How should I choose between claim development and an expected-loss method?

Start with the maturity and reliability of the experience. Development methods project observed paid or incurred claims using selected factors, so unstable early emergence can distort the result. An expected-loss method relies more on an independent loss expectation. Bornhuetter-Ferguson combines that expectation with claims reported or paid to date; Benktander updates the blend as experience matures.

What makes a classification model useful for ratemaking?

A useful model predicts new experience and produces rating differences that can be explained and implemented. Check exposure, distribution and link choices, interactions, residual patterns, stability across periods, and performance on held-out data. A statistically significant factor may have little rate impact, while a large indicated change may be too uncertain or confounded to adopt without further analysis.

Why compare VaR with tail value at risk in capital decisions?

VaR identifies a loss threshold at a chosen probability level. Tail value at risk averages losses beyond that threshold, so it distinguishes portfolios with similar thresholds but different extreme outcomes. Assess either measure alongside dependence, liquidity, the time horizon and the cost of capital before using it to allocate capital or judge a mitigation strategy.

FCAS Study Guide: Claims, Ratemaking and Risk

Start with concept 1 Choose a topic

FCAS study builds on the ACAS foundation and brings together three distinct technical problems. In Exam 7, you turn claim triangles and exposure information into liability estimates, test development assumptions, and explain uncertainty around a selected reserve, including the effect of reinsurance. Exam 8 asks how experience becomes a defensible rate: model frequency and severity, validate classifications, apply credibility, and price deductibles, limits and individual risk plans. Exam 9 asks whether those underwriting risks fit an insurer’s capital and risk appetite. You compare tail measures, value cash flows, allocate capital, and test dependence and mitigation. Throughout the guide, the useful habit is to state assumptions, calculate the result, and explain what evidence would change the decision.

Exam 7: Advanced Estimation of Claims Liabilities

1. Build a reliable development triangle

A development triangle groups claims by origin period in rows and maturity in columns. Incremental payments belong to one development interval; cumulative payments add every interval through that age. Build each cumulative cell by adding the current increment to the preceding cumulative cell, then reconcile row totals and the latest diagonal with the claim system. Check that valuation dates, coverage, currency, recoveries, and treatment of reopened claims are consistent. A discrepancy can change both the current paid amount and the age-to-age factor derived from that row. For each factor, compare the sum of cumulative claims at the later age with the sum at the earlier age for rows observed at both ages. Investigate negative increments, missing cells, duplicate claims, and abrupt changes before deciding whether to correct, exclude, or retain an observation. The Exam 7 content outline includes data diagnostics and unpaid claim estimation.

Worked example: Payments for one origin year are 80, 30, and 15 thousand dollars in successive intervals. Cumulative payments must be 80, 110, and 125 thousand dollars. A supplied second-age cumulative value of 105 is 5 thousand dollars below 80 + 30. Using it would make the second-to-third factor 125 ÷ 105 = 1.190 instead of 125 ÷ 110 = 1.136. Trace the missing 5 before selecting a factor.

Watch the distinction: A triangle cell is not reliable merely because later cumulative values increase. An omitted payment can preserve that pattern while inflating a development factor.

Reference: [2]

2. Detect changes in case reserving

Incurred claims equal paid claims plus case reserves at a valuation date. A payment normally reduces the case reserve on that claim, so paid growth alone need not increase incurred claims. When incurred development turns negative while paid development is positive, separate settlement activity from changes in estimated outstanding amounts. Compare closed and open claim counts, average case reserves on open claims, payments on closed claims, and large claim movements. A faster settlement pattern may shift amounts from case to paid without changing expected ultimate claims. A deliberate case-reserve strengthening or release changes incurred amounts even without a corresponding change in ultimate cost. Examine whether the change affects one origin year, a claim type, or several calendar periods. That diagnosis determines whether a historical incurred factor remains appropriate and whether paid data offer a more stable cross-check.

Worked example: At the first valuation, paid claims are 60 and case reserves are 50, so incurred claims are 110 thousand dollars. At the next, paid claims are 75 and case reserves are 25, giving incurred claims of 100 thousand dollars. Payments increased by 15, while case reserves fell by 25; incurred claims declined by 10. Review closures and case-reserve changes before extrapolating that decline.

Watch the distinction: Positive paid development does not prove ultimate claims increased. Payment can replace an existing case reserve, and a separate case-reserve release can reduce incurred claims.

Reference: [2]

3. Make unlike claim cohorts comparable

Claim cohorts can differ because the amount insured, number of exposures, coverage terms, deductible, or business mix changed. Raw claim counts and losses then combine changes in risk with changes in portfolio size. Put observations on a comparable basis before pooling development: calculate frequency per exposure, severity per claim, or loss per consistently defined exposure, as the question requires. For reserve work, also ask whether older cohorts represent the claims still open in newer cohorts. A coverage expansion may add large claims with slower reporting, so a simple exposure adjustment cannot restore comparable development by itself. Segment material differences or adjust a model for them, then test whether enough credible experience remains. Document the exposure definition and whether claims are reported, closed, or ultimate. A difference in those definitions can imitate a change in frequency.

Worked example: Cohort A has 120 reported claims from 10,000 exposures, or 120 ÷ 10,000 × 100 = 1.2 claims per 100 exposures. Cohort B has 150 claims from 20,000 exposures, or 0.75 per 100. Counts rose by 30, yet observed frequency fell by 0.45 per 100. Compare coverage and reporting practices before interpreting the lower rate as a genuine risk improvement.

Watch the distinction: More claims do not necessarily mean a riskier cohort. When exposure grows faster than claim count, frequency falls even as the raw count rises.

Reference: [2]

4. Recognize calendar-period distortions

Every triangle cell has an origin period, development age, and calendar period. Calendar period equals origin period plus development interval under a consistent indexing convention. A change in payment processing, inflation, legal environment, or settlement strategy can affect cells on one diagonal across several origin years. That pattern differs from an origin-year effect, which follows one row, and an age effect, which follows one column. Inspect incremental amounts and diagnostic ratios by all three dimensions. If a calendar shock accelerated payments that otherwise would have occurred later, its current spike may be followed by lower future payments. Applying an age-only pattern that treats the spike as repeatable can overstate ultimate claims. Before adjusting, test competing explanations such as a large isolated claim, a portfolio change, or changes in data capture. The adjustment should match the mechanism identified.

Worked example: Three origin years each have an extra 10 thousand dollars of incremental payments during calendar 2025, for 30 thousand dollars in total. Earlier and later calendar diagonals lack that pattern. If a one-time settlement initiative caused the payments, a factor calibrated to the elevated age cells could project too much future payment. Compare claim closures and later development before removing or redistributing the effect.

Watch the distinction: A high payment cell is not automatically unusual development at its age. When several ages spike in one calendar year, investigate a shared calendar cause.

Reference: [2]

5. Project paid claims to ultimate

The paid development method estimates ultimate claims by multiplying paid claims at the current maturity by a selected cumulative factor to ultimate. Unpaid claims equal estimated ultimate less paid to date, with both amounts on the same coverage and recovery basis. The cumulative factor is the product of selected age-to-age factors and any supported tail factor. Selection requires more than averaging historical ratios: inspect volume, outliers, settlement speed, changes in claim mix, and whether older origin years are representative. A factor near one implies most expected payments have emerged; it does not prove there are no late claims. Reconcile the resulting unpaid amount with case reserves, expected losses, and recent payments. If settlement accelerated temporarily, historical paid factors may overstate development still to come. Apply the factor at the specified age, not to an unrelated paid total.

Worked example: At age 24 months, an origin year has paid claims of 120 thousand dollars. Its selected factor from age 24 to ultimate is 1.25. Estimated ultimate claims are 120 × 1.25 = 150 thousand dollars, leaving unpaid claims of 150 − 120 = 30 thousand dollars. If recent settlements have accelerated, test whether 1.25 still represents the remaining payment pattern.

Watch the distinction: The factor's excess over one is not itself the unpaid amount. Multiply paid claims by the full factor, then subtract paid claims to obtain unpaid claims.

Reference: [2]

6. Project incurred claims to ultimate

Incurred development starts with paid claims plus case reserves. Multiply current incurred claims by a selected incurred factor to estimate ultimate claims, then subtract paid claims to obtain total unpaid claims. Subtracting incurred claims instead gives development beyond current case reserves, a different quantity. Because case reserves reflect claim handlers' current estimates, incurred data may react earlier than paid data to new information. They can also shift abruptly when case-setting rules, large-loss review, or claim mix changes. Compare paid and incurred indications, case reserve adequacy, open counts, and reported claim severity before choosing a result. For a strengthening of current case reserves, the old historical factor may double-count some expected emergence if applied unchanged. The Exam 7 study guide provides a focused route through these reserving methods.

Worked example: Current incurred claims at age 24 months are 180 thousand dollars, including 130 thousand dollars paid and 50 thousand dollars in case reserves. A selected incurred-to-ultimate factor of 1.10 gives ultimate claims of 180 × 1.10 = 198 thousand dollars. Total unpaid is 198 − 130 = 68 thousand dollars: 50 in case reserves plus 18 of estimated development beyond them.

Watch the distinction: Ultimate less incurred is not the total unpaid reserve. It omits the case reserves already included in incurred claims; subtract paid claims for total unpaid.

Reference: [2]

7. Combine an expected loss with emergence

The paid Bornhuetter-Ferguson method combines observed payments with an independent expected ultimate loss. If the selected paid emergence proportion is p, the expected unpaid proportion is 1 − p. Estimated unpaid claims equal expected ultimate claims times 1 − p; estimated ultimate equals that unpaid amount plus paid claims. This method limits the influence of immature actual payments because only the emerged portion of ultimate responds directly to experience. Its prior expectation should come from a relevant exposure and loss-cost basis, adjusted for coverage, trend, and mix. The emergence pattern must match that basis and the current settlement process. As maturity rises, the expected unpaid share falls and observed payments carry more of the ultimate estimate. Compare the implied ultimate with chain ladder and investigate material differences rather than treating the prior as automatically correct.

Worked example: Suppose expected ultimate claims are 200 thousand dollars and the selected paid emergence proportion is 40%. Paid claims to date are 70 thousand dollars. The expected unpaid share is 60%, so estimated unpaid claims are 200 × 0.60 = 120 thousand dollars. Estimated ultimate is 70 + 120 = 190 thousand dollars. The difference from the 200 prior reflects payments below the prior's implied emerged amount of 80.

Watch the distinction: Bornhuetter-Ferguson does not multiply current paid claims by the unpaid percentage. That percentage applies to the independent expected ultimate loss, while actual paid claims enter separately.

Reference: [2]

8. Update a reserve with Benktander credibility

Benktander reserving applies emergence credibility repeatedly. Start with a Bornhuetter-Ferguson ultimate based on an expected-loss prior. Use that first result as the revised prior in the same emergence formula to obtain the next estimate. With paid emergence proportion p and initial expected ultimate E, the first ultimate is P + (1 − p)E. The second is P + (1 − p)[P + (1 − p)E]. When the same p is used, this second result moves toward the chain-ladder ultimate P ÷ p, while retaining some influence from E. The direction depends on whether actual payments exceed or fall below the prior's implied emerged payments. Check that p describes the same paid basis and cohort throughout. Repeated updating cannot cure a biased emergence pattern or an expected-loss prior drawn from unlike risks.

Worked example: Paid claims are 70 thousand dollars, expected ultimate claims are 200 thousand dollars, and p = 0.40. The first Bornhuetter-Ferguson ultimate is 70 + 0.60 × 200 = 190 thousand dollars. The next Benktander ultimate is 70 + 0.60 × 190 = 184 thousand dollars. Chain ladder gives 70 ÷ 0.40 = 175 thousand dollars, so the updated estimate moves from 190 toward 175.

Watch the distinction: The second update does not use the original 200 prior again. It uses the first ultimate estimate as its revised expectation, changing the result.

Reference: [2]

9. Test development-factor assumptions

A selected age-to-age factor assumes the observed relationship can forecast later cohorts at that maturity. Examine individual link ratios against origin period, prior cumulative size, claim count, and earlier development. A systematic origin-period trend may signal changing settlement speed or case adequacy. A pattern tied to claim volume may reveal instability in small cohorts or a misspecified variance relationship. Correlation with prior factors can indicate that apparently separate development stages are linked. Calculate a volume-weighted factor using sums of later and earlier cumulative amounts for matched rows, but do not mistake that pooled result for evidence of stability. Compare residuals and alternative selections, then assess how each choice changes ultimate claims. If a genuine operational change occurred, a recent-data selection or an explicit adjustment may be preferable to a long historical average.

Worked example: Suppose three equally sized origin-year bases each show age-to-age factors of 1.8, 1.6, and 1.4 in chronological order. Their equal-base pooled factor is (1.8 + 1.6 + 1.4) ÷ 3 = 1.6. The steady decline suggests a newer cohort could develop below 1.6. Investigate settlement timing and mix before choosing a forecast; unequal bases would require volume weighting instead.

Watch the distinction: A pooled factor is a calculation, not a stability test. A consistent downward trend can make its forecast unsuitable even when the average is computed correctly.

Reference: [2]

10. Estimate claims by coverage layer

For a layer with attachment a and limit l, claim-level layer loss is min(max(X − a, 0), l), where X is ground-up claim size on the contract's defined basis. This function is nonlinear: a claim below attachment contributes nothing, while a small size change can create a layer payment. Trend and development can therefore change the number of claims reaching the layer, not just the amounts already in it. Estimate the ground-up claim-size distribution, apply appropriate trend and development, then apply the layer terms to projected claims or to the distribution. Check whether expenses, deductibles, aggregate limits, and claim definitions affect attachment. A factor estimated for ground-up dollars generally does not transfer unchanged to layer dollars. Compare modeled layer results with observed threshold crossings and large-claim diagnostics, recognizing that high layers may have sparse experience.

Worked example: A ground-up claim of 90 thousand dollars increases by a hypothetical 20%, reaching 90 × 1.20 = 108 thousand dollars. For a 100 thousand excess of 100 thousand layer, current layer loss is zero. After trend it is min(max(108 − 100, 0), 100) = 8 thousand dollars. Trend created a layer claim rather than merely increasing an existing one.

Watch the distinction: Applying a ground-up percentage directly to observed layer losses misses claims that cross attachment after development or trend. Transform claim sizes before calculating layer payments.

Reference: [2]

11. Reserve a high-deductible obligation

Under a per-claim deductible d, the insured's retained part of a ground-up claim X is min(X, d), and the insurer's part is max(X − d, 0), subject to any applicable limit and contract terms. Estimate both portions at claim level because aggregate ground-up losses do not reveal how many claims cross the deductible. For unpaid claims, project each claim's ultimate size or an appropriate size distribution, calculate each party's expected share, then subtract amounts already paid on the same basis. Consider whether the insurer pays the claim first and seeks reimbursement from the insured; that arrangement creates collection exposure alongside claim liability. Claim counts, reported sizes, development patterns, deductible changes, and insured reimbursement history are useful diagnostics. A high deductible can make the insurer's share especially sensitive to large claims and threshold crossings.

Worked example: Three ground-up claims are 30, 80, and 140 thousand dollars, with a 50 thousand dollar per-claim deductible and no other limits. Insured portions are 30, 50, and 50, totaling 130 thousand dollars. Insurer portions are 0, 30, and 90, totaling 120 thousand dollars. The 250 thousand dollar ground-up total alone would not determine this split.

Watch the distinction: Multiplying aggregate losses by an assumed deductible percentage can misstate both obligations. A per-claim deductible must be applied separately to each claim or its size distribution.

Reference: [2]

12. Forecast retrospective premium

A retrospective premium responds to losses under a contractual formula, often with a basic charge, a loss conversion factor, and minimum and maximum premiums. Project the relevant ultimate losses first, apply the formula and caps, then compare the resulting premium with amounts already billed or collected. This produces an estimated future premium adjustment, which is distinct from the unpaid claim liability. Its realizable value can depend on audit timing, contract wording, disputes, and collectibility. Losses near a cap deserve special attention because a further loss change may have little or no premium effect. If the formula uses a loss basis different from the reserve basis, reconcile claim expenses, limits, and development before calculating. Work through a few alternative loss outcomes to understand sensitivity; one central estimate can conceal a capped or asymmetric result. Additional FCAS practice questions can reinforce this distinction.

Worked example: A hypothetical plan sets premium equal to 100 + 0.8 times ultimate losses, with a 150 minimum and 250 maximum, all in thousands of dollars. Projected losses of 180 give 100 + 0.8 × 180 = 244, within the limits. If 200 has been billed, projected additional premium is 44 thousand dollars, subject to collection. This 44 is a premium adjustment, not an unpaid claim reserve.

Watch the distinction: Do not add a projected retrospective premium adjustment to unpaid claims as though both were claim liabilities. They arise from different contract cash flows.

Reference: [2]

13. Estimate stochastic development parameters

A volume-weighted development factor divides total cumulative claims at the later age by total cumulative claims at the earlier age, using only cohorts observed at both ages. In a stochastic reserving model, that fitted factor is a parameter estimate, not a certain constant. Compare each cohort's later amount with its fitted value from the earlier amount; the differences are residuals. Their dispersion helps assess how variable observed development is relative to the model's assumptions. Residuals also need diagnostic checks: trends, outliers, changing variance with volume, and dependence across cells can invalidate a simple uncertainty estimate. Parameter risk comes from estimating the factor with limited data; process risk concerns future claims even if the factor were known. A model may combine both, but its uncertainty calculation depends on the assumed variance structure and independence conditions.

Worked example: Two origin years move from 100 to 150 and from 200 to 260 thousand dollars. The volume-weighted factor is (150 + 260) ÷ (100 + 200) = 410 ÷ 300 = 1.3667. It predicts 136.67 and 273.33 at the later age, giving residuals of +13.33 and −13.33 thousand dollars. Two observations show variation but cannot establish a reliable dispersion estimate.

Watch the distinction: The pooled factor does not erase differences between cohorts. Their residuals carry information about fit and uncertainty, although very few observations cannot support a stable variance estimate.

Reference: [2]

14. Translate distribution parameters into tails

A reserve distribution describes possible unpaid claim outcomes rather than a single estimate. For a lognormal variable R, log R is normal with mean μ and standard deviation σ. Its median is exp(μ), its arithmetic mean is exp(μ + σ²/2), and its pth percentile is exp(μ + z_pσ), where z_p is the corresponding standard normal quantile. The mean exceeds the median because high outcomes in the right tail pull it upward. Use the distribution on a clearly stated monetary scale and check whether fitted parameters describe reserve uncertainty, ultimate losses, or individual claims; those are different targets. A mathematically calculated high percentile is only as credible as the fitted tail and dependency assumptions. Compare implied moments and percentiles with observed large claims, selected reserve ranges, and plausible stress scenarios before relying on a margin.

Worked example: Assume unpaid claims R, measured in thousands of dollars, are lognormal with μ = 4 and σ = 0.2. The mean is exp(4 + 0.2²/2) = exp(4.02) ≈ 55.7 thousand dollars. Using z₀.₉₅ ≈ 1.645, the 95th percentile is exp(4 + 1.645 × 0.2) = exp(4.329) ≈ 75.9 thousand dollars. The higher percentile reflects right-tail outcomes.

Watch the distinction: The log-scale mean μ is not the reserve's arithmetic mean. Exponentiating μ gives the median; the mean also includes the σ²/2 adjustment.

Reference: [2]

15. Interpret reserve prediction error

Prediction error measures how far future unpaid claims may fall from today's estimate. Process variation remains even if model parameters are known: claim occurrence, size, and payment timing can still vary. Parameter estimation risk comes from learning factors and other parameters from finite data. Under a stated model with zero covariance between these components, their variances add to mean squared prediction error, and its square root is root mean squared prediction error. If components are dependent, include their covariance rather than assuming additivity. Model uncertainty is broader: choosing a different development structure, tail form, or data adjustment may change the forecast and can lie outside a reported model-based error. State which sources the figure covers and whether it is conditional on selected assumptions. Comparing errors across models requires a consistent reserve basis, horizon, and treatment of parameter uncertainty.

Worked example: Suppose a hypothetical reserve model assigns process variance 64 and parameter variance 36, both in squared thousands of dollars, with zero covariance. Mean squared prediction error is 64 + 36 = 100 squared thousands of dollars. Its square root is 10 thousand dollars. The component standard deviations are 8 and 6 thousand dollars; adding them to get 14 would overstate the stated combined measure.

Watch the distinction: Adding component standard deviations is not variance aggregation. Square them, account for any covariance, add the resulting terms, and take a square root if reporting root mean squared prediction error.

Reference: [2]

16. Construct a bootstrap predictive view

A reserving bootstrap starts with a fitted model and residuals that describe how observed triangle cells differ from fitted values. After adjusting residuals as required by the model, resample them to create a plausible alternative historical triangle, refit development parameters, and project its future cells. To obtain a predictive distribution rather than only a distribution of fitted means, simulate future process variation as well. Repeat the sequence many times and sum future cells in each trial to obtain a reserve outcome. Check that simulated triangles preserve meaningful features of the data, including volume patterns and any modeled calendar effects. A bootstrap reproduces variation represented by the fitted residuals; it cannot discover an omitted structural change. Inspect central values, tail behavior, and sensitivity to influential cells, since extreme percentiles can be unstable even when the simulated mean looks reasonable.

Worked example: Three illustrative bootstrap trials produce reserve outcomes of 90, 105, and 120 thousand dollars after refitting and simulating future payments. Their average is (90 + 105 + 120) ÷ 3 = 105 thousand dollars. The observed range is 30 thousand dollars. These trials illustrate the calculation, but three outcomes provide almost no reliable information about a 95th percentile or rare tail event.

Watch the distinction: Resampling fitted reserve estimates alone omits future process variation. A predictive bootstrap must represent both parameter changes from refitting and possible future claim outcomes.

Reference: [2]

17. Model development with a GLM

An incremental-claims generalized linear model can express expected payment in a cell through origin-year, development-age, and calendar-period effects. With a log link, additive coefficients on the log scale become multiplicative factors on the claim scale, ensuring a positive fitted mean. Choose a response and variance family appropriate to the data, then inspect residuals for systematic patterns. Calendar period is determined by origin period and development age, so a model with unrestricted effects for all three dimensions has an identifiability problem, particularly in their linear trends. Apply an explicit constraint or structured trend before interpreting coefficients. Different constraints can yield the same fitted historical cells but different extrapolations, making the forecast assumption consequential. A fitted calendar factor may capture inflation, process change, or another shared effect; investigate its cause rather than assigning a name solely from its statistical position.

Worked example: A hypothetical log-link model has a baseline expected incremental payment of 20 thousand dollars, an origin multiplier of 1.2, and a development multiplier of 0.5. Their product gives 20 × 1.2 × 0.5 = 12 thousand dollars. A calendar multiplier of 1.1 raises it to 13.2 thousand dollars. The calendar effect needs an identifying constraint and a defensible forecast beyond observed periods.

Watch the distinction: Separate origin, age, and calendar coefficients are not automatically identifiable. Calendar timing follows from origin and age, so unrestricted fitted trends can be confounded.

Reference: [2]

18. Incorporate expert judgment probabilistically

A conditional reserve forecast assumes a particular set of parameters and judgments is correct. A predictive distribution also allows uncertainty about those choices and future claim outcomes. One practical approach assigns probabilities to plausible development regimes, fits or specifies a conditional reserve distribution for each, then combines them as a mixture. The mixture mean is the probability-weighted average of conditional means. By the law of total variance, overall variance equals the average conditional variance plus the variance of conditional means. This second term represents uncertainty about which regime applies. Weights should reflect available evidence and expert reasoning, and sensitivity to them should be shown. Avoid treating scenarios as independent claim events or adding their reserve means. Check that each scenario shares the same coverage, valuation date, and monetary basis before mixing results.

Worked example: Two plausible regimes each have 50% weight. Their reserve means are 80 and 120 thousand dollars, each with conditional variance 100 squared thousands of dollars. The mixture mean is 0.5 × 80 + 0.5 × 120 = 100 thousand dollars. Between-regime variance is 0.5 × (80 − 100)² + 0.5 × (120 − 100)² = 400. Total variance is 100 + 400 = 500 squared thousands of dollars.

Watch the distinction: Averaging the conditional variances alone understates predictive uncertainty when scenarios have different means. Include the variance between those scenario means.

Reference: [2]

19. Evaluate output and select a risk margin

Reserve output needs diagnostic review before a range or margin is selected. Compare methods on the same claim and expense basis, inspect whether their differences arise from data, factor selections, prior expectations, or model form, and test key assumptions. Examine fitted versus observed development, influential cells, and whether simulated tails are plausible relative to large claims and coverage limits. A working range can exclude an indication whose mechanism is demonstrably unrepresentative, but the reason should be specific and quantified. A risk margin defined by a percentile is the selected percentile of the unpaid-claim distribution less its selected mean, when both use the same basis. Its size changes with tail shape, dependence, and parameter uncertainty. A percentile is not validated merely by calculation; the intended risk measure and the model's tail behavior must fit the decision being made.

Worked example: Three methods indicate unpaid claims of 95, 105, and 125 thousand dollars. Investigation shows the 125 result projects a one-time calendar payment surge as recurring, supporting a working range of 95 to 105. If a separately validated distribution has selected mean 100 and 75th percentile 112 thousand dollars, the stated percentile margin is 112 − 100 = 12 thousand dollars. The range and margin answer different questions.

Watch the distinction: The highest point estimate is not automatically a prudent margin. First assess why methods differ, then calculate any margin from a defined, credible distribution.

Reference: [2]

20. Translate primary losses into reinsurance reserves

A ceded reserve is the projected future recovery from reinsurance under the applicable contract, not a fixed percentage of primary unpaid claims unless the cover is proportional and all relevant terms align. For per-risk excess cover, calculate each claim's ceded amount at current and projected ultimate ground-up size, then take the difference on a consistent paid or incurred basis. The layer payment is min(max(X − attachment, 0), limit), subject to contract provisions. A quota share instead cedes a stated share of covered losses, with possible adjustments for exclusions or participation terms. Reinsurance reserving needs claim-level sizes, coverage dates, attachment and limits, reinstatement or aggregate terms where relevant, and the status of prior recoveries. Separate gross, ceded, and net estimates; reconcile gross minus ceded to net. Test large claims near attachment because small projection changes can create disproportionate ceded changes.

Worked example: A claim is currently valued at 80 thousand dollars and projected to reach 150 thousand dollars. A 100 thousand excess of 100 thousand cover cedes zero now and min(150 − 100, 100) = 50 thousand dollars at ultimate, so projected ceded unpaid is 50. Under a separate 40% quota share, ceded unpaid would be 0.40 × (150 − 80) = 28 thousand dollars.

Watch the distinction: A quota-share percentage cannot be applied to an excess layer's unpaid amount. The contracts respond to different parts of claim size and produce different recoveries.

Reference: [2]

Exam 8: Advanced Ratemaking

21. Calculate a class loss cost

A class loss cost is projected claim cost divided by the earned exposure that produced it. Start by putting losses on a common basis: develop immature claims, trend costs to the prospective period, and adjust for changes in benefits, limits, or deductibles. Match those losses to the same class definitions and exposure period in the denominator. A relativity compares class loss costs only after these adjustments; a difference caused by coverage or time is not a class effect. The class pure premium covers expected claims per exposure. To obtain a charged rate, the pricing framework must also account for expenses, profit or risk provisions, and the way those provisions vary with premium or exposure. The Exam 8 content outline places class loss costs within broader evaluation of rating schemes.

Worked example: Suppose projected losses are $360,000 for 1,200 earned exposures in class A and $240,000 for 600 in class B, all on the same coverage and cost basis. Loss costs are $360,000 ÷ 1,200 = $300 and $240,000 ÷ 600 = $400 per exposure. B's indicated relativity to A is $400 ÷ $300 = 1.333. That ratio compares expected claims; it is not yet a complete rate.

Watch the distinction: Dividing developed losses by written exposures mixes time bases. Policies written near year-end may contribute little earned exposure, making the apparent loss cost misleading.

Reference: [3]

22. Specify a classification GLM

For claim frequency, define the response as a count over a measured period and include the logarithm of earned exposure as an offset in a log-link GLM. Its coefficient is fixed at one, so doubling exposure doubles the expected count while leaving the annual frequency unchanged. Rating variables such as class and territory enter as estimated coefficients; exponentiating one gives a multiplicative frequency effect relative to its reference level. Choose a count variance model after examining dispersion, since Poisson variance equal to the mean may be too restrictive. Also decide how claims, cancellations, and zero-exposure records are treated before fitting. The fitted response is a claim count, not a premium: severity and other pricing components are needed to convert it to expected cost.

Worked example: Assume a baseline frequency of 0.04 claims per exposure-year and a class coefficient of log(1.5). For a risk with two earned exposure-years, the fitted count is exp(log(2) + log(0.04) + log(1.5)) = 2 × 0.04 × 1.5 = 0.12 claims. The class raises annual frequency to 0.06; the offset then scales it over two years.

Watch the distinction: Estimating a free coefficient on log exposure can let the model treat observation length as a risk characteristic. An offset preserves the intended frequency interpretation.

Reference: [3]

23. Choose models for frequency and severity

A frequency model predicts claims per exposure; a severity model predicts cost per claim conditional on a claim occurring. Their product estimates loss cost per exposure when both use consistent claim definitions, coverage terms, and target risk mix. Counts often call for Poisson or overdispersed count models. Positive, right-skewed claim amounts call for a suitable positive-severity model, with explicit treatment of limits, deductibles, and very large losses. Diagnose the components separately: a portfolio can have the right total cost for the wrong frequency and severity reasons. Separate models are useful when rating factors affect the chance of a claim differently from its size. A direct loss-cost model can be a useful benchmark, particularly if component models have sparse claim data or unstable multiplication effects.

Worked example: A hypothetical group has modeled frequency 0.08 claims per earned exposure and mean severity $5,000 per covered claim. Its pure premium is 0.08 × $5,000 = $400 per exposure. If the severity sample excludes small claims but the frequency model counts them, $400 is inconsistent: either align the claim definitions or model the excluded payments separately.

Watch the distinction: A good fit to total losses does not prove both component models are sound. Offsetting frequency and severity errors can conceal poor deductible or limit pricing.

Reference: [3]

24. Balance base rates and relativities

A rating plan usually multiplies a base rate by class relativities. If the selected relativities have an exposure-weighted average above one, using the desired portfolio average premium as the base rate will overshoot the target. Calculate the average relativity using the prospective exposure distribution, then divide the target average premium by it. Recheck that the resulting class rates reproduce the target when weighted by those exposures. This balancing step changes the scale of the rates while preserving their ratios; it cannot establish that the ratios themselves are appropriate. If the portfolio mix changes, the average premium can change even with unchanged filed rates. For multi-factor plans, calculate the average of each risk's combined relativity rather than simply multiplying averages of separate factors.

Worked example: Suppose prospective exposures are 75% class A and 25% class B, with relativities 1.00 and 1.40. Their weighted average is 0.75 × 1.00 + 0.25 × 1.40 = 1.10. For a target average premium of $550, the base is $550 ÷ 1.10 = $500. Class rates are $500 and $700; 0.75 × $500 + 0.25 × $700 checks to $550.

Watch the distinction: Setting the base rate to $550 would produce a $605 portfolio average under this mix. Relativities must be balanced before interpreting a base rate.

Reference: [3]

25. Interpret rating-factor interactions

An interaction means one rating factor's effect depends on the level of another factor. With a log link, main-effect coefficients and the interaction coefficient add on the log scale, so their exponentiated relativities multiply. Compare the combined effect with what the main effects alone predict; the interaction factor below one moderates the joint effect, while one above one amplifies it. Interpret each main effect at the reference level of the other variable. Before adopting an interaction, examine exposure and claims in every relevant combination: a large apparent effect from a tiny cell may be unstable. Check that the combination also makes insurance sense and improves validation results. Interactions complicate rate tables, so evaluate the gain against volatility and implementation burden.

Worked example: A hypothetical model assigns urban territory 1.20, commercial use 1.30, and their joint interaction 0.90. A commercial urban risk has relativity 1.20 × 1.30 × 0.90 = 1.404 against the reference risk. Without the interaction it would be 1.56. The 0.90 factor reduces the combined indication by 1 − 0.90 = 10%, conditional on both characteristics.

Watch the distinction: Adding 1.20, 1.30, and 0.90 gives no valid relativity. In a log-link model, coefficients add before exponentiation; the displayed factors multiply.

Reference: [3]

26. Separate significance from rate impact

Statistical significance asks whether the data distinguish a factor effect from zero on the fitted scale. Practical rate impact asks how much premium would change and whether that change improves decisions. A narrow interval around a small relativity can show convincing evidence of a modest effect. Evaluate its portfolio premium impact, cell-level changes, stability across periods, predictive improvement, and cost of administering the classification. Also check whether the variable is a proxy for another rating characteristic or reflects a temporary mix shift. An interval describes uncertainty under the fitted model and data assumptions; it does not establish causal validity or guarantee future performance. Conversely, a large estimated effect with wide uncertainty may warrant pooling, more data, or a cautious transition rather than immediate full implementation.

Worked example: Suppose a fitted factor has relativity 1.02 and an illustrative 95% interval from 1.01 to 1.03. The interval excludes 1, yet the central indication is a 2% increase. On a $600 base premium, that is $12. Review whether the factor improves validation and supports a stable, usable class before deciding that $12 warrants a separate rating variable.

Watch the distinction: A small p-value is not a pricing decision. A well-estimated effect can be too small, unstable in use, or poorly justified as a class distinction.

Reference: [3]

27. Diagnose model fit

Model fit must be checked where prices are applied, not just in aggregate. Compare observed and expected claims or losses by territory, size, time, and other material segments, using enough exposure to interpret differences. Residual plots can reveal curvature, omitted interactions, and changes in variance; deviance comparisons can show whether added terms improve fit relative to complexity. Start with data integrity and exposure definitions when a pattern appears. Then assess whether a missing factor, inappropriate link, or unstable experience explains it. A model may balance perfectly overall because positive and negative segment errors cancel. Persistent directional errors matter for cross-subsidy even when a single goodness-of-fit statistic looks acceptable. Use validation data to judge whether a proposed correction persists outside the fitting sample.

Worked example: A count model predicts 100 claims in each of two territories; observed counts are 130 and 70. Across both, predicted and observed totals equal 200, but observed-to-expected ratios are 1.30 and 0.70. The opposite errors suggest a territorial pattern. Check that exposure and claim coding are comparable, then test a territorial effect and its performance on separate data.

Watch the distinction: A portfolio observed-to-expected ratio of 1.00 does not establish calibration by class. Segment errors can cancel while systematically underpricing one group.

Reference: [3]

28. Validate a model on new experience

Validation asks whether a candidate model improves predictions for experience that did not determine its coefficients or tuning choices. Set aside data by time or another design suited to the intended use, then compare the candidate with a meaningful benchmark on the same observations. Examine overall calibration, material segments, predictive error, and stability of proposed rate changes. An observed-to-expected ratio near one overall is useful but can hide bias in a segment; a candidate can improve one segment while worsening another. Respect changes in coverage, claim reporting, and business mix between training and validation periods. Repeatedly selecting models based on one holdout set weakens its independence. The Exam 8 study guide can help organize these model evaluation tasks.

Worked example: On hypothetical holdout data, the benchmark has observed-to-expected ratios of 1.00 overall and 1.30 for large risks. The candidate has 1.02 overall and 1.05 for large risks. Large-risk calibration improves by 0.25 in ratio terms, while overall calibration moves 0.02 above one. Compare prediction error and other segments before choosing; these two ratios alone do not settle implementation.

Watch the distinction: Refitting or tuning on the validation period, then presenting that same period as independent evidence, overstates performance. Keep final evaluation separate from model selection.

Reference: [3]

29. Apply class credibility

Credibility blends a class's observed indication with a complement representing broader, relevant experience. Write the selected estimate as Z times the class indication plus (1 − Z) times the complement, where Z expresses the information assigned to the class data. Sparse or volatile experience generally receives less weight; the exact credibility procedure depends on the variance structure and purpose. Put both inputs on the same prospective cost and exposure basis first. The complement should reflect risks comparable to the class after known rating differences are handled. Otherwise shrinkage draws the estimate toward a biased target. Assess the resulting rate for stability and predictive performance; a mechanically computed weight does not make an invalid class definition sound. Credibility is also useful when a class's apparent change is driven by a few unusually large claims.

Worked example: A hypothetical small class indicates $600 loss cost per exposure. A comparable complement indicates $400, and the selected credibility weight is 30%. The blended indication is 0.30 × $600 + 0.70 × $400 = $180 + $280 = $460 per exposure. It lies $60 above the complement, reflecting limited support for the class's raw $200 difference.

Watch the distinction: Credibility is not a discount applied to observed losses. The remaining weight goes to a specified complement, whose relevance is essential to the result.

Reference: [3]

30. Assess a risk class's validity

A useful risk class separates expected costs in a way that remains credible when applied prospectively. Test whether the apparent difference survives adjustment for territory, limits, coverage, time, and other correlated factors. Examine exposure volume, claim count, uncertainty, and stability across periods; an extreme relativity based on a few risks may be sampling noise. Consider whether the characteristic has a plausible connection to insured loss, can be measured consistently, and produces a workable rate structure. Predictive association alone does not answer every classification question, especially when a variable mainly stands in for another factor. Compare the proposed class system with a simpler benchmark on new experience and inspect resulting rate changes. A sound decision weighs predictive value, credibility, and practical consequences together.

Worked example: A proposed class has fitted relativity 1.40 but only 20 earned exposures, mostly in one high-loss territory. After territory enters the model, its relativity falls to 1.05. The drop suggests that much of the original 40% difference reflected geography. With so little independent class experience, retaining a separate factor needs stronger validation.

Watch the distinction: A large fitted relativity does not prove a distinct risk class. Correlated rating factors and sparse exposure can create an apparent separation that disappears after adjustment.

Reference: [3]

31. Control territorial confounding

Territorial comparisons are confounded when territories contain different mixes of limits, classes, or other risk features. A raw loss-cost ratio combines the geographic effect with those mix differences. To isolate a territorial indication, compare risks on a common rating-factor basis through stratification, standardization, or a predictive model that includes the relevant variables. Ensure the same geographic coding, exposure basis, and cost level in each area. Inspect whether the model has enough overlapping risks to support the comparison; a territory containing only high-limit policies offers weak evidence about a low-limit counterfactual. Geographic effects may also change with catastrophe exposure, claims practice, or time, so validate the adjusted indication prospectively. The controlled result supports a decision only if the adjustment variables themselves are reliable.

Worked example: Territory X has raw loss cost $500 per exposure and Y has $400, a ratio of 1.25. X contains more high-limit policies. After standardizing both territories to the same limit mix, their costs are $420 and $400. The controlled territorial ratio is $420 ÷ $400 = 1.05; the raw comparison attributed part of the limit effect to location.

Watch the distinction: Applying the raw 1.25 territorial factor on top of a separate limit factor can charge twice for the high-limit mix in territory X.

Reference: [3]

32. Handle many sparse classes

A variable with many levels creates cells with little experience, so unregularized estimates can swing sharply after one claim. First check whether adjacent or similar levels can be grouped using meaningful risk characteristics. Partial pooling or penalization then shrinks unstable level effects toward a common pattern while allowing well-supported levels to differ. The penalty strength can be chosen using separate validation data, with attention to how often levels appear and how new levels will be rated. Compare the resulting rates with a simpler group model and inspect whether rare classes receive implausible changes. Pooling trades some apparent fit in the training sample for lower prospective variance. It does not cure inconsistent coding or a large loss assigned to the wrong class, so data checks come first.

Worked example: One vehicle model has a single $20,000 claim over 20 earned exposures, giving a raw loss cost of $20,000 ÷ 20 = $1,000 per exposure. Its broader vehicle group indicates $400. A pooled estimate could fall between $400 and $1,000, depending on the model's credibility or penalty. One claim alone gives weak support for a stand-alone $1,000 indication.

Watch the distinction: Adding a parameter for every sparse level may improve training fit while worsening future rates. More detailed classes need enough information to support their distinctions.

Reference: [3]

33. Compare component and direct loss-cost models

A direct pure-premium model predicts expected claim cost per exposure. Component models instead predict claim frequency and conditional severity, whose product gives expected cost when definitions and target populations align. Compare both approaches on prospective calibration and stability, not just training fit. Components can reveal offsetting behavior that a direct cost model hides: one class may have more small claims, another fewer large claims. That difference matters when a deductible removes small claims or a higher limit admits more severe loss. Component models also introduce extra assumptions about claim counts, positive severity, and their relationship; sparse severity data can make the product unstable. Use consistent treatment of large losses, limits, and zero-claim risks when comparing indications. The preferable model is the one that supports the actual rating decision reliably.

Worked example: Class A has frequency 0.10 claims per exposure and mean severity $4,000; class B has 0.05 and $8,000. Both have pure premium $400: 0.10 × $4,000 = 0.05 × $8,000. A direct model sees equal mean cost. A per-claim deductible could affect the classes differently because their claim-size patterns differ.

Watch the distinction: Equal pure premiums do not imply equal responses to limits or deductibles. Those terms act on individual claim amounts, which may differ markedly between classes.

Reference: [3]

34. Integrate limits into rating

A policy limit caps each covered claim payment, so a limit factor compares expected capped loss at two limits on a consistent exposure and coverage basis. With ground-up severity X, expected payment at limit L is E[min(X, L)], subject to any deductible or other contract terms. Divide the higher-limit expectation by the base-limit expectation to obtain the factor. The factor depends on the severity tail: a class with more large claims can have a larger limit factor even if its base-limit loss cost matches another class's. Check whether claim data are censored at existing limits; observed paid amounts alone may conceal losses above them. Trend and inflation can also move claims across a limit, changing the factor prospectively. Validate the modeled tail because high-limit indications may be sensitive to few observations.

Worked example: Suppose a stated model gives expected capped loss of $300 per exposure at a $100,000 limit and $360 at a $200,000 limit. The additional expected cost is $60 per exposure, and the higher-limit factor is $360 ÷ $300 = 1.20. If another class has a heavier severity tail, applying this same 1.20 factor may understate its higher-limit cost.

Watch the distinction: A $200,000 limit does not imply twice the cost of a $100,000 limit. Only claims above the lower limit contribute extra payment.

Reference: [3]

35. Build an aggregate loss distribution

Aggregate loss is the sum of claim amounts across a random claim count. Conditional on a fixed count, independent claim severities contribute their means and variances add. Random count adds another source of variance because the number of severity amounts changes. If count N is independent of identically distributed severity X, then E[S] = E[N]E[X] and Var(S) = E[N]Var(X) + Var(N)(E[X])². A Poisson count has equal mean and variance, simplifying the second expression to E[N]E[X²]. Check dependence, exposure heterogeneity, and catastrophic common shocks before relying on that simplification; they can enlarge the tail beyond an independent compound model. For layer pricing, transform each claim payment by the layer terms before building its aggregate distribution.

Worked example: Assume annual count is Poisson with mean 2 and independent claim severity has mean $100 and variance $2,500. Aggregate mean is 2 × $100 = $200. Aggregate variance is 2 × $2,500 + 2 × ($100)² = $25,000 in squared dollars. Its standard deviation is √$25,000 ≈ $158.11; the variance includes severity and count uncertainty.

Watch the distinction: Multiplying count by severity variance alone gives $5,000 and omits random-count variation. The claim count can differ even when individual severity behavior stays fixed.

Reference: [3]

36. Price an excess layer

An excess layer pays only the part of an individual covered claim above attachment a, capped at layer width w: min(max(X − a, 0), w). Apply that function to ground-up severity before averaging; applying it to average severity generally gives the wrong answer because the function is nonlinear. Expected layer loss per exposure equals expected ground-up claim frequency times expected layer payment per ground-up claim, provided frequency and severity describe the same portfolio. Include zero layer payments from claims below attachment when estimating that mean. Then assess inflation, policy limits, claim handling, and tail data, since small shifts around attachment can materially change cost. Aggregate variability and expenses may affect the final price beyond expected layer loss. The FCAS practice questions provide another setting for working through such contract calculations.

Worked example: For a $100 excess of $100 layer, hypothetical ground-up claims of $80, $150, and $260 produce payments of $0, $50, and $100. Assuming these three severities are equally likely among claims, mean layer payment is ($0 + $50 + $100) ÷ 3 = $50 per claim. At 0.2 ground-up claims per exposure, expected layer loss is 0.2 × $50 = $10 per exposure.

Watch the distinction: Applying the layer formula to the average ground-up claim loses tail information. Claims below attachment and above exhaustion affect the mean differently.

Reference: [3]

37. Measure a deductible's effect

A per-claim deductible makes the insured retain min(X, d) and leaves the insurer max(X − d, 0), subject to any further limit. Calculate those amounts claim by claim because the deductible resets for every claim. Summing ground-up losses first and subtracting one deductible would treat several claims as one and overstate insured payment. For prospective rating, the relevant ratio is expected insured loss divided by expected ground-up loss under a suitable claim-size distribution. An observed ratio from a small sample may shift when severity mix or inflation changes. Also distinguish a per-claim deductible from an aggregate deductible, which accumulates losses across claims under different terms. Check whether expense treatment, claim counts, or behavior changes alter the final rate credit beyond the expected loss reduction.

Worked example: For hypothetical claims of $20, $70, and $130 under a $50 per-claim deductible, insurer payments are max($20 − $50, 0) = $0, $20, and $80. They total $100 against ground-up losses of $220, or $100 ÷ $220 = 45.45%. The insured retains $120. This sample ratio is not a universal deductible factor.

Watch the distinction: Subtracting one $50 deductible from the $220 aggregate gives $170, which is wrong. Three separate claims each face the deductible.

Reference: [3]

38. Experience-rate an individual risk

Experience rating asks how much an individual risk's own history improves the class prediction. First place that history on a prospective basis: adjust for exposure changes, claim development, trends, and any losses handled specially under the plan. Blend its indicated loss cost with the class expectation using a credibility weight tied to the information in the risk's experience. More stable, extensive experience can support a higher weight; a few volatile claims generally support less. The result is an expected loss cost, which the plan can translate into a premium modification while preserving its overall pricing objectives. Assess whether repeated changes encourage sensible risk control or merely respond to random fluctuations. Review how caps, large-loss limits, and changes in operations affect the indication.

Worked example: A risk's adjusted history indicates $700 loss cost per exposure; its class expectation is $500. With a selected credibility weight of 25%, the blended indication is 0.25 × $700 + 0.75 × $500 = $175 + $375 = $550 per exposure. The result is $50 above the class expectation, rather than the full $200 historical difference.

Watch the distinction: A single unusually large claim should not automatically move the full class rate to the risk's observed cost. Credibility reflects how informative its experience is.

Reference: [3]

39. Analyze retrospective rating

Retrospective rating adjusts premium after policy expiration according to covered loss experience and stated contract terms. A simplified calculation starts with a basic premium, adds a loss conversion factor times the defined loss measure, then applies minimum and maximum premium bounds. The loss measure may be limited, developed, or adjusted under the plan; using an inconsistent definition can dominate the arithmetic. Within the uncapped range, each additional dollar of covered loss changes premium by the conversion factor. At a bound, that marginal relationship changes until the formula returns inside the range. Examine how claim development, timing of adjustments, collectibility, and incentives affect the economic result. A retrospective premium indication is distinct from an unpaid claim liability: one concerns amounts charged, the other concerns claim payments owed.

Worked example: Assume a hypothetical plan: premium = min($400, max($180, $120 + 1.1 × covered losses)). At $40 of losses, the formula gives $164 and the minimum makes premium $180. At $100, premium is $230. At $300, the formula gives $450 and the maximum makes premium $400. Above the cap, further covered loss does not raise premium under these terms.

Watch the distinction: Using the uncapped formula at every loss level misses the plan's risk transfer. Minimum and maximum premiums change who bears additional loss.

Reference: [3]

40. Judge an individual rating plan

An individual rating plan can combine an experience adjustment, based on the risk's own claims, with a schedule adjustment for documented characteristics that the class rate misses. Evaluate whether each adjustment adds predictive information without charging twice for the same feature. Specify the order and basis of application: percentage changes applied multiplicatively differ from percentage-point changes added to one factor. Test resulting premium against expected claims, expenses, and the plan's target margin, both for the individual risk and across the portfolio. Also examine volatility, incentives for loss control, consistency of schedule judgments, and how the plan handles unusual claims. A plan can produce reasonable relative changes while leaving the overall price inadequate if its base level or caps are poorly set.

Worked example: A hypothetical plan starts with a $1,000 premium, applies a 10% experience debit, then a 5% schedule credit multiplicatively. Final premium is $1,000 × 1.10 × 0.95 = $1,045. If prospective expected claims plus expenses are $1,080, the result falls short by $35, or $35 ÷ $1,080 ≈ 3.24% of that cost target.

Watch the distinction: Adding 10% and subtracting 5% to get $1,050 ignores the stated multiplication rule. It also says nothing about whether either adjustment improves adequacy.

Reference: [3]

Exam 9: Risk Management for Actuaries

41. Fit and use an exposure curve

An exposure curve converts a ground-up severity model into expected loss at different policy limits. If X is claim severity, expected loss capped at limit u is E[min(X,u)]. The expected loss in a layer of width w attaching at a is E[min(X,a+w)] − E[min(X,a)]. Dividing that difference by the expected loss at a chosen reference limit gives a layer share. To fit a curve, first put losses on a comparable basis for claim size trend, deductibles and coverage, then check how well the fitted distribution reproduces observed limited losses at several limits. Sparse large claims make the upper curve especially uncertain. The Fall 2026 Exam 9 outline includes fitting loss distribution based exposure curves; a smooth fit alone does not establish that its tail is credible.

Worked example: Suppose a fitted severity model gives expected capped loss of 60 currency units at limit 100 and 90 at limit 200. A 100 excess of 100 layer has expected loss 90 − 60 = 30 units per claim. It represents 30 ÷ 90 = 33.3% of expected loss within the first 200 units. Compare that share with observed large claims before pricing the layer.

Watch the distinction: Subtracting attachment from the mean ground-up claim does not produce layer loss. Apply both caps to each severity outcome, then take the difference in expectations.

Reference: [4]

42. Price different reinsurance forms

Reinsurance cost depends on where the contract attaches and what unit it covers. A quota share transfers a stated fraction of each covered loss and usually a related share of premium. Per-risk excess transfers the part of an individual risk's loss above an attachment, up to the layer limit. Aggregate excess applies an attachment and limit to the combined covered losses over the contract period. Consequently, identical total losses can produce different ceded amounts when their sizes are distributed differently among risks. For pricing, calculate ceded loss under each form across a distribution of outcomes, then include expenses, uncertainty and contract provisions appropriate to the quote. A surplus share also depends on each risk's retained line; facultative coverage is negotiated for specified risks, while a treaty generally covers a defined portfolio. The Exam 7 study guide provides related reserving context.

Worked example: Two eligible risks have losses of 80 and 170 currency units, totaling 250. A 40% quota share cedes 0.40 × 250 = 100. A 100 excess of 100 per-risk layer cedes 0 on the first risk and 70 on the second. An aggregate cover attaching at 200, limited to 100, cedes min(250 − 200,100) = 50.

Watch the distinction: Applying a per-risk attachment to the portfolio total overstates or understates recovery. Calculate each covered risk separately before summing its ceded losses.

Reference: [4]

43. Measure contract provision effects

An excess layer may have only one original limit available during a contract period. A reinstatement restores some or all of that capacity after a covered loss, often for an additional premium. The economic recovery to the cedant is the layer payment less any reinstatement charge, although the charge and recovery should also be tracked separately for accounting and pricing. Wording matters: a charge can be proportional to the amount reinstated, the time remaining, or both. A loss corridor instead leaves the cedant responsible for losses within a specified band; it changes the loss transferred even when the headline attachment and limit appear unchanged. To evaluate a provision, model claims in sequence, apply each attachment and limit, update remaining capacity, then calculate charges and net cash flows. Price the resulting distribution, not a single illustrative loss.

Worked example: Assume a layer pays its full 100-unit limit on a first loss and is fully reinstated. Its original premium is 30 units, and the hypothetical charge is 20% of that premium for full reinstatement. The charge is 0.20 × 30 = 6 units. Net recovery from this first payment is 100 − 6 = 94 units before other contract terms.

Watch the distinction: A reinstated limit is not free additional cover. Ignoring its charge overstates net protection, while assuming full reinstatement can overstate capacity after a partial or excluded loss.

Reference: [4]

44. Evaluate a catastrophe model

A catastrophe model starts with a catalogue of possible events and their frequencies. Hazard describes event intensity at each location; exposure identifies insured property and values; vulnerability maps intensity and property characteristics to damage. Policy terms then convert physical damage into insured loss, including deductibles, limits and reinsurance where relevant. Summing event loss times event probability gives modeled annual expected loss, provided the event representation and probability basis are consistent. Tail measures require the full event loss distribution and treatment of multiple events, not merely that mean. Diagnose results by tracing large contributors through these stages: stale insured values affect exposure, inaccurate building attributes affect vulnerability, and incorrect deductibles affect financial loss. Compare model versions on identical portfolios to isolate changes. A plausible total can conceal offsetting errors among regions or construction types.

Worked example: A hypothetical event has a 1% annual occurrence probability and causes 20 million currency units of insured loss. Its contribution to annual expected loss is 0.01 × 20 = 0.20 million. Updating property values raises its modeled insured loss to 25 million, making the contribution 0.25 million. The increase is 0.05 million, conditional on unchanged event probability and financial terms.

Watch the distinction: Treating one event's expected-loss contribution as the portfolio's entire catastrophe provision omits other events. Also, a higher modeled loss may reflect updated values rather than worsening hazard.

Reference: [4]

45. Bring catastrophe loss into rates

A modeled annual catastrophe loss can be expressed as a rate provision by dividing it by the exposures expected to earn the rate, such as policy years or insured value. Match the model's coverage, territory, period and gross or net basis to the policies being priced. If rates include reinsurance cost separately, avoid also using a net catastrophe provision that embeds the same protection. Test changes in hazard frequency, property values, vulnerability and policy terms separately, because each can alter the indication for a different reason. An exposure-based average is a portfolio starting point: geographic concentration and construction type may require different provisions by class. Compare the modeled result with experience where credible, while recognizing that a short claim history has limited information about rare events.

Worked example: Suppose modeled annual catastrophe loss is 6 million currency units for 30,000 comparable policy years. The average provision is 6,000,000 ÷ 30,000 = 200 per policy year. Under a stated alternative vulnerability assumption, loss rises to 7.5 million, giving 250 per policy year. The 50 difference measures model sensitivity, assuming exposures and coverage stay fixed.

Watch the distinction: Dividing a gross modeled loss by net exposures mixes bases. Keep losses and exposures consistent, and do not interpret a change in vulnerability assumptions as observed claim development.

Reference: [4]

46. Classify and measure insurer risk

A risk taxonomy helps identify causes before selecting measures. Underwriting risk includes claim frequency, severity and reserve deterioration; market risk includes interest rate and asset price changes; credit risk includes counterparty failure; liquidity risk concerns meeting cash needs when due; operational risk includes system and process failures. Some events span categories: a catastrophe can generate underwriting loss, strain liquidity and coincide with reinsurer credit exposure. Record the exposure, owner, controls, horizon and plausible dependencies in a risk register. Then choose a measure suited to the decision. Expected loss supports routine pricing, a cash-flow stress tests liquidity, and a tail measure can inform capital. Aggregation requires joint scenarios or dependence assumptions; adding standalone tail figures without that structure can misstate company risk. The Exam 9 study guide is a useful companion for connecting these decisions.

Worked example: Under one hypothetical simultaneous scenario, a portfolio incurs a 10 million currency unit catastrophe loss, its bonds decline by 3 million, and a reinsurer default creates another 2 million loss. The scenario total is 10 + 3 + 2 = 15 million. This is one joint outcome, not an annual expected loss or a capital percentile.

Watch the distinction: Classifying risks in separate boxes does not make them independent. A common event can affect claims, investments, collections and available cash at the same time.

Reference: [4]

47. Compare VaR and tail value at risk

Value at risk at confidence level p is a loss threshold: under the left-quantile convention, it is the smallest amount whose cumulative probability reaches p. It reports a boundary but says little about loss severity beyond that boundary. Tail value at risk averages losses in the worst 1 − p share of outcomes. With a discrete distribution, a probability mass can straddle the cutoff, so specify how much of that mass enters the tail average. Both measures depend on loss definitions, time horizon and the modeled distribution. VaR is useful for a clear threshold; tail value at risk responds to increasingly severe outcomes above it. Examine stress scenarios alongside either measure, since scarce extreme observations and model choices can make a numerical tail estimate unstable.

Worked example: Let loss, in millions of currency units, be 0 with probability 90%, 10 with 5%, and 50 with 5%. Cumulative probability reaches 95% at 10, so 95% VaR is 10. The worst 5% consists entirely of the 50 outcome, making 95% tail value at risk 50 million. VaR hides that extra 40 million of tail severity.

Watch the distinction: VaR is not the largest possible loss. It identifies a quantile; outcomes beyond it can be much larger, which is why a tail average can matter.

Reference: [4]

48. Value insurance cash flows over time

Value each insurance cash flow at its expected payment date, then bring it to a common valuation date using a rate consistent with the cash flow's currency, timing and risk treatment. A cash amount due in one year has present value equal to that amount divided by 1 plus the annual discount rate. Premiums received early can earn investment income before claims are paid, while expenses may be immediate. Keep expected cash flows and risk adjustments distinct: discounting accounts for time value; a risk charge compensates for uncertainty or capital use under a stated framework. Be careful with inflation and rates: nominal cash flows require nominal discount rates, while real cash flows require real rates. For uncertain claim timing, discount each projected payment period separately rather than applying one average date without checking its effect.

Worked example: A policy brings in 120 currency units now, incurs 20 of expenses now, and has expected claims of 105 payable in one year. At a hypothetical annual discount rate of 5%, claim present value is 105 ÷ 1.05 = 100. Expected present value profit is 120 − 20 − 100 = 0 units before any risk charge.

Watch the distinction: A zero expected discounted profit does not imply adequate compensation for risk. The discount calculation handles payment timing; uncertainty and required capital may call for separate treatment.

Reference: [4]

49. Set a risk-loaded technical premium

A technical premium starts with expected claim payments and expenses on a consistent policy and time basis, then adds a defined risk load. The load may reflect a capital charge, a distortion of the loss distribution, or another explicit risk measure. State the method before calculating: otherwise the same label can hide different meanings. If capital is used, identify how much is allocated to the policy, the required return period, and whether investment income or discounting has already been credited. Avoid charging twice for the same uncertainty through both conservative expected losses and an additional margin. Compare the resulting target with the proposed premium and explain the gap. A positive expected underwriting margin can still fail a risk-adjusted target, especially for concentrated or highly volatile coverage.

Worked example: Per hypothetical policy, expected claims are 80 currency units, expenses are 15, and a stated capital method assigns an 8-unit risk load. Technical premium is 80 + 15 + 8 = 103. A quote of 100 is 3 below that target, although it exceeds expected claims plus expenses of 95 by 5. The issue is the required risk compensation.

Watch the distinction: Calling the risk load an additional expected claim confuses average cost with compensation for uncertainty. Show the components separately so the pricing decision remains interpretable.

Reference: [4]

50. Decompose underwriting return

Underwriting return can be decomposed into profit margin on premium and premium-to-equity leverage. Profit divided by premium measures how much of each premium unit remains after the costs included in the profit definition. Premium divided by equity measures business volume supported by each equity unit. Their product equals profit divided by equity, the corresponding return on equity, provided all three quantities cover the same period and profit basis. This identity explains why a modest premium margin can produce a high equity return when leverage is high. It also exposes risk: an adverse margin moves through the same multiplier. Before comparing insurers, check whether profit includes investment income, whether premium is written or earned, and whether equity reflects capital actually supporting that business. Differences in those definitions can overwhelm an apparent performance gap.

Worked example: An insurer has premium of 500 currency units, equity of 100, and profit of 20 on a consistent annual basis. Margin is 20 ÷ 500 = 4%; premium-to-equity leverage is 500 ÷ 100 = 5. Their product is 4% × 5 = 20% return on equity, matching 20 ÷ 100. A negative 4% margin would likewise imply negative 20% return.

Watch the distinction: A high return on equity does not by itself show superior underwriting. It may arise from greater leverage, which also magnifies losses when the margin turns unfavorable.

Reference: [4]

51. Explain the cost of insurance capital

Insurers hold capital to support losses that exceed expected claims and to maintain capacity to pay obligations. Supplying that capital has a cost: investors require compensation for its risk and for committing funds that could be used elsewhere. The required amount and return can vary with portfolio concentration, diversification, asset risk and the credibility of risk transfer. Financial distress adds costs beyond direct claim payments, including disrupted operations and reduced ability to write business. Information differences between insurers and buyers can also affect prices and access to capital, while competition limits how much cost can be passed through. For pricing, calculate a capital charge under a stated allocation and return assumption, then reconcile it with investment income and other margins already included. Capital cost is a decision input, not a universal percentage added to every policy.

Worked example: Two hypothetical insurers each expect claims of 90 currency units and expenses of 10 per policy. Insurer A allocates 40 of capital at a 10% annual cost, giving a 4-unit charge and a 104-unit target. Insurer B allocates 25 at the same rate, giving 2.5 and a 102.5-unit target, assuming no other adjustments.

Watch the distinction: Equal expected claims do not imply equal capital needs. Concentration, dependence, diversification and financial structure can change the risk borne by investors.

Reference: [4]

52. Allocate capital across lines

Standalone capital measures each line as though it carried its risk alone. Company capital can be lower than the sum because losses across lines do not always peak together. An allocation distributes the diversified company total among units so pricing and performance measures can reflect their contributions to overall risk. Marginal contribution asks how company risk changes when a line's exposure changes slightly; it may differ markedly from that line's standalone risk. The chosen allocation method should reconcile to total company capital and should be tested against dependence assumptions, especially common catastrophe or economic shocks. A line that looks modest alone can consume substantial capital if it rises in the same states as the rest of the company. Explain whether allocated amounts are intended for pricing, performance assessment or management incentives, since those uses can call for different judgments.

Worked example: Suppose standalone capital is 60 currency units for line A and 50 for line B, while company capital is 80. An illustrative allocation of 45 to A and 35 to B reconciles to 80 and assigns 30 units of diversification benefit relative to 60 + 50 = 110. The split still needs a contribution-based rationale; arithmetic reconciliation alone does not establish fairness.

Watch the distinction: Assigning each line its full standalone capital and then calling the sum company capital discards diversification. An arbitrary split that sums correctly can also misstate marginal risk.

Reference: [4]

53. Price for allocated risk capital

A risk-adjusted price can include expected claims, expenses and a charge for the capital tied up by the policy. The simplest charge multiplies allocated capital by a required annual return, but that is appropriate only when the allocation and return refer to the same period. Discount expected cash costs when timing is material, and account for expected investment income consistently. If the capital method already nets expected investment earnings, subtracting them again would understate the target. Allocation matters: two policies with equal expected claims may have different prices if one adds more to company tail risk. Check whether the proposed price is commercially achievable and whether any additional profit or tax components are specified by the framework. State each component so a reviewer can distinguish a change in expected loss from a change in risk appetite.

Worked example: A hypothetical one-year policy has expected claims of 70 currency units, expenses of 15, allocated capital of 30, and a required annual capital return of 12%. With costs treated on the same undiscounted annual basis, the charge is 0.12 × 30 = 3.60. Indicated price is 70 + 15 + 3.60 = 88.60 per policy.

Watch the distinction: Multiplying expected claims by the required capital return uses the wrong base. Apply that return to allocated capital, while keeping expected claims as a separate cost.

Reference: [4]

54. Assess business-unit economic performance

Economic profit compares a unit's realized or projected profit with the return required on the capital allocated to it. Calculate the capital charge as allocated capital times the required return for the matching period, then subtract that charge from profit. Positive economic profit means the unit earned more than this specified hurdle; negative economic profit means it fell short, even if accounting profit is positive. Compare units using consistent profit definitions, capital allocations and valuation dates. A unit with a high raw profit may consume enough capital to have lower economic profit than a smaller unit. Review whether unusual catastrophe experience, reserve changes or investment results are included before drawing a persistent performance conclusion. The measure can guide pricing and portfolio choices, but changing the allocation method can change the apparent ranking without changing underlying cash flows.

Worked example: Line A earns 18 currency units on allocated capital of 100 with a hypothetical annual required return of 12%. Its charge is 0.12 × 100 = 12 and economic profit is 18 − 12 = 6. Line B earns 10 on capital of 50, giving a charge of 6 and economic profit of 4. A creates 2 more units above its hurdle.

Watch the distinction: Ranking lines only by accounting profit ignores capital use. Ranking them only by percentage return can also obscure how many currency units they create above the hurdle.

Reference: [4]

55. Manage interest rate and credit exposure

Interest rate risk changes the market value of assets and liabilities as yields move. Modified duration gives a first-order price approximation: percentage value change is about negative duration times the yield change expressed as a decimal. Compare asset and liability durations and cash-flow timing, because an asset decline can be offset by a liability decline. Credit risk concerns a counterparty's failure or deterioration. A simple expected credit loss equals exposure at default times default probability times loss given default, subject to consistent horizons and recovery assumptions. Reinsurance receivables create credit exposure even though the underlying claims are insurance obligations. For decisions, supplement expected loss with concentration and stress measures: a small average default cost can coexist with a large loss if a major reinsurer fails after a catastrophe. Prepayment can alter both cash-flow timing and credit exposure.

Worked example: Assets worth 100 million currency units have modified duration 4. For a one percentage point yield rise, approximate value change is −4 × 0.01 × 100 = −4 million. Separately, a 10 million reinsurance receivable with 2% one-year default probability and 50% loss given default has expected credit loss 10 × 0.02 × 0.50 = 0.10 million.

Watch the distinction: Duration is an approximation for a specified yield move, not a forecast of investment return. Expected credit loss is an average, not the loss in a default event.

Reference: [4]

56. Unpack a structured-finance waterfall

A structured-finance pool channels asset cash flows to investors in a stated order. Senior tranches receive priority; junior tranches absorb initial losses and thereby provide subordination. To allocate a pool loss, reduce the junior balance first, then pass any excess to the next tranche. This waterfall changes who bears loss, but it does not eliminate the pool's total loss. Senior investors therefore face low losses over some scenarios and potentially sharp losses once junior protection is exhausted. Assess the pool's loss distribution, dependence among underlying assets, and the size and quality of credit enhancement. Correlated defaults can consume a junior buffer much faster than an independent-risk model suggests. Also distinguish principal loss allocation from interest payment priority, since the contract may specify different cash-flow rules.

Worked example: A hypothetical 100-unit asset pool funds a 20-unit junior tranche and an 80-unit senior tranche. With a 15-unit pool principal loss, junior investors lose 15 and retain 5; senior investors lose zero. With a 30-unit loss instead, juniors lose their full 20 and seniors lose 30 − 20 = 10, leaving senior principal of 70.

Watch the distinction: Calling a senior tranche protected does not make it loss-free. Protection lasts only while the subordinate buffer can absorb the pool's losses.

Reference: [4]

57. Evaluate a catastrophe bond

A catastrophe bond transfers defined catastrophe risk to investors through a security whose principal or interest can be reduced when a trigger is met. Trigger design determines how closely the payout tracks the insurer's actual claims. An indemnity trigger follows the sponsor's losses more closely but can involve reporting delay and moral hazard; an index or modeled-loss trigger can settle differently and create basis risk. Collateral arrangements affect the security of funds available after a trigger. Compare the bond with reinsurance using expected payout, tail protection, price, timing and counterparty exposure on a consistent basis. Model both insurer loss and trigger payout in the same events. A large nominal bond amount is not equivalent to that much effective protection when the trigger releases less than the insurer loses.

Worked example: Assume a hypothetical catastrophe bond has 50 million currency units of available principal. A qualifying event gives the insurer 70 million of claims, while the trigger releases 40 million from the bond. The insurer retains 70 − 40 = 30 million of those claims before other cover. The 10 million of unreleased principal shows why nominal size alone overstates protection for this event.

Watch the distinction: Equating the insurer's claim amount with the bond payout ignores trigger design. Basis risk can leave substantial retained loss even after a severe qualifying event.

Reference: [4]

58. Turn an ERM process into decisions

An enterprise risk management process begins by identifying material risks and assigning people accountable for them. It then measures exposure using data, scenarios and models that fit the decision horizon. Management chooses responses such as reducing exposure, strengthening controls, transferring risk or holding capital. Monitoring checks whether those responses actually change outcomes and whether new risks appear. Strategic and operational risks belong in the same decision process as insurance and financial risks, even when they are harder to quantify. A register should record causes, controls, indicators, owners and dependencies; a single score cannot replace that detail. Evaluate a proposal against a defined objective, such as expected value, liquidity or tail capital. Revisit assumptions after implementation, since behavior, systems and the external environment can change the risk distribution.

Worked example: A hypothetical digital claims system saves 2 million currency units in annual handling cost but could cause 5 million of disruption loss in a severe scenario. The 5 million is a scenario amount, not an expected annual cost. An ERM review assigns an operational owner, estimates the event probability and controls, and compares resulting savings and risk against the insurer's objective.

Watch the distinction: Subtracting the severe scenario loss directly from annual savings treats a possible event as certain. Estimate its likelihood and examine its liquidity and tail effects separately.

Reference: [4]

59. Model dependence across risks

Diversification depends on how risks move together, not just on their individual sizes. For two losses with standard deviations σ1 and σ2 and correlation ρ, aggregate variance is σ1² + σ2² + 2ρσ1σ2. Taking the square root gives aggregate standard deviation. Positive correlation raises the combined figure; negative correlation lowers it within the mathematical limits of correlation. Use a dependence assumption appropriate to adverse states, because ordinary-period estimates may miss common catastrophe or market shocks. Correlation captures linear association and does not fully describe joint tail behavior. For capital, test alternative dependence structures and compare the resulting tail measure, not only standard deviation. The FCAS practice questions can help reinforce the arithmetic, while scenario analysis tests whether the assumption makes business sense.

Worked example: Two hypothetical annual loss risks each have standard deviation 10 million currency units. At zero correlation, combined standard deviation is √(10² + 10²) = √200 ≈ 14.14 million. At correlation 0.8, it is √(100 + 100 + 2 × 0.8 × 10 × 10) = √360 ≈ 18.97 million. Stronger co-movement removes much of the diversification benefit.

Watch the distinction: Adding standard deviations assumes perfect positive correlation. Assuming zero correlation without evidence can understate aggregate risk when a common shock affects both lines.

Reference: [4]

60. Test a mitigation strategy

Evaluate a mitigation strategy against the insurer's stated objective and across several plausible models. A reinsurance cover may lower a tail loss or required capital while adding premium cost, basis risk, reinstatement cost and counterparty exposure. Calculate gross and net outcomes under the same scenarios, including contract limits and exclusions. Then compare the change in expected cost, liquidity needs and tail risk with the cover price. A tail-loss reduction is measured in loss units; it cannot be compared directly with a premium as though both were expected profit. Translate the reduction into the chosen decision metric, such as capital released or probability of breaching a risk limit. Test sensitivity to dependence and event severity, and monitor actual coverage after purchase. A strategy is effective when its residual risk and cost fit the objective, not merely when one modeled tail number falls.

Worked example: A hypothetical cover costs 6 million currency units and lowers modeled tail loss from 100 to 65 million, a 35 million reduction. Under an alternative dependence scenario, net tail loss is 85 million, so reduction is only 15 million. Both exceed the premium numerically, but that comparison does not establish value; assess capital relief, expected recoveries and residual risk under each scenario.

Watch the distinction: A large fall in one tail statistic does not prove that a cover pays for itself. Premium, expected recovery, capital benefit and remaining exposures answer different questions.

Reference: [4]

Official references and further reading

All study guides

Choose your exam version

CAS Exam 7

CAS Exam 8

CAS Exam 9

Exam 7: Advanced Estimation of Claims Liabilities

1. Build a reliable development triangle

2. Detect changes in case reserving

3. Make unlike claim cohorts comparable

4. Recognize calendar-period distortions

5. Project paid claims to ultimate

6. Project incurred claims to ultimate

7. Combine an expected loss with emergence

8. Update a reserve with Benktander credibility

9. Test development-factor assumptions

10. Estimate claims by coverage layer

11. Reserve a high-deductible obligation

12. Forecast retrospective premium

13. Estimate stochastic development parameters

14. Translate distribution parameters into tails

15. Interpret reserve prediction error

16. Construct a bootstrap predictive view

17. Model development with a GLM

18. Incorporate expert judgment probabilistically

19. Evaluate output and select a risk margin

20. Translate primary losses into reinsurance reserves

Exam 8: Advanced Ratemaking

21. Calculate a class loss cost

22. Specify a classification GLM

23. Choose models for frequency and severity

24. Balance base rates and relativities

25. Interpret rating-factor interactions

26. Separate significance from rate impact

27. Diagnose model fit

28. Validate a model on new experience

29. Apply class credibility

30. Assess a risk class's validity

31. Control territorial confounding

32. Handle many sparse classes

33. Compare component and direct loss-cost models

34. Integrate limits into rating

35. Build an aggregate loss distribution

36. Price an excess layer

37. Measure a deductible's effect

38. Experience-rate an individual risk

39. Analyze retrospective rating

40. Judge an individual rating plan

Exam 9: Risk Management for Actuaries

41. Fit and use an exposure curve

42. Price different reinsurance forms

43. Measure contract provision effects

44. Evaluate a catastrophe model

45. Bring catastrophe loss into rates

46. Classify and measure insurer risk

47. Compare VaR and tail value at risk

48. Value insurance cash flows over time

49. Set a risk-loaded technical premium

50. Decompose underwriting return

51. Explain the cost of insurance capital

52. Allocate capital across lines

53. Price for allocated risk capital

54. Assess business-unit economic performance

55. Manage interest rate and credit exposure

56. Unpack a structured-finance waterfall

57. Evaluate a catastrophe bond

58. Turn an ERM process into decisions

59. Model dependence across risks

60. Test a mitigation strategy

Official references and further reading

Frequently Asked Questions

Related Study Guides

Associate of the Casualty Actuarial Society (ACAS)

Fellow of the Society of Actuaries (FSA)

Associate of the Society of Actuaries (ASA)