The prior contributes information about claim intensity, while exposure and observed claims update it. Future predictive counts mix Poisson outcomes over this posterior and therefore include parameter uncertainty. Be explicit about rate versus scale parameterization because their numerical meanings differ. Conjugacy is an instructional and computational convenience, not evidence that claim counts are Poisson or that one shared intensity adequately represents a heterogeneous portfolio.
Worked example: A Gamma(3, rate 2) prior is updated with 12 claims over exposure 4. The posterior is Gamma(15, rate 6), with mean intensity 2.5. For future exposure 2, expected count is five and predictive variance exceeds five.
Common mistake: Using the gamma rate as though it were a scale reverses the posterior mean calculation and can produce seriously incorrect predictions.
Official syllabus reference: [2]
44. MCMC convergence and effective information
Markov chain Monte Carlo approximates a posterior through dependent draws. Inspect multiple chains, trace behavior, convergence diagnostics, and effective sample sizes for quantities that matter, including aggregate reserve and tail parameters. Many stored draws can contain little independent information when autocorrelation is strong. Apparent convergence does not validate the likelihood or prior, and a stable mean does not guarantee stable tail quantiles. Compare results across dispersed initial values and assess Monte Carlo error. Sampling pathologies may indicate weak identification, poor parameterization, or a posterior geometry that requires a different computational approach.
Worked example: An analysis stores 100,000 draws but has effective sample size 300 for a tail parameter. Its reserve mean appears stable, yet repeated chain summaries produce noticeably different 99th percentiles. More informative sampling and tail-specific diagnostics are needed.
Common mistake: Equating the number of stored iterations with independent posterior information ignores autocorrelation and can overstate the precision of reported reserve percentiles.
Official syllabus reference: [2]
45. Posterior predictive reserve generation
Posterior prediction integrates a future-claims model over the posterior distribution of its parameters. For each retained parameter draw, generate future increments conditional on that draw and sum them into the desired reserve quantity. Keep shared parameters common across cells within a simulation so their dependence is preserved. The distribution of conditional expected reserves alone excludes future process variation. Posterior predictive checks compare replicated observed-data features with actual features to diagnose fit; they complement, rather than replace, out-of-sample testing and sensitivity to alternative structural assumptions.
Worked example: One posterior draw contains a high common inflation parameter. All affected future cells should use that same draw before process simulation. Redrawing inflation independently for every cell would remove the common shock and artificially diversify the aggregate reserve.
Common mistake: Generating each cell from an independently selected posterior parameter set can destroy dependence created by portfolio-wide parameters and understate aggregate uncertainty.
Official syllabus reference: [2]
46. Verrall-style expert-opinion integration
Expert opinion can enter a stochastic reserve model through probability distributions on ultimate levels, development behavior, or other clearly defined quantities. Translate elicited medians and intervals into a coherent prior, checking whether those summaries can belong to the selected family. The prior should represent uncertainty, not merely a fixed override of unfavorable data. Verrall's predictive-distribution topic emphasizes incorporating such information within the statistical framework. Avoid double counting when the expert used the same claims data as the likelihood. Sensitivity analysis should show how prior strength changes both central estimates and predictive tails.
Worked example: An expert gives a median ultimate of 100 and a symmetric log-scale 90% interval from 80 to 125. Because 80 × 125 = 100², these summaries are consistent with a lognormal median and symmetric log-quantiles.
Common mistake: Treating an expert estimate derived from the same triangle as independent external evidence can count the observed experience twice.
Official syllabus reference: [2]
47. Bayesian hierarchical partial pooling
Hierarchical reserving gives related cohorts or segments distinct parameters governed by a shared population distribution. Sparse groups borrow information from stronger groups, while credible differences can remain. The degree of pooling depends on within-group information and estimated between-group variation. This mechanism is useful in Bayesian MCMC reserving of the kind studied in Meyers, but the exact model must match the application. Exchangeability is a substantive assumption: groups with different coverage or development mechanisms should not be pooled merely because data are scarce. Check sensitivity to population priors and grouping choices.
Worked example: A small excess segment has only two developed years, while ten related segments have longer histories. A hierarchical tail model stabilizes its estimate toward the population pattern but allows a distinct tail if its own evidence is sufficiently informative.
Common mistake: Assuming partial pooling makes heterogeneous risks comparable overlooks the exchangeability assumption that determines whether borrowing information is appropriate.
Official syllabus reference: [2]
48. Dependence through common reserve drivers
Portfolio reserve dependence can arise from shared inflation, judicial changes, catastrophes, claims operations, and estimated parameters. Model these drivers explicitly where feasible or use a justified dependence structure for marginal reserve distributions. Linear correlation alone does not fully describe joint tail behavior, and independence should not be selected solely because empirical estimates are noisy. In simulation, draw common shocks once per portfolio scenario and conditional segment outcomes afterward. Compare aggregate tails under alternative plausible structures. Diversification is a model result conditional on those assumptions, not an automatic benefit of having several segments.
Worked example: Two liability segments both respond to the same legal-cost shock. A joint simulation applies one annual shock to both, then adds segment-specific variation. Independent shock draws would make simultaneous deterioration too rare and reduce the modeled aggregate tail.
Common mistake: Low historical correlation in a short dataset does not establish independence, especially when rare common shocks drive the outcomes of greatest concern.
Official syllabus reference: [2]
49. Backtesting reserve predictions without look-ahead
A valid backtest reconstructs the information available at an earlier valuation, produces the model's stated prediction, and compares it with the matching subsequent outcome. Distinguish ultimate runoff from next-year payments or one-year reserve deterioration. Evaluate bias, interval coverage, tail misses, and performance by segment rather than relying on one aggregate score. Overlapping tests share claims and calendar conditions, so they are not necessarily independent. Avoid selecting parameters using later observations and then claiming those same outcomes as out-of-sample validation. Changes in portfolio structure also affect the relevance of historical performance.
Worked example: At each year-end from 2018 through 2022, an analyst rebuilds the then-available triangle and predicts next-year payments. Comparing those original intervals with realized payments tests that horizon; using today's revised triangles would introduce information unavailable at the original dates.
Common mistake: Backtesting with fully revised historical data can leak later information into earlier predictions and make a weak model appear better calibrated.
Official syllabus reference: [2]
50. Scenario ranges versus probability intervals
A range of reasonable indications can reflect alternative methods, assumptions, or scenarios without assigning probabilities. A prediction interval instead has a stated probability interpretation under a specified model. Do not relabel the lowest and highest method estimates as particular percentiles without a probability framework. Explain what drives the range and whether its endpoints represent jointly plausible assumptions. Method disagreement can reveal structural uncertainty absent from a single stochastic model. Combining models requires justified weights and attention to dependence and overlapping information, rather than treating each method as independent new evidence.
Worked example: Three defensible methods indicate reserves of 90, 105, and 130. Those numbers describe method variation, not automatically a 90% prediction interval. A separate scenario analysis explains that the high estimate assumes slower settlement and stronger future inflation.
Common mistake: Calling the minimum and maximum of selected methods a confidence interval invents probability coverage that the calculations have not established.
Official syllabus reference: [2]
51. Percentile-based risk-margin calibration
A percentile-based margin can be defined as a selected predictive percentile less a central estimate, commonly the mean when that is the specified basis. The confidence level is a policy or framework input, not a universal actuarial constant. Confirm that the predictive distribution includes the intended risks and uses consistent discounting and expense treatment. Aggregate the distribution before selecting the portfolio percentile. An increase in the margin can result from greater uncertainty even when the mean is unchanged. Report the central estimate, margin, total provision, and key model limitations separately.
Worked example: A stipulated policy targets the 80th percentile. Mean unpaid claims are 250 and the aggregate 80th percentile is 280, in millions. The margin is 30 and provision 280; another confidence target would produce a different margin.
Common mistake: Presenting a chosen percentile as a universal reserving requirement confuses an explicit calibration decision with a generally applicable technical rule.
Official syllabus reference: [2]
52. Cost-of-capital margin mechanics
A cost-of-capital approach values the future cost of holding capital against specified reserve risks. Under a simple formulation, sum each period's required capital times the stipulated capital-cost rate, discounted to valuation. Define the risk measure, runoff pattern, timing, diversification, and eligible risks before calculating. The capital-cost rate is an assumption, not automatically the discount rate or expected investment return. This approach differs from adding a fixed percentile excess at valuation. Results can be sensitive to how capital runs off and whether future capital is recalculated under adverse scenarios.
Worked example: Required capital is 40 at the ends of years one and two, the assumed annual capital cost is 6%, and discounting is 3%. The margin is 2.4/1.03 + 2.4/1.03², approximately 4.59 in matching monetary units.
Common mistake: Multiplying today's capital by a cost rate once ignores the duration and changing size of capital required throughout the liability runoff.
Official syllabus reference: [2]
53. Internal and external systemic reserve risk
A structured risk-margin assessment distinguishes random claim variation from systematic influences that affect many claims together. Internal systemic risks can include model specification, data quality, and operational assumptions; external systemic risks can include economic or legal changes outside the insurer's control. Marshall and colleagues provide the syllabus's risk-margin framework context. The classification helps identify risks missing from a statistical model and prevents indiscriminate diversification assumptions. Quantification should connect each allowance to a mechanism and avoid counting the same inflation or parameter uncertainty in both the base model and a supplementary load.
Worked example: A bootstrap captures cell-level fluctuation and estimated development uncertainty but contains no legal-change scenario. The analyst assesses that external common driver separately, then checks that the proposed allowance does not duplicate an inflation scenario already included in the predictive simulation.
Common mistake: Adding several broadly labeled uncertainty loads without tracing their mechanisms can double count shared risks and obscure what the final margin covers.
Official syllabus reference: [2]
54. Discounted reserves and payment-timing risk
A discounted reserve weights future payments by discount factors appropriate to their timing and the stipulated valuation basis. Even with a fixed ultimate amount, changes in payment speed alter present value. Uncertainty about interest rates, payment amounts, and timing can be correlated, so discounting a mean reserve at a mean duration is not generally equivalent to averaging discounted scenarios. State whether rates are fixed or stochastic and whether the exercise concerns nominal or discounted claims. Do not import discounting into a problem that specifies an undiscounted liability without explaining the basis change.
Worked example: A fixed $100 payment discounted at 5% has present value 95.24 after one year and 86.38 after three years. Equal probabilities give expected present value 90.81, which differs from discounting at the mean two-year timing.
Common mistake: Treating payment timing as irrelevant once ultimate claims are fixed overlooks its direct effect on discounted liability and associated uncertainty.
Official syllabus reference: [2]
55. Quota share versus surplus-share cessions
Quota share applies a specified proportional cession to covered business, while surplus share typically varies the cession by policy according to the relationship between insured amount and retained line, within treaty capacity. Consequently, surplus-share claims cannot always be projected using one portfolio-wide historical cession. Both forms provide proportional risk transfer, but exposure mix and treaty terms determine ceded amounts. Separate claim recoveries from ceding commissions and other premium terms. A uniform quota-share shortcut is valid only when coverage, expenses, periods, and prior ceded payments align with the same percentage.
Worked example: With a retained line of $1 million and adequate surplus capacity, a $4 million insured amount implies a 75% cession under a simple proportional surplus arrangement. A $400,000 covered loss then cedes $300,000, whereas a separate 30% quota share would cede $120,000.
Common mistake: Applying one surplus-share percentage to every policy ignores the policy-specific retained line and can distort ceded reserves after exposure mix changes.
Official syllabus reference: [2]
56. Per-risk and occurrence excess recovery units
Excess treaties apply attachments and limits to a contractual unit, such as an individual risk or a defined occurrence. That unit determines whether losses are combined before the layer function is applied. A catastrophe occurrence can involve many underlying risks, while a per-risk treaty may respond separately to each. Definitions, hours clauses, expense treatment, and inuring arrangements can materially affect recoveries; use the provided wording rather than assuming one standard structure. Reserve analysis needs ultimate losses grouped consistently with the treaty unit and must account for prior recoveries and remaining limits.
Worked example: Two risks each lose $600,000 in one event. A $1 million attachment applied separately to each risk produces no recovery. The same attachment applied to the combined $1.2 million occurrence produces $200,000, assuming adequate limit and matching coverage.
Common mistake: Aggregating all event losses for a per-risk treaty, or separating them under an occurrence treaty, changes the contractual recovery unit.
Official syllabus reference: [2]
57. Aggregate exhaustion and remaining ceded liability
An aggregate limit caps total treaty recovery over the defined period, while an aggregate deductible delays recovery until qualifying losses exceed a threshold. Apply these terms to the contractually defined loss measure after any required occurrence calculations. Remaining ceded liability equals expected ultimate recovery less recoveries already paid or otherwise settled on the selected basis. Because caps are nonlinear, applying them to expected losses can differ from averaging recoveries across scenarios. Track exhaustion using cumulative eligible losses and recoveries, not merely the current gross unpaid reserve.
Worked example: Occurrence recoveries total 8, but the annual aggregate recovery limit is 6, in millions. With 4 already recovered, only 2 remains payable under the treaty. A gross unpaid estimate of 5 does not create additional limit beyond that balance.
Common mistake: Applying an aggregate limit independently to each accident-year reserve component can allow total modeled recoveries to exceed the treaty's actual capacity.
Official syllabus reference: [2]
58. Reinsurance reporting and settlement lag
Reinsurance data can contain multiple delays: the claimant reports to the cedant, the cedant recognizes development, the cedant submits a bordereau or notice, and the reinsurer records or pays recovery. Changes in any stage can alter observed reinsurance triangles without changing underlying ultimate losses. Diagnose lag by cedant, treaty, reporting frequency, and claim size. Payment development may also reflect settlement procedures or disputes. Keep economic claim development distinct from information transmission and cash settlement. Exposure-based or expected-loss methods can help immature periods when observed reports are temporarily unrepresentative.
Worked example: A cedant switches from monthly to quarterly bordereaux. At an interim valuation, reported recoveries fall below the historical pattern even though its internal claim estimates are unchanged. Adding the known reporting pipeline is more defensible than declaring an ultimate improvement.
Common mistake: Interpreting delayed cedant submissions as lower ultimate claims confuses missing information with a genuine reduction in covered loss exposure.
Official syllabus reference: [2]
59. Treaty-year exposure and coverage mapping
Reinsurance reserves require mapping losses to the correct treaty and coverage trigger, such as risks attaching or losses occurring, as specified in the scenario. Underwriting year, accident year, and reporting year need not coincide. Historical experience may also require adjustment for attachment, limit, premium rate, mix, and inflation changes before it informs current treaty estimates. Maintain an explicit mapping from underlying policy or occurrence to contract terms. A development factor from one treaty year can be unsuitable for another if coverage changes alter which claims penetrate the layer or how exposure earns.
Worked example: A policy starts in December and suffers a covered loss the following March. A risks-attaching treaty may associate it with the policy inception year's treaty, while a losses-occurring treaty uses the loss date, subject to the actual contract wording.
Common mistake: Assigning every recovery to the accident-year treaty without checking the coverage trigger can place losses under the wrong terms and limits.
Official syllabus reference: [2]
60. Reinstatement premium and net recovery value
A reinstatement restores exhausted treaty capacity under specified terms and may require additional premium. Model both the extra recovery capacity and the associated premium obligation when valuing the treaty's economic result. Reinstatement charges may depend on the amount restored, time remaining, or other contractual rules; use explicit terms rather than assuming a universal formula. Gross claim recoveries, reinstatement premium, and collection adjustments should remain separately visible. In stochastic analysis, calculate reinstatement effects within each scenario because loss order and exhaustion can influence both available cover and additional premium.
Worked example: A scenario generates $5 million of recoveries using initial and reinstated capacity. The contract's stipulated reinstatement charge is $0.4 million. Economic recovery net of that charge is $4.6 million, before discounting and credit effects, although claim recoveries remain separately reported.
Common mistake: Counting restored capacity without its contractual premium cost overstates the economic benefit of the treaty in scenarios that use reinstatement.
Official syllabus reference: [2]
Official syllabus sources
- Exam 7 - Advanced Estimation of Claims Liabilities | Casualty Actuarial Society
- CAS Exam 7 Fall 2026 current content outline