Build technical fluency in stochastic reserving, excess-layer development, premium liabilities, risk margins, monitoring, specialized ratemaking, reinsurance, and catastrophe risk. This guide contains twenty original practice tasks and sixty distinct technical lessons, with worked calculations and assessable answer criteria. It follows the March/July 2027 GI 301 syllabus supplied as published on October 13, 2026, for upcoming administrations, rather than substituting the former track examination. The supplied evidence does not establish the examination's response format; these written exercises are educational practice. Their counts and suggested times are study allocations, not published examination logistics. Citations establish syllabus relevance; scenario assumptions supply calculation conventions.
What this guide covers
Study GI 301 with 20 original worked tasks and 60 concepts covering reserving, pricing, reinsurance and catastrophe risk.
Stochastic Reserving
1. Chain ladder conditional-mean assumptions
A stochastic chain ladder model makes assumptions beyond multiplying cumulative claims by selected factors. In the usual Mack formulation, the next cumulative amount has conditional expectation equal to the current amount multiplied by an age-specific factor. Accident years are independent, and conditional variance follows a specified relationship with current cumulative claims. These assumptions support estimation of prediction uncertainty without requiring a complete claims distribution. They can fail when development depends on claim size, calendar conditions, or operational changes. Diagnose those features before interpreting the model's standard error as a complete uncertainty measure.
Worked example: Two accident years have identical cumulative paid claims at 24 months. The model gives them the same expected next-age multiplier. If one year contains substantially more unresolved large claims, that common conditional development assumption may require revision.
Common mistake: A stable average development factor does not establish accident-year independence or validate the assumed conditional variance relationship.
Official syllabus reference: [2]
2. Volume weighting and factor normalization
The conventional volume-weighted age-to-age factor is the sum of next-age cumulative claims divided by the sum of corresponding current-age claims. It is also a weighted average of individual development ratios, with current cumulative claims supplying the weights. Correspondence matters: both sums must contain the same accident years. This estimator differs from an arithmetic average of ratios, which gives small and large observations equal influence. Selecting weights changes the implicit model and can materially affect immature-year projections when historical observations vary greatly in size or development experience.
Worked example: One year develops from $100 to $150 and another from $900 to $990. The volume-weighted factor is $1,140/$1,000 = 1.14. The arithmetic average is 1.30, giving much greater influence to the smaller year's unusually high ratio.
Common mistake: Do not combine all available next-age observations with a different set of current-age observations when calculating the factor.
Official syllabus reference: [2]
3. Heteroscedasticity in development ratios
Under a variance specification Var(C_next | C) = sigma²C, the development ratio C_next/C has conditional variance sigma²/C. Larger cumulative observations therefore have less variable ratios, which explains the role of volume weighting. Residuals must be standardized for this changing variability before comparing observations of different sizes. An unscaled residual plot may otherwise make large accident years appear systematically problematic simply because they contain more dollars. Alternative variance powers can be investigated when standardized residual dispersion still changes with volume, but changing variance alone does not repair a misspecified mean.
Worked example: Suppose sigma² is 9 and current cumulative claims are measured consistently. A year with C = 100 has ratio variance 0.09, while a year with C = 900 has ratio variance 0.01. Equal raw ratio deviations carry different diagnostic significance.
Common mistake: Similar dollar residuals across differently sized accident years do not imply similar standardized errors or support an equal-variance ratio model.
Official syllabus reference: [2]
4. Residual plots as structural diagnostics
Residual analysis compares observed development with the fitted conditional mean after allowing for the model's variance structure. Plot standardized residuals against development age, accident year, calendar year, fitted amount, and cumulative volume. Curvature can indicate an unsuitable mean relationship; changing spread can indicate a variance problem; diagonal clustering can suggest calendar effects. A satisfactory overall average residual can conceal substantial structure because positive and negative deviations cancel. Residual diagnostics guide alternative models, but isolated outliers should also be reconciled to claim information before being deleted or assigned a special adjustment.
Worked example: A model has nearly zero average residual overall, yet early development ages are mostly negative and late ages mostly positive. The cancellation hides a systematic development mismatch. An age-specific adjustment deserves investigation before the reserve distribution is used.
Common mistake: A near-zero overall residual mean does not demonstrate that the fitted development pattern is unbiased within important subsets.
Official syllabus reference: [2]
5. Testing proportional development with an intercept
Chain ladder's proportional mean relationship can be challenged by fitting a regression that permits an intercept: expected next cumulative claims equal a + bC. A material intercept can indicate an additive development component, differences related to claim volume, or other omitted structure. It provides a diagnostic comparison rather than automatic permission to replace the model. Assess residual behavior, uncertainty, and predictive performance, especially when the sample is small. An intercept may also reflect data problems or changing composition, and extrapolating it to very small or very large portfolios can produce implausible results.
Worked example: Several mature cohorts each add approximately $200,000 of reopening payments despite very different current cumulative totals. A proportional-only model struggles with that additive pattern. An intercept model and investigation of reopening practices offer a more targeted comparison.
Common mistake: A statistically noticeable intercept identifies a departure worth investigating; it does not by itself prove a particular operational explanation.
Official syllabus reference: [2]
6. Calendar effects and dependent accident years
A calendar shock can affect observations from several accident years simultaneously. Inflation, settlement initiatives, judicial developments, and claim-system changes may create diagonal residual patterns in a development triangle. Such effects can challenge both the mean projection and accident-year independence. Because calendar year equals accident year plus development age under a chosen indexing convention, unrestricted effects for all three dimensions create identification problems. Use defensible constraints, parsimonious interventions, or explicit external information. Distinguish an effect on payment timing from an effect on ultimate cost before translating the diagnosis into reserve changes.
Worked example: A settlement campaign increases payments across four accident years during one calendar year. Paid development accelerates, but case estimates fall correspondingly. Treating every positive payment deviation as higher ultimate severity would confuse timing with a change in total claim cost.
Common mistake: Adding unrestricted accident-year, development-year, and calendar-year parameters can create identification problems rather than a uniquely interpretable fitted explanation.
Official syllabus reference: [2]
7. Process variance in unpaid-claims prediction
Process variance describes randomness in future claims even when model parameters are known. In a one-period Mack-style model, conditional variance is sigma² times the latest cumulative amount. Across several future periods, uncertainty propagates through subsequent development factors, so early future deviations can influence later cumulative amounts. Process variance belongs in a predictive distribution, whereas a confidence interval for the conditional mean addresses a different target. The distinction matters when communicating reserve adequacy: precise estimation of average development cannot remove the randomness of actual settlements and claim emergence.
Worked example: With current cumulative claims of 500 and sigma² of 8 in consistent monetary units, next-period process variance is 4,000. Even if the development factor were known exactly, the next cumulative outcome would still vary around its conditional expectation.
Common mistake: A large historical data set can reduce parameter uncertainty substantially while leaving meaningful process uncertainty in future claim payments.
Official syllabus reference: [2]
8. Shared parameter uncertainty and reserve covariance
Parameter uncertainty arises because development factors or curve parameters are estimated from finite data. When several accident years use the same estimated factor, their forecast errors share estimation uncertainty even if their underlying claim processes are independent. Aggregate prediction variance therefore generally requires covariance terms as well as individual variances. In a simple one-factor setting, parameter variance for a forecast C times f_hat is C²Var(f_hat). For an aggregate forecast using the same factor, the combined sensitivity is the sum of the cumulative amounts, which must be squared before multiplying by factor variance.
Worked example: Two years with cumulative amounts of 100 and 200 use one estimated factor with variance 0.01. Aggregate parameter variance is (100 + 200)² × 0.01 = 900. Adding individual parameter variances gives only 500 and omits their shared estimation covariance.
Common mistake: Independent accident-year claim processes do not make forecasts independent when those forecasts share development factors estimated from the same data.
Official syllabus reference: [2]
9. Residual bootstrap and leverage adjustment
A residual bootstrap approximates sampling uncertainty by resampling appropriately standardized residuals, constructing pseudo-observed data, and refitting the reserving model. Fitted residuals are often less variable than underlying errors because parameter estimation absorbs some variation. Leverage or degrees-of-freedom adjustments help address that shrinkage. Residuals may also require centering and exclusion of structurally uninformative cells. Exchangeable resampling is defensible only after checking that the residual pool lacks systematic patterns. Bootstrap output inherits the fitted model's structural assumptions and cannot automatically represent omitted calendar shocks or an unsuitable tail specification.
Worked example: An analyst resamples raw dollar residuals from a triangle dominated by large mature cohorts. The resulting pseudo-data distort variability for small cohorts. Standardizing by modeled variance and applying suitable leverage adjustments makes the residual pool more consistent with the assumed error structure.
Common mistake: Resampling observed residuals without considering fitted variance, leverage, and exchangeability can produce a precise-looking but poorly calibrated bootstrap distribution.
Official syllabus reference: [2]
10. Two-stage predictive bootstrap
A predictive bootstrap separates parameter estimation from future process variation. First, generate pseudo-history and refit the model to obtain a new set of projected means. Second, simulate future incremental claims around those means using the selected process distribution. Repeating both stages produces a predictive reserve distribution. Stopping after refitting generally describes variability in estimated conditional means rather than actual future outcomes. Distributional choices also matter: a process family should accommodate the modeled variance and data characteristics, including whether negative increments are possible. Tail and model uncertainty may still require separate treatment.
Worked example: Across refitted bootstrap histories, an insurer's mean reserve varies between $9 million and $11 million. Simulating future claims around each mean produces a wider distribution. That additional spread represents process randomness rather than further uncertainty in the fitted parameters.
Common mistake: A distribution of refitted reserve means is not automatically a predictive distribution for the claims that will actually emerge.
Official syllabus reference: [2]
11. Clark growth curves: LDF and Cape Cod means
Clark-style parametric reserving models describe cumulative emergence through a growth function G(t), approaching one as development matures. Expected increments are linked to differences in G between ages. In an LDF formulation, accident-year ultimate amounts are estimated as parameters. In a Cape Cod formulation, ultimate expectations are tied to an exposure measure, commonly appropriately adjusted premium, through an expected loss ratio. The distinction changes how immature years borrow information. Growth-family selection, exposure adequacy, and treatment of within-period development ages affect results; a smooth fitted curve does not remove these modeling choices.
Worked example: Two immature years have little paid data but reliable adjusted premiums. A Cape Cod structure connects their ultimate expectations to those premiums and a common loss ratio. An LDF structure instead gives each year its own ultimate parameter, potentially allowing greater instability.
Common mistake: Cape Cod does not make premium an automatically valid exposure measure when rate changes or underwriting mix have not been addressed.
Official syllabus reference: [2]
12. Delta-method linearization of a nonlinear reserve estimate
A reserve calculated from fitted development parameters is generally a nonlinear function R(θ). The delta method approximates its parameter variance by gᵀΣg, where g is the gradient of R evaluated at the fitted parameters and Σ is their estimated covariance matrix. This follows from a first-order Taylor expansion: small parameter changes produce approximately linear reserve changes. Sensitivity signs determine whether covariance increases or decreases variance. The result measures uncertainty in the fitted reserve mean; future process variability requires separate treatment. Strong curvature or imprecise parameters can make the approximation unreliable.
Worked example: At age one year, R(U,k)=Ue^(−k), measured in $000. With U=1,000 and k=ln(2), R=500 and sensitivities are (0.5,−500). Given Var(U)=400, Var(k)=0.0004 and Cov(U,k)=0.2, parameter variance is 0.5²×400+500²×0.0004+2×0.5×(−500)×0.2=100. Thus the parameter standard error is $10,000.
Common mistake: Adding only the two marginal variance contributions gives 200 here. Omitting the covariance term overstates parameter uncertainty and does not produce a complete prediction variance.
Official syllabus reference: [2]
Development Analysis for Excess Limits and Layers
13. Layer losses as differences of limited losses
For attachment a and layer width l, the layer payment is min(max(X − a, 0), l). Equivalently, it is min(X, a + l) minus min(X, a). This identity supports claim-level calculations and expected layer costs through differences of limited expected values. The transformation must respect contract definitions, including whether expenses erode limits and whether limits apply per claim or occurrence. Applying it to an aggregate average generally fails because the payment function is nonlinear. Individual loss development can cross the attachment or exhaust the layer at different times.
Worked example: A $200,000 layer above $100,000 pays $0 on a $70,000 claim, $80,000 on a $180,000 claim, and $200,000 on a $500,000 claim. The corresponding differences of capped ground-up amounts produce exactly the same three payments.
Common mistake: Applying a layer payment function to total portfolio losses does not reproduce a sum of per-claim layer payments.
Official syllabus reference: [2]
14. Attachment penetration and layer-specific development
Excess-layer development reflects both changes in claim size and the timing at which claims penetrate attachment points. A claim initially below attachment can later create its first layer loss, while a claim already above exhaustion can grow without further layer impact. Higher layers often emerge later and have fewer observations, but the actual pattern depends on reporting, settlement, and contract structure. Ground-up development factors therefore require validation before use on a layer. Build consistent layer triangles where possible and distinguish current reported layer amounts from expectations of future attachment penetration.
Worked example: A reported claim increases from $400,000 to $650,000 under a layer attaching at $500,000. The layer moves from zero to $150,000 even though ground-up development is only 62.5%. A layer factor cannot be inferred from this claim's initial zero layer amount.
Common mistake: A claim currently below attachment can still contribute substantial ultimate layer loss, especially when its development remains uncertain.
Official syllabus reference: [2]
15. Inflation leverage in fixed excess layers
With fixed nominal attachments and limits, ground-up severity inflation can produce a larger proportional increase in excess losses. Inflation pushes additional claims across attachment and increases payments on claims already within the layer. Exhaustion limits the effect on very large claims, so leverage is not uniform across all layers or severity distributions. Trend claims to a common cost level before applying prospective treaty terms when the analysis requires it. Applying a ground-up percentage directly to historical layer losses can miss both newly penetrating claims and changes in exhaustion.
Worked example: A $600,000 claim produces $100,000 in a layer attaching at $500,000 with ample remaining limit. Ten percent severity inflation raises the claim to $660,000 and layer loss to $160,000. The layer increase is 60%, despite 10% ground-up inflation.
Common mistake: Ground-up severity trend and excess-layer loss trend are generally different because fixed attachments change the proportional effect of inflation.
Official syllabus reference: [2]
16. Expected-emergence methods for sparse layers
Sparse excess layers can make pure development projections unstable, especially when current reported layer losses are near zero. An expected-emergence approach combines observed layer amounts with an independently selected ultimate expectation and a layer-specific unreported proportion. A Bornhuetter–Ferguson-style estimate is reported layer loss plus expected ultimate layer loss multiplied by the expected unreported fraction. The external expectation may use exposure analysis, severity distributions, or comparable portfolios. Independence is conceptual rather than automatic: an expectation derived from the same emerging claims can double-count information. Sensitivity to layer emergence assumptions should be explicit.
Worked example: Reported layer loss is $200,000, independently expected ultimate loss is $1 million, and the selected reported proportion is 30%. A reported-based Bornhuetter–Ferguson estimate is $200,000 + 70% × $1 million = $900,000, rather than $200,000/30% = $666,667.
Common mistake: A zero reported layer amount does not justify a zero ultimate estimate when exposure and expected future attachment penetration remain material.
Official syllabus reference: [2]
Premium Liabilities
17. Coverage-period boundaries for premium liabilities
Premium liabilities concern obligations arising from coverage that remains after the valuation date under the adopted measurement boundary. Unpaid-claims liabilities concern insured events already incurred, including events not yet reported. Settlement dates do not determine this distinction: a future-covered event may be paid years later, while an already-incurred claim may be paid tomorrow. Estimation therefore requires separating exposure periods and identifying which future premiums, expenses, claims, and recoveries belong to remaining coverage. Contract boundaries and accounting presentation vary, so establish the exercise or reporting definition before calculating a numerical liability.
Worked example: At December 31, an annual policy has six months of coverage remaining. A November accident reported in January belongs to unpaid claims. An accident occurring the following March belongs to the remaining-coverage analysis, even if its payment occurs much later.
Common mistake: Classifying liabilities solely by when cash is paid confuses already-incurred claims with obligations generated by future insured events.
Official syllabus reference: [2]
18. Future premiums, expenses, and the UPR comparison
A prospective premium-liability calculation assembles future cash flows attributable to remaining coverage. Expected claims, claim-handling costs, maintenance expenses, future premium receipts, and related reinsurance must be treated consistently. Acquisition costs already paid are not automatically future outflows, although accounting comparisons can require separate adjustments. An unearned premium reserve allocates written premium across coverage periods; it is not necessarily the same as the economic cost of remaining obligations. Compare the two only after aligning gross or net basis, expense treatment, discounting, and any margin requirements. Avoid deducting an accounting reserve as though it were a receipt.
Worked example: Remaining coverage is expected to generate $500,000 of claims and expenses, with $80,000 of future premium receipts. Before discounting and other adjustments, net future cost is $420,000. An existing $450,000 unearned premium balance is a comparison amount, not another deduction.
Common mistake: Subtracting both future premium receipts and the full unearned premium reserve can double-count premium financing in a prospective cash-flow calculation.
Official syllabus reference: [2]
19. Seasonality and discounted coverage cash flows
Remaining exposure may not be proportional to the unexpired fraction of policy time. Seasonal hazards, changing insured values, and uneven activity can concentrate expected claims in particular months. Estimate event incidence over remaining coverage, then project settlement and expense timing to discount the resulting cash flows consistently. Reinsurance premiums and recoveries may follow different timing patterns and contract boundaries. A simple pro rata premium allocation can therefore differ from prospective liability even when annual pricing is adequate. Discounting should follow the specified valuation basis, rather than an assumed universal rate or settlement convention.
Worked example: Half of a property's annual policy term remains, but that half includes most of its modeled storm exposure. Allocating exactly half the annual expected claims understates remaining risk. The analyst distributes claims seasonally and then applies the expected settlement-delay pattern.
Common mistake: An unexpired time fraction is not automatically the correct proportion of annual expected claims when exposure is strongly seasonal.
Official syllabus reference: [2]
Risk Margins for Unpaid Claims
20. Process, parameter, and model uncertainty
Process uncertainty is randomness in future outcomes conditional on a model and its parameters. Parameter uncertainty concerns estimating those parameters from finite information. Model uncertainty concerns whether the structure, assumptions, or selected method adequately represents the business. These sources call for different tools: conditional simulation, statistical estimation, and structured alternative-model or scenario analysis. A fitted model's prediction error can include process and parameter components while excluding important model risk. A risk-margin framework should identify what each measure covers and avoid adding overlapping allowances as if all components were independently measured.
Worked example: A reserve simulation varies future payments and refits development factors, capturing process and parameter effects. It retains the same settlement pattern in every run. A plausible permanent settlement slowdown therefore needs an additional model-sensitivity analysis rather than being assumed included.
Common mistake: A stochastic reserve distribution does not automatically cover every uncertainty merely because it was generated through a large number of simulations.
Official syllabus reference: [2]
21. Internal and external systemic uncertainty
Systemic uncertainty can move many claims or segments together, limiting diversification. Internal sources arise from the insurer's own systems and practices, such as case-reserving changes, data errors, or claims-management shifts. External sources include inflation, legal developments, and broader social or economic conditions. This distinction helps organize investigation and scenario construction; it does not imply that either category is necessarily unquantifiable. Evaluate affected portfolios, duration, direction, and dependence. An uncertainty register should connect each source to an estimated impact or sensitivity and identify whether existing stochastic models already capture any part of it.
Worked example: A claims-system conversion understates outstanding amounts across several liability products, while unexpected repair inflation affects motor and property. The first is an internal systemic source and the second external. Both may create correlated reserve movements that independent segment simulations miss.
Common mistake: Diversifying all segment errors as independent can understate uncertainty when a common operational or economic source affects several segments simultaneously.
Official syllabus reference: [2]
22. Quantile and tail-mean risk margins
A quantile-based margin is a selected predictive quantile minus the expected liability. It targets a probability of sufficiency under the modeled distribution. A tail-mean measure also considers the severity of outcomes beyond a specified quantile, making it sensitive to upper-tail magnitude. Discrete distributions require an explicit quantile and tail-averaging convention, particularly when probability mass sits at the cutoff. Neither approach selects its own confidence level or validates the underlying model. Explain the valuation purpose, distributional assumptions, and sensitivity to tail changes when translating a statistical measure into a chosen margin.
Worked example: Two portfolios both have a 90th-percentile unpaid amount of $12 million and mean of $10 million. One has much larger losses in its worst 5% of outcomes. Their 90% quantile margins match, but an upper-tail mean measure can distinguish the severity difference.
Common mistake: Equal quantiles at one confidence level do not imply equal tail severity or equivalent risk under every margin method.
Official syllabus reference: [2]
23. Cost-of-capital margins and runoff duration
A cost-of-capital margin discounts future charges for capital supporting uncertain liability runoff. Its components are a projected capital path, an annual charge rate, payment timing, and discount factors. The capital balances themselves are not expenses to be added to the liability; the charges compensate for providing those balances over time. Longer runoff can create a larger margin even with the same initial capital. Results depend on how future capital is estimated and which risks it covers. State whether the approach includes diversification, future management actions, and changes in uncertainty as claims settle.
Worked example: Two portfolios each initially require $50 million of capital. One releases nearly all capital after one year, while the other requires substantial support for five years. With identical annual charge and discount rates, the longer runoff generally produces the larger present value of capital charges.
Common mistake: Multiplying only today's capital by one annual charge rate omits future support costs when the liability remains uncertain for several years.
Official syllabus reference: [2]
24. Correlation and diversification of reserve uncertainty
For two reserve deviations, aggregate variance is sigma_A² + sigma_B² + 2rho sigma_A sigma_B. Correlation therefore determines how much individual uncertainty diversifies. Summing standard deviations assumes perfect positive linear dependence; summing variances assumes zero covariance. Neither is universally appropriate. Historical estimates can be unstable and may omit shared future shocks, so supplement them with source-based reasoning and sensitivity analysis. A variance-based aggregation does not fully describe tail dependence or nonlinear risk measures. If a margin is allocated back to segments, distinguish total aggregation from the chosen allocation rule.
Worked example: Segments have standard deviations of $30 million and $40 million, with correlation 0.25. Aggregate standard deviation is sqrt(900 + 1,600 + 600) = $55.68 million. It lies between the independent result of $50 million and the perfectly correlated result of $70 million.
Common mistake: Using zero correlation because historical estimates are noisy treats uncertainty about dependence as evidence that dependence is absent.
Official syllabus reference: [2]
Monitoring Results
25. Expected emergence from a frozen valuation
Actual-versus-expected monitoring begins with expectations fixed at an earlier valuation date. Project payments, reported claims, or other measures over a clearly specified future interval, then compare actual emergence on the same basis. Changing the expected benchmark after observing actual results obscures forecast performance. Consistency includes gross versus net treatment, expenses, currency, claim definitions, and exposure periods. Present monetary deviations alongside ratios because they answer different questions. Segmenting by accident year, development age, and business type helps distinguish isolated movements from broad departures and prevents favorable offsets from concealing adverse experience.
Worked example: A December valuation predicts $8 million of first-quarter payments. Actual payments are $8.8 million. The ratio is 110% and the deviation $0.8 million. Replacing the original expectation with a March-updated $8.7 million forecast would no longer measure the December forecast's performance.
Common mistake: Updating the expected benchmark with information from the monitoring period can make a weak original forecast appear artificially accurate.
Official syllabus reference: [2]
26. Reserve runoff and ultimate-estimate reconciliation
For the same closed cohort of claims and a consistent valuation basis, the prior unpaid estimate should equal subsequent payments plus the new unpaid estimate if ultimate expectations are unchanged. The difference between those quantities measures development in the selected ultimate estimate. Payment acceleration alone can increase actual paid claims while reducing remaining reserves by an offsetting amount. Reconciliation must separately identify new business, currency changes, discount unwind, and other basis changes. This bridge connects emergence monitoring with reserve performance and prevents a payment-only comparison from being interpreted as a complete measure of deterioration.
Worked example: Prior unpaid claims were $10 million. During the year, $3 million is paid and the new unpaid estimate is $8 million. Payments plus remaining reserves total $11 million, indicating $1 million of adverse development for the unchanged cohort and basis.
Common mistake: Payments above expectation do not necessarily imply adverse ultimate development when the remaining unpaid estimate falls by an offsetting amount.
Official syllabus reference: [2]
27. Frequency–severity decomposition with exposure mix
Aggregate claim cost can be decomposed into exposure, claim frequency per exposure, and average claim severity. Monitoring these components helps identify whether a cost deviation reflects more business, more claims, or more expensive claims. Mix changes can alter aggregate frequency and severity even when each underlying class performs as expected. Compare actual results with expectations using actual exposure where appropriate, and retain a separate exposure-volume bridge. Claim maturity, large-loss treatment, and reporting delays must be aligned before comparing average severity. Component analysis diagnoses mechanisms but does not establish causation without supporting information.
Worked example: Expected frequency is 0.05 and severity $4,000. Actual exposure rises from 1,000 to 1,200 units with unchanged performance. Expected claims on actual exposure are $240,000, so comparing actual $240,000 with the original $200,000 budget would misidentify growth as deterioration.
Common mistake: An increase in aggregate average severity can arise from a shift toward higher-severity classes even when every class remains on expectation.
Official syllabus reference: [2]
Specialized Ratemaking Topics
28. Classification pure premiums and confounding
A class pure premium is expected loss per exposure unit. Simple observed class loss costs can suggest relativities, but correlated rating variables can confound those comparisons. A class concentrated in high-cost territories may appear intrinsically expensive even when territory explains the difference. Multivariate classification separates effects under an explicit model structure, such as multiplicative class and territory factors. Exposure definitions, loss development, trend, credibility, and the selected base level must be consistent. Classification effects describe modeled cost differences; they do not automatically establish causation or determine whether a variable is appropriate for use.
Worked example: Class A has observed loss cost of $300 and Class B $200. Most Class A exposures are in a territory with twice the underlying cost. Adjusting for territory may substantially reduce or reverse the apparent class difference before rates are selected.
Common mistake: A raw loss-cost ratio can reflect unequal distributions of other rating characteristics rather than the isolated effect of the class variable.
Official syllabus reference: [2]
29. Minimum-bias iterations and identifiability
Minimum-bias methods iteratively update rating factors so fitted premiums or losses satisfy selected balance equations. In a multiplicative model, one factor set is updated while others are held fixed, then the process cycles through variables. The exact equations depend on the chosen criterion, so a ratio-of-totals update should not be assumed interchangeable with every least-squares or likelihood procedure. Multiplicative factors require normalization because scaling one set upward and another downward can leave fitted cell values unchanged. Check convergence in fitted values and objective behavior, not merely whether one displayed factor changes little.
Worked example: A class update gives amounts of $120 and $90 while territory factors remain 1 and 2. Representing the result with base $120 and class relativities 1 and 0.75 preserves all cell fits. Further territory updates may still change the eventual solution.
Common mistake: One completed factor update is not a converged multivariate solution, and normalization does not itself improve the fitted loss estimates.
Official syllabus reference: [2]
30. GLM links, exposure offsets, and distributions
A generalized linear model connects a response mean to a linear predictor through a link function and specifies a mean–variance relationship through a distributional family. For Poisson claim counts with a log link, log exposure is an offset with coefficient fixed at one, so expected count equals exposure times modeled frequency. Modeling severity or aggregate loss cost requires a suitable response and variance structure rather than reusing the count model automatically. Log-link coefficients exponentiate into multiplicative effects, conditional on other variables. Overdispersion, zero patterns, and sparse cells should influence diagnostics and uncertainty estimates.
Worked example: A count model predicts frequency 0.08 per vehicle-year. With an exposure offset, 250 vehicle-years produce expected count 20. Doubling exposure to 500 doubles expected count to 40 while leaving the modeled per-exposure frequency unchanged, assuming the same rating characteristics.
Common mistake: Treating exposure as an ordinary fitted predictor allows an unintended exposure elasticity instead of enforcing proportional expected counts.
Official syllabus reference: [2]
31. Interaction effects and predictive validation
An interaction allows one rating variable's effect to depend on another. In a log-link model, a class-by-territory interaction changes the multiplicative territory effect across classes. Interactions can address genuine structure but consume information and may become unstable in sparse cells. Evaluate their contribution with suitable model comparisons, residual diagnostics, and predictive validation on data not used for fitting. Temporal validation is especially useful when rates will apply prospectively. Better in-sample fit alone does not justify complexity, and statistical significance does not resolve operational feasibility, stability, or permitted use of a classification variable.
Worked example: Territory doubles expected cost for private vehicles but increases commercial-vehicle cost by only 20%. A main-effects model forces a common territory multiplier. An interaction can represent the difference, provided enough data support both class–territory combinations and predictive improvement persists.
Common mistake: An interaction that reduces training error can still worsen prospective predictions when its apparent pattern is driven by sparse-cell noise.
Official syllabus reference: [2]
32. Risk-classification variable selection under ASOP 12
Risk-classification analysis considers how characteristics distinguish expected outcomes for the intended purpose. ASOP 12 is a named syllabus reading addressing actuarial risk-classification work; its inclusion does not make every variable legally permissible in every jurisdiction. Evaluate statistical relationships, data quality, practicality, potential adverse selection, and consistency with the classification system's purpose. Distinguish predictive association from a causal claim, and consider whether effects remain stable across relevant populations. Applicable law and implementation constraints are separate requirements. A defensible selection explains why the characteristic contributes useful classification information and how limitations affect its interpretation.
Worked example: A candidate variable predicts higher claim frequency but is inconsistently recorded and strongly correlated with an existing variable. The actuary evaluates its incremental contribution, recording reliability, and permitted use before recommending it, rather than treating a small p-value as sufficient justification.
Common mistake: Syllabus inclusion of ASOP 12 does not establish that a particular rating variable is lawful or suitable in every market.
Official syllabus reference: [2]
33. Credibility as a blend of risk and collective information
Credibility combines an individual risk's observed experience with a collective expectation. A common form is Z times individual experience plus 1 − Z times the collective estimate. Larger or more stable experience generally supports greater weight, but Z depends on the selected credibility model rather than an arbitrary preference. Under a Bühlmann-style structure, Z = n/(n + K), where K reflects within-risk variation relative to between-risk variation under the model. Align exposure, development, trend, and coverage before blending. Credibility reduces sampling noise but cannot correct systematically mismatched experience bases.
Worked example: A risk's adjusted loss cost is $140, the collective expectation is $100, and credibility is 0.25. The blended estimate is $110. The limited weight recognizes that a small individual history may contain substantial random variation despite its adverse observed result.
Common mistake: Increasing credibility because an observed result is unusually adverse confuses the amount of deviation with the statistical reliability of the experience.
Official syllabus reference: [2]
34. Loss elimination ratios for ordinary deductibles
For an ordinary per-claim deductible d and unlimited ground-up severity X, the insurer pays (X − d)+. Expected retained loss per ground-up claim is E[X] − E[min(X,d)]. The loss elimination ratio is E[min(X,d)]/E[X], using a consistent comparison basis. Multiplying retained severity by ground-up frequency gives expected insured loss cost without separately estimating the positive-payment frequency. If coverage also has a limit, adjust the payment function and denominator accordingly. Deductible credits based on losses should not automatically reduce fixed expenses or other premium components by the same percentage.
Worked example: Ground-up expected severity is $12,000 and expected loss below a $2,000 deductible is $1,500. Expected insurer payment is $10,500 per ground-up claim and the loss elimination ratio is 12.5%. With frequency 0.04, insured loss cost is $420 per exposure.
Common mistake: A deductible's face amount is not its expected loss reduction because some claims are smaller than the deductible.
Official syllabus reference: [2]
35. Per-claim deductibles versus annual aggregate deductibles
A per-claim deductible applies separately to each covered claim, while an annual aggregate deductible applies to the specified annual loss total. Their expected costs depend on different distributions. Per-claim pricing can use individual severity and frequency; aggregate-deductible pricing generally requires the annual aggregate loss distribution, including count variability and dependence. Contract order also matters when per-claim limits or deductibles combine with an aggregate provision. Applying a deductible to expected annual losses is not equivalent to averaging scenario payments because the positive-part function is nonlinear. Define the exact covered aggregate before calculating.
Worked example: Two claims of $8,000 each produce $6,000 total payment after a $5,000 per-claim deductible. A single $5,000 annual aggregate deductible instead produces $11,000 payment, assuming no other limits. Identical deductible labels therefore conceal substantially different payment structures.
Common mistake: An annual aggregate deductible cannot generally be priced by multiplying a per-claim deductible credit across the expected number of claims.
Official syllabus reference: [2]
36. Increased-limit factors from limited expected values
For first-dollar coverage capped at limit L, expected claim payment is the limited expected value E[min(X,L)]. With unchanged frequency and comparable coverage, the loss increased-limit factor from basic limit B to L is E[min(X,L)]/E[min(X,B)]. A severity distribution or credible empirical excess analysis supplies these quantities. Tail assumptions become increasingly influential at high limits, and historical limits can censor the data needed for estimation. Expense treatment and risk loads can make premium factors differ from loss factors. Deductibles, aggregate limits, and expense erosion require explicit modifications to the payment function.
Worked example: Limited expected severity is $15,000 at the basic limit and $24,000 at a higher limit. The loss factor is 1.60. If basic-limit expected loss cost is $600, the higher-limit expected loss cost is $960 before expenses and risk loads.
Common mistake: Multiplying every component of a basic-limit premium by a severity-based factor can incorrectly scale fixed expenses and unrelated charges.
Official syllabus reference: [2]
37. Premium equations with fixed and variable expenses
When variable expenses are a proportion v of premium and fixed expenses are F per exposure, a simple premium equation is P = expected loss cost + F + vP, before any additional specified loads. Solving gives P = (expected loss cost + F)/(1 − v). This differs from multiplying cost by 1 + v because the expense percentage applies to premium itself. Deductible and limit changes should first alter the loss component, with expense components adjusted only when their drivers change. State how profit or contingency loads are defined before extending the equation.
Worked example: Expected loss cost is $700, fixed expenses are $100, and variable expenses are 20% of premium. Required premium is $800/0.80 = $1,000. Adding 20% to $800 gives $960, whose $192 variable expense leaves only $768 for the $800 costs.
Common mistake: A percentage-of-premium expense allowance requires solving the premium equation; adding that percentage to costs uses the wrong denominator.
Official syllabus reference: [2]
38. Claims-made maturity and the reporting-delay distribution
Claims-made coverage generally responds to claims made within a specified reporting period, subject to retroactive dates and other terms. Early policy years may cover a shorter history of eligible incidents, producing lower expected reported losses than mature coverage. Maturity factors reflect how the eligible incident window interacts with reporting delays. They are distinct from settlement development on claims already reported. Derive or select factors using consistent incident, report, and policy definitions, and consider changes in reporting behavior or exposure. A maturity factor is not a universal discount schedule applicable to every insured or profession.
Worked example: A newly insured professional has a retroactive date at policy inception. During the first year, only incidents after that date can generate covered reports. Several renewals later, reports may arise from a longer eligible incident history, raising expected claims-made cost toward maturity.
Common mistake: Claims-made maturity is not simply the percentage of reported claims that have been paid by a particular development age.
Official syllabus reference: [2]
39. Retroactive dates and extended reporting coverage
A retroactive date limits the incident history eligible under a claims-made contract. An extended reporting provision can allow reporting after ordinary coverage ends for eligible earlier incidents, according to the contract's terms. It generally does not create unrestricted coverage for new incidents after termination. Pricing requires the distribution of reporting delays, eligible exposure history, remaining unreported claims, and any reporting-window limit. Distinguish extended reporting coverage from prior-acts coverage when an insured changes carriers. Exact triggers and conditions are contractual, so the payment eligibility function must be established before applying maturity or tail factors.
Worked example: A policy ends December 31 and has an extended reporting period for eligible earlier incidents. A November incident first reported the following February may qualify under the stated terms. An incident occurring in February is outside the earlier incident window despite being reported during the extension.
Common mistake: An extended reporting period does not automatically insure new professional activity performed after the underlying policy's incident-eligibility period ends.
Official syllabus reference: [2]
40. Experience-rating claim limits and expected-loss consistency
Experience-rating plans often limit individual claim contributions to reduce the influence of unusually large, volatile losses. Some distinguish primary and excess components because frequency and severity can carry different predictive information. The experience measure and expected losses must be defined on compatible bases; otherwise the actual-to-expected comparison is biased by construction. Development, trend, exposure changes, and the experience period also require alignment. A modification adjusts prospective pricing under the plan and is not a direct repayment of past losses. Use the specified plan formula rather than assuming one universal experience-rating structure.
Worked example: Unlimited actual losses are $400,000, but the plan's claim limits produce ratable losses of $250,000. If expected ratable losses are $200,000, the relevant ratio is 1.25. Comparing the unlimited $400,000 with the limited expectation would exaggerate the plan-defined adverse experience.
Common mistake: Comparing limited actual losses with an unlimited expected-loss amount creates a mismatch even when both numbers are individually estimated correctly.
Official syllabus reference: [2]
41. Schedule rating and self-insurance funding allocation
Schedule rating modifies a risk's rate for defined characteristics not adequately reflected in the base classification or experience measure. Credits and debits should connect to expected cost and avoid duplicating effects already recognized elsewhere. In self-insurance funding, a related allocation problem assigns expected losses, expenses, and selected risk charges among participating units. Exposure-based expected costs and credible individual experience can serve different roles in that allocation. Keep prospective incentives distinguishable from reimbursement of realized losses. A funding formula should reconcile to the total required contribution while making any deliberate cross-subsidies explicit.
Worked example: A self-insured group allocates expected claim cost using payroll and job hazard, then adds each unit's share of administration. A documented safety feature may justify a separate adjustment only if its effect is not already embedded in the hazard class or experience modification.
Common mistake: Applying a schedule credit for a characteristic already fully reflected in another factor can double-count its expected loss benefit.
Official syllabus reference: [2]
42. Retrospective premium assets and nonlinear adjustments
A retrospective premium asset represents expected future additional premium under an experience-sensitive contract, measured relative to premium already recognized or another explicitly defined baseline. Final premium may depend on ultimate ratable losses, basic premium, conversion factors, taxes, and minimum or maximum premiums. Forecast the premium adjustment using the joint contract and loss-development structure. When caps or floors bind, expected final premium is generally not the formula evaluated at expected losses. Collection timing, collectibility, and correlation between loss development and premium adjustments can matter. Distinguish additional expected premium from a simple percentage of outstanding losses.
Worked example: Two equally likely final premiums are $200,000 and $600,000 after contract limits. Expected final premium is $400,000. If $350,000 is already accrued, the defined asset is $50,000, before any specified discount or collectibility adjustment.
Common mistake: Applying a capped retrospective formula only to average ultimate losses can misstate expected premium because the contract responds separately in each outcome.
Official syllabus reference: [2]
Reinsurance
43. Quota share and surplus share proportions
Quota-share reinsurance cedes a fixed percentage of covered premiums and losses, subject to contract terms. Surplus-share reinsurance determines the ceded proportion risk by risk using the insurer's retained line and the amount insured, within treaty capacity. Both are proportional forms, but surplus share does not imply one constant portfolio-wide percentage. Commissions, exclusions, limits, and participation rules affect the resulting economics. Distinguish the proportion of an individual risk ceded from an excess-of-loss attachment applied to each loss. A proportional share ordinarily applies from the first covered dollar rather than only above a loss threshold.
Worked example: With a $200,000 retained line and sufficient surplus capacity, a $500,000 insured risk has 40% retained and 60% ceded. A covered $100,000 loss then produces a $60,000 cession. A separate $150,000 risk may be retained entirely under the same structure.
Common mistake: A surplus-share retained line determines a proportional risk share; it is not necessarily a deductible subtracted from every claim.
Official syllabus reference: [2]
44. Per-risk, per-occurrence, and aggregate excess triggers
Excess-of-loss contracts differ in the unit to which attachment and limit apply. Per-risk protection responds to covered losses on an individual risk, per-occurrence protection aggregates losses arising from a defined event, and aggregate protection responds to a specified period's cumulative covered losses. Event definitions, hours clauses, expense treatment, and inuring reinsurance can materially change recoveries. Apply the contract's aggregation rules before calculating layer payments. A single catastrophe may involve many risks and claims, so using claim-level attachments for an occurrence treaty can severely understate recoveries, while inappropriate event grouping can overstate them.
Worked example: One storm causes three covered losses of $400,000 each. A $1 million layer above $1 million per occurrence pays $200,000 on the $1.2 million event total. Applying the same attachment separately to each claim would incorrectly produce no recovery.
Common mistake: Identical attachment and limit numbers do not make per-claim and per-occurrence treaties economically equivalent when one event affects several claims.
Official syllabus reference: [2]
45. Treaty versus facultative placement and protection purpose
Treaty reinsurance applies to a defined portfolio under agreed automatic terms, while facultative reinsurance involves individual risk placement and acceptance. The choice interacts with objectives such as capacity, catastrophe protection, earnings stabilization, and access to expertise. A broad proportional treaty may support portfolio growth, whereas a facultative placement can address an unusually large or specialized risk. Evaluate attachment, exclusions, counterparty exposure, cost, and retained volatility against the objective. Reinsurance changes the insurer's net risk profile but ordinarily does not remove its obligations to policyholders under the original insurance contracts.
Worked example: An insurer writes mostly ordinary commercial properties but receives one unusually large industrial submission. A facultative placement can address that specific capacity need. Its catastrophe treaty still serves a separate purpose by protecting accumulation across many properties in one event.
Common mistake: Purchasing protection for one unusually large risk does not necessarily address the insurer's portfolio-wide catastrophe accumulation or annual earnings volatility.
Official syllabus reference: [2]
46. Reinsurance reserving with reporting lags and changing terms
Reinsurance reserving often involves delayed bordereaux, incomplete claim detail, changing cedent practices, and sparse large losses. Historical triangles may mix attachments, limits, shares, currencies, or underwriting years that are not directly comparable. Development methods require a stable meaning for each cell, so adjust data where feasible or segment the analysis. Reinsurance reporting delay is additional to the original insured claim's delay. Commutations, late advice, and changes in case-reserving practices can also distort apparent development. Supplement mechanical projections with exposure information, contract review, and claim-specific analysis while retaining a clear reconciliation to reported balances.
Worked example: A treaty attachment doubles between underwriting years. Older years contain many smaller ceded claims, while newer years report only severe losses and emerge later. Applying one pooled development factor would mix different layers and reporting patterns, potentially producing a misleading reserve.
Common mistake: A long reinsurance development pattern can reflect cedent reporting delay as well as the underlying time required to settle claims.
Official syllabus reference: [2]
47. Experience rating and burning-cost adjustments
Burning-cost pricing divides adjusted historical ceded losses by a suitable historical exposure base, often subject premium. Losses should reflect prospective treaty terms, development, and claim-cost level; premium should reflect comparable rate level and exposure. The historical ratio is a loss rate rather than a complete price, so expenses, commissions, risk charges, and profit require separate treatment. Sparse losses can make the estimate volatile, and changing underwriting mix can undermine representativeness. Compare experience-based indications with exposure-based methods rather than assuming a loss-free history proves negligible expected layer cost.
Worked example: Adjusted ceded losses are $600,000 on $12 million of rate-level-adjusted subject premium, giving a 5% loss rate. If expenses are 25% of treaty premium and no other load applies, the indicated gross rate is 5%/0.75 = 6.67%.
Common mistake: A historical loss ratio based on unadjusted premium can confuse primary rate changes with changes in underlying reinsurance risk.
Official syllabus reference: [2]
48. Exposure rating through severity curves
Exposure rating estimates expected reinsurance loss from the underlying insured portfolio rather than relying solely on the treaty's observed claims. For excess layers, severity or exposure curves allocate ground-up expected losses to attachments and limits, with adjustments for insured values, policy limits, deductibles, and risk characteristics. Results depend strongly on curve suitability and consistent exposure definitions. This approach is useful when historical ceded experience is sparse or terms have changed. It is not assumption-free: underlying expected loss ratios, severity tails, and accumulation patterns can all materially affect the indicated layer cost.
Worked example: A portfolio has expected ground-up losses of $10 million. A suitable exposure analysis allocates 8% to the proposed layer, producing expected layer losses of $800,000 before treaty expenses or risk loads. A different property-value mix may require a different allocation curve.
Common mistake: An exposure curve calibrated to one occupancy or insured-value mix should not be treated as universally representative of all portfolios.
Official syllabus reference: [2]
49. Compound aggregate distributions
An aggregate model combines claim count N with individual ceded severities Y to form S = sum of Y over N claims. Under independent, identically distributed severities independent of count, E[S] = E[N]E[Y] and Var(S) = E[N]Var(Y) + Var(N)E[Y]². For Poisson counts this becomes lambda E[Y²]. Simulation, recursion, or numerical convolution can estimate the full distribution when pricing depends on aggregate deductibles, limits, or tail risk. Preserve dependence and contract order when relevant. Matching the mean alone does not validate a model intended to price high-layer or aggregate protection.
Worked example: Expected count is two, expected severity is $100,000, and severity variance is $4 billion in squared dollars. With Poisson count, aggregate variance is 2 × ($4 billion + $10 billion) = $28 billion, showing the additional variability created by random count.
Common mistake: Multiplying severity variance by expected count alone omits count uncertainty unless the number of claims is fixed rather than random.
Official syllabus reference: [2]
50. Property catastrophe treaty pricing
Property catastrophe pricing translates modeled event losses into contract recoveries and a sustainable price. Apply occurrence definitions, attachments, limits, inuring covers, annual aggregate provisions, and reinstatements to an event-year representation where available. Expected ceded loss is only one component; expenses, capital cost, uncertainty, and market considerations may also affect pricing. Exposure quality and geographic accumulation are critical because rare events dominate results. Compare model versions and alternative assumptions on a consistent financial basis. A single return-period loss is useful for capacity analysis but cannot replace integration over the loss distribution when estimating expected treaty cost.
Worked example: A catastrophe model gives event losses before reinsurance. The analyst applies the occurrence layer to each event, then applies the annual limit and reinstatement rules in sequence. Averaging those annual recoveries estimates expected ceded loss more faithfully than applying the layer to average annual gross loss.
Common mistake: Applying an occurrence layer once to average annual gross loss does not reproduce expected recoveries from individual catastrophe events.
Official syllabus reference: [2]
51. Reinstatement premiums and restored limit
A reinstatement restores some or all exhausted treaty limit for subsequent covered losses, subject to the number and terms permitted. The additional premium may be proportional to the amount reinstated, time remaining, both, or another contract rule. It is a cash flow to the reinsurer and affects both cedent cost and economic risk-transfer analysis. In pricing, expected reinstatement premium can depend on the original premium, requiring an equation rather than a simple fixed addition. Distinguish restored available limit from an increase in the amount recoverable for the event that already exhausted it.
Worked example: A $10 million limit has a full original premium of $2 million. A contract charges 100% additional premium pro rata as to amount only. Restoring $4 million of exhausted limit produces $800,000 reinstatement premium, assuming the available reinstatement capacity has not been exhausted.
Common mistake: A reinstatement generally replenishes protection for subsequent covered losses; it does not automatically increase the exhausted event's own occurrence limit.
Official syllabus reference: [2]
52. Present-value cash-flow risk-transfer testing
Economic risk-transfer testing examines the reinsurer's possible financial outcomes using the complete contract cash flows. Include premiums, claims, commissions, refunds, reinstatement premiums, expenses when relevant, and settlement timing on a consistent present-value basis. Experience-sensitive features can reduce downside or return favorable experience to the cedent, changing the distribution materially. Scenario probabilities and loss timing require support. A significant accounting conclusion also depends on the applicable framework and contract substance. Avoid replacing that assessment with an unsupported universal numerical rule, and distinguish positive expected profit from the absence of meaningful adverse outcomes.
Worked example: A treaty receives $100 million and may pay $130 million of claims several years later. A retrospective premium adjustment recovers part of the adverse amount from the cedent. The test must include both the adjustment and payment dates before measuring the reinsurer's downside.
Common mistake: Testing only ceded claims against initial premium can overstate transferred risk when adverse experience triggers substantial additional cedent payments.
Official syllabus reference: [2]
53. Expected reinsurer deficit
Expected reinsurer deficit measures downside by averaging the negative part of modeled reinsurer profit, often expressed relative to a specified premium measure. If profit is P, the numerator is E[max(−P,0)]. It combines the probability of loss with the average loss severity conditional on loss. It differs from expected profit because favorable scenarios do not offset deficits in the numerator. Define the denominator and present-value convention explicitly. The statistic supports economic analysis but does not, by itself, establish compliance with every accounting framework or prescribe a universal threshold for sufficient risk transfer.
Worked example: A reinsurer loses $20 million with probability 10% and earns a profit otherwise. Expected deficit is $2 million. Relative to $50 million of premium, the measure is 4%, regardless of how large the positive profits are in the remaining scenarios.
Common mistake: Expected profit and expected deficit answer different questions; profitable scenarios do not cancel adverse scenarios when calculating the negative-part expectation.
Official syllabus reference: [2]
54. Sliding commissions and experience-sensitive economics
Sliding-scale commissions vary with treaty experience, commonly returning more premium to the cedent when losses are favorable and less when losses are adverse. Profit commissions, loss corridors, retrospective premiums, and deficit carryforwards can create additional nonlinear cash flows. These provisions change both price and the distribution of risk retained by each party. Apply each formula within scenarios before averaging, respecting caps, floors, timing, and interactions. A contract's label does not establish its economic substance. Analyze how downside is shared across plausible outcomes and whether connected agreements alter the apparent transfer.
Worked example: A commission falls from 30% to 10% when the treaty loss ratio rises. The cedent therefore absorbs part of adverse experience through reduced commission, even before any explicit additional premium. A scenario model includes this offset when calculating reinsurer losses and cedent net cost.
Common mistake: Treating a sliding commission as one fixed average percentage can conceal how it reduces or reshapes risk in adverse scenarios.
Official syllabus reference: [2]
Catastrophe Models and Risk Loads
55. Hazard, vulnerability, exposure, and financial modules
A catastrophe model links event occurrence and physical intensity to insured loss. The hazard component describes events and site-level intensity; exposure data locate and characterize insured assets; vulnerability relationships translate intensity into damage; financial calculations apply insurance terms. Errors can propagate between components, so accurate policy terms cannot compensate for poor geocoding or unsuitable vulnerability assumptions. Models may also include uncertainty around damage and event parameters. Understanding the module chain helps identify why two models disagree and which input improvements are likely to matter for underwriting, accumulation management, or reinsurance analysis.
Worked example: Two identical buildings have the same replacement value but different modeled flood depths because one is geocoded to the wrong elevation. The vulnerability and financial calculations may operate correctly while the location error still produces a substantially incorrect insured loss estimate.
Common mistake: A detailed financial module does not eliminate uncertainty or bias introduced by inaccurate locations, building characteristics, or hazard assumptions.
Official syllabus reference: [2]
56. Average annual loss from event frequencies
Average annual loss is the expected total catastrophe loss per year on a specified gross, insured, or net basis. In an event-loss representation with compatible event frequencies, it can be calculated as the sum of each event's annual occurrence rate multiplied by its expected loss. In an annual simulation, average the total loss across simulated years using their probabilities. Do not average event losses without considering their frequencies. AAL supports expected-cost pricing and budgeting but says little by itself about volatility, annual exhaustion, or extreme-loss capital needs. Preserve the financial basis when comparing model outputs.
Worked example: One event has annual rate 0.02 and expected loss $100 million; another has rate 0.01 and expected loss $300 million. Their contributions to AAL are $2 million and $3 million, totaling $5 million under the stated event representation.
Common mistake: The simple average of event loss amounts is not average annual loss unless event probabilities and the annual aggregation structure justify it.
Official syllabus reference: [2]
57. Occurrence and aggregate exceedance curves
An occurrence exceedance curve describes the annual probability that the largest event loss exceeds a threshold. An aggregate exceedance curve describes the probability that the year's total event loss exceeds it. Aggregate exceedance can be greater because several smaller events may combine to cross the threshold. For an annual exceedance probability p, a conventional return period is 1/p, but it is not a timetable or a guarantee of one event per interval. State whether thresholds are exceeded strictly and how interpolation is handled. Contract basis and annual event dependence influence both curves.
Worked example: A year contains two event losses of $80 million each. At a $100 million threshold, that year contributes to aggregate exceedance because total loss is $160 million, but not occurrence exceedance because neither individual event exceeds $100 million.
Common mistake: A 100-year return-period loss does not imply that a comparable loss cannot occur in consecutive years under the modeled annual process.
Official syllabus reference: [2]
58. Secondary uncertainty and model sensitivity
Catastrophe output reflects uncertainty in event occurrence and in losses conditional on an event. Secondary uncertainty commonly refers to variability in damage or loss given the modeled event conditions, while terminology can vary by model. Exposure errors, vulnerability assumptions, and model-form limitations add further uncertainty. Including conditional loss variability can materially affect tail estimates and nonlinear insurance recoveries even when mean event damage is unchanged. Compare alternative settings using the same exposure and financial basis. Model blending or sensitivity analysis should have an explicit rationale rather than assuming an arithmetic average removes bias.
Worked example: An event has mean insured loss of $50 million, but plausible conditional outcomes range widely. A treaty attaching at $60 million may have positive expected recovery once that variability is included, even though applying the attachment only to the $50 million mean gives zero.
Common mistake: Applying a nonlinear insurance layer to mean event loss can understate expected recovery when conditional event-loss variability crosses the attachment.
Official syllabus reference: [2]
59. Covariance allocation of a variance-based risk load
A variance-based portfolio risk load depends on account variances and their covariances. If the total load is k Var(S), one full-allocation rule assigns account i the amount k Cov(X_i,S), since these covariances sum to Var(S). For a standard-deviation load k SD(S), an Euler allocation is k Cov(X_i,S)/SD(S), provided the relevant conditions hold. The chosen risk functional determines the allocation formula. Accounts with similar standalone variability can receive different allocations because their dependence with the existing portfolio differs. These calculations require joint loss information rather than standalone summaries alone.
Worked example: Account variances are 100 and 225, with covariance 30. Portfolio variance is 385. For k = 1, covariance allocations are 100 + 30 = 130 and 225 + 30 = 255, which sum exactly to the specified variance-based load of 385.
Common mistake: Allocating solely in proportion to standalone variances ignores covariance and need not represent each account's contribution to portfolio catastrophe risk.
Official syllabus reference: [2]
60. Shapley values and portfolio diversification
A Shapley allocation treats the portfolio risk-load function as a cooperative game. For each account, calculate its incremental contribution when added to every possible predecessor set, weighted equivalently by averaging across all entry orders. The allocations sum to the total portfolio value when the empty portfolio has zero value. Averaging removes dependence on one arbitrary entry order and shares diversification according to marginal contributions. For nonlinear risk functions, coalition values are needed; standalone loads and the grand total alone are insufficient for larger portfolios. The result is conditional on the selected risk-load function and modeled dependence.
Worked example: Account A has standalone load $40,000, B has $30,000, and their combined load is $50,000. A's contributions are $40,000 when first and $20,000 when second, averaging $30,000. B receives the remaining $20,000 by the same calculation.
Common mistake: For three or more accounts, standalone values and the total portfolio load generally do not provide all coalition values needed for Shapley allocation.
Official syllabus reference: [2]
Official syllabus sources
- GI 301 Further Topics in General Insurance
- Syllabus with Learning Objectives/Outcomes and Readings – Mar 2027/Jul 2027