Build the skills needed to turn insurance data into defensible predictive decisions: define targets, prepare exposure-consistent records, select and diagnose GLMs, and explain financial implications. This original preview contains twenty questions and sixty worked concepts allocated to Dealing with Data, Model Diagnostics and Selection, and Model Interpretation and Presentation using examination weights of 35%, 35% and 30%. It follows the supplied September 2026 version 8 outline for study as of October 16, 2026; no future version is included. Multiple-choice, multiple-select, numeric-entry and matching exercises represent the published examination mixture. Matching and numeric-entry answers use text-entry practice representations. Citations establish syllabus relevance, rather than independently verifying every technical explanation.
What this guide covers
Practice data preparation, GLM selection and actuarial communication with 20 questions and 60 concepts.
Dealing with Data
1. Turn a business decision into a defined target
A predictive target must specify the outcome, observation unit, time horizon and valuation basis. Claim count, positive claim severity and aggregate loss answer different questions even when drawn from the same claims system. Stakeholder discussion establishes when predictions will be used and which financial decisions they support. Define whether losses are paid or incurred, capped or uncapped, and measured at a common development age. These choices determine eligible records and interpretation. A technically accurate model of the wrong target can still produce an unsuitable business recommendation.
Worked example: A pricing team requests next-year expected water-loss cost per insured property. The analyst defines one property-year per row, includes zero-loss years, and uses consistently developed uncapped losses. A model of positive payments alone would omit claim occurrence.
Common mistake: Treating a convenient database field as the target can silently substitute paid-to-date amounts or positive claims for the business outcome.
Official syllabus reference: [2]
2. Prediction timestamps control feature eligibility
A predictor is eligible only if its value can be obtained at the moment the model will score a record. Historical extracts often contain later revisions, settlements or administrative codes that were unavailable at the original decision date. Such information creates leakage even if it is stored beside legitimate predictors. Define a scoring timestamp and reconstruct feature values as of that timestamp. Distinguish when an event occurred from when the organization learned about it. This temporal discipline makes retrospective evaluation resemble operational prediction.
Worked example: A quotation model uses prior claims reported before the quote. A claim that occurred earlier but was reported two months after quotation is excluded from that quote's historical claim count, because the insurer did not yet know about it.
Common mistake: Using an event date alone can admit information recorded later, overstating the performance achievable at the actual decision time.
Official syllabus reference: [2]
3. Declare the observation grain before combining data
The grain states what one row represents, such as a policy-year, coverage-period or individual claim. Variables must be summarized or attached consistently with that unit. A policy-level exposure repeated across multiple claim rows can inflate denominators if later summed. Conversely, collapsing claim rows prematurely can discard severity variation needed for individual-claim modeling. Write down the intended unique key and test it after each major transformation. Grain is a modeling decision because it determines the response distribution, weighting and meaning of one observation.
Worked example: A frequency file requires one row per policy-year. Three claim transactions for one claim are first consolidated to a claim identifier, then unique claims are counted by policy-year. Exposure is attached once, preserving the intended denominator.
Common mistake: Assuming every claims-system row represents a separate claim can confuse transactions with events and inflate the modeled frequency.
Official syllabus reference: [2]
4. Import types preserve identifiers and units
Import schemas assign meaning before modeling begins. Identifiers often belong in string fields because leading zeros and punctuation are part of the key. Dates need an explicit convention, while numeric amounts need consistent currency and scale. Missing-value tokens must be distinguished from legitimate categories or zeros. After parsing, inspect conversion failures, category counts and plausible ranges. A file that loads without an error can still contain silently altered information. Typed imports make these assumptions explicit and reduce downstream join and transformation failures.
Worked example: An exposure extract stores policy 000812, premium in cents and dates in day/month/year order. The analyst preserves the policy as text, divides monetary values by 100, and specifies the date format before checking annual premium totals.
Common mistake: Successful file loading does not establish semantic correctness; automatic conversions can alter keys, dates and monetary scales without stopping execution.
Official syllabus reference: [2]
5. Join cardinality determines whether totals survive
A join matches records through keys, and its cardinality describes whether each key appears once or many times in each table. Joining two tables with repeated keys can produce every matching combination, multiplying rows and financial amounts. Validate key uniqueness where expected, summarize detail tables to the required grain, and inspect unmatched records. Reconcile row counts, exposure and target totals before and after joining. A left join preserves left-side records but does not guarantee preservation of left-side row counts when the right key repeats.
Worked example: A policy has two exposure rows and three claims. Joining both tables directly on policy ID creates six rows. Summarizing each table to policy-year first produces one row with the correct total exposure and claim count.
Common mistake: Choosing a left join does not prevent duplication when several right-side records match each left-side key.
Official syllabus reference: [2]
6. Exposure defines the denominator of insurance rates
Exposure measures the amount of risk observed, such as earned vehicle-years or insured property-years. Frequency is claim count divided by compatible exposure, while pure premium is aggregate loss divided by exposure. The exposure definition must align with coverage, observation period and target eligibility. Partial periods require proportional exposure when that assumption is appropriate. Zero or negative exposure records require investigation because ordinary rates become undefined or misleading. Comparing raw claim totals across groups with different exposure confounds risk intensity with portfolio size.
Worked example: One territory has 30 claims over 300 vehicle-years and another has 45 over 900 vehicle-years. Their frequencies are 0.10 and 0.05 claims per vehicle-year, so the larger claim total belongs to the lower-rate territory.
Common mistake: Comparing claim counts without their exposure denominators can reverse the conclusion about which portfolio has the higher underlying rate.
Official syllabus reference: [2]
7. Reconciliation detects extraction and duplication errors
Reconciliation compares modeling-data totals with independently understood source totals under matching definitions. Useful checks include record counts, unique policies, earned exposure, claim counts and loss amounts by period or coverage. Differences should be decomposed into explained exclusions, timing differences and unresolved discrepancies. Duplicate detection requires a business key: identical amounts alone do not establish duplicate claims. Reconciliation does not prove every field is correct, but it can reveal broad extraction or joining errors that a model may otherwise absorb without warning.
Worked example: The ledger reports $12 million of paid losses, while the extract has $12.8 million. A period-by-period reconciliation finds $800,000 repeated after a transaction table join. Removing duplicated transaction keys restores the expected total without deleting legitimate equal-sized payments.
Common mistake: Deleting records merely because their amounts match can remove valid claims while leaving duplicates with slightly different fields untouched.
Official syllabus reference: [2]
8. Match claim valuation and development maturity
Insurance outcomes change as claims develop, so valuation timing can create artificial differences across cohorts. Recent accident periods may have lower reported counts or paid amounts simply because less time has elapsed. Define a consistent development age, use an appropriate developed target, or explicitly model the observation process. Paid, case-incurred and ultimate estimates are not interchangeable measurements. A calendar trend fitted to inconsistent maturity can capture reporting or settlement delay rather than underlying risk. Any adjustment should remain consistent with the intended prediction horizon.
Worked example: An analyst compares policies written last year with policies written four years earlier using current paid losses. The recent cohort looks cheaper partly because claims remain unsettled. Measuring both at twelve months removes one maturity difference before assessing remaining development needs.
Common mistake: A lower paid-loss average for recent policies does not by itself demonstrate improved risk when claim development ages differ.
Official syllabus reference: [2]
9. Distinguish missing values from observed zeros
Missingness means information is unavailable or inapplicable; zero is an observed value on a defined scale. Converting missing losses or claim counts to zero asserts that no loss or claim occurred. Investigate whether missingness arises from collection design, operational processes or the value itself. These mechanisms affect whether complete-case analysis or imputation can be credible. An unavailable mileage value may also carry predictive information through the collection process, but that relationship can change when procedures change. Preserve the original missingness state during preparation.
Worked example: A broker feed omits mileage when customers skip a question, while another feed records zero for vehicles in storage. Combining both as zero would make nonresponse look like nonuse. Separate missingness and verify how stored vehicles should enter the target population.
Common mistake: Replacing every blank numeric field with zero converts uncertainty about a value into a substantive claim about the policy.
Official syllabus reference: [2]
10. Fit imputation rules within training partitions
Imputation estimates replacement values from observed data, making it part of the fitted modeling pipeline. Medians, category modes and predictive imputers must be learned from training observations and then applied unchanged to validation data. During cross-validation, refit the imputer inside each training fold. A missingness indicator can distinguish an imputed value from an identical observed value. Simple imputation can reduce variability and does not automatically represent uncertainty about missing values. Evaluate whether the resulting pipeline remains appropriate for the deployment collection process.
Worked example: Training mileage has a median of 9,000 miles, while validation mileage has a median of 11,000. The deployed pipeline fills missing values with 9,000 in both sets and adds an indicator, reproducing what future scoring would actually do.
Common mistake: Computing an imputation median from the full data set allows assessment observations to influence the fitted preprocessing procedure.
Official syllabus reference: [2]
11. Separate erroneous observations from valid extremes
An outlier is unusual relative to a distribution or model, whereas a data error is incorrect. Large insurance losses may be valid and central to the expected-cost objective. Investigate source records, coverage terms, units and transaction structure before altering an extreme value. Correct confirmed errors using defensible evidence. For valid extremes, examine model influence, tail fit and sensitivity rather than applying automatic deletion. The appropriate treatment also depends on whether the target is uncapped loss, limited loss or another explicitly defined financial quantity.
Worked example: A $900,000 payment is flagged against a portfolio median of $4,000. Source review confirms a covered fire loss. A separate $90 million entry is traced to cents imported as dollars. The first remains relevant; the second receives a unit correction.
Common mistake: An observation's distance from the median is evidence of unusualness, not proof that the recorded insurance event is wrong.
Official syllabus reference: [2]
12. Capping losses changes the estimand
Capping replaces amounts above a limit with that limit and therefore changes the response being modeled. A model of min(loss, limit) estimates expected limited loss rather than expected uncapped loss. This may match a coverage layer or improve stability for a specific decision, but the removed tail must not disappear from an uncapped-cost recommendation. Distinguish legitimate contractual limits from analytical winsorization. Document the cap, its rationale and how predictions connect to the business target. Sensitivity analysis can quantify dependence on the chosen cap.
Worked example: Two claims of $200,000 and $800,000 become $200,000 and $250,000 under a $250,000 cap. Their average falls from $500,000 to $225,000. That new average describes limited payments, so an uncapped pricing target needs an additional tail treatment.
Common mistake: Presenting capped-model predictions as total expected loss ignores the portion deliberately removed from the response during preparation.
Official syllabus reference: [2]
13. Log transformations encode proportional predictor changes
Taking the logarithm of a positive predictor compresses large values and makes equal proportional changes equal distances on the transformed scale. For example, doubling from 1,000 to 2,000 produces the same log increment as doubling from 2,000 to 4,000. This is useful when diminishing absolute effects or multiplicative relationships are plausible. A predictor transformation is separate from the response family and link. Zeros and negative values require a justified alternative; adding a constant changes interpretation and depends on the variable's measurement scale.
Worked example: A building-value predictor enters as ln(value/100,000). Values of $100,000 and $200,000 map to zero and approximately 0.693. A coefficient then describes the modeled response change associated with a proportional value increase, rather than each additional dollar.
Common mistake: Logging a predictor does not mean the response is lognormal, and adding one to every variable is not a scale-neutral solution.
Official syllabus reference: [2]
14. Centering and scaling change coordinates, not information
Centering subtracts a reference value, while scaling divides by a chosen spread or unit. These operations can improve numerical conditioning and make polynomial terms or penalties easier to handle. An unpenalized model with an equivalent specification retains the same fitted predictions under a reversible linear rescaling, although coefficients and intercept interpretation change. Penalized models require deliberate scaling because penalties act on coefficient magnitudes. Learn empirical means and standard deviations within training data. Preserve the transformation parameters so scoring uses exactly the same coordinate system.
Worked example: Building area has training mean 2,000 square feet and standard deviation 500. An area of 2,500 becomes one standardized unit. The stored mean and spread also transform future properties; recalculating them on each incoming batch would change model meaning.
Common mistake: Standardizing a predictor does not remove skewness or create normality, and batch-specific scaling changes the coordinates expected by the model.
Official syllabus reference: [2]
15. Categorical encoding defines interpretable comparisons
A categorical predictor represents distinct levels rather than an assumed numeric progression. With an intercept, reference coding typically uses one indicator per nonreference level; each coefficient compares that level with the reference, conditional on other predictors. Including every level indicator alongside an intercept creates linear dependence. The reference choice changes coefficient presentation without changing equivalent fitted predictions. Specify handling for unseen scoring levels and check category consistency across sources. Numeric-looking labels such as territory codes should not acquire an ordinal slope unless that structure is substantively justified.
Worked example: Territories coded 10, 20 and 30 are geographic labels. Encoding them as two indicators with territory 10 as reference allows separate effects. Treating the codes as numeric instead forces the effect from 10 to 20 to equal that from 20 to 30.
Common mistake: A numeric storage type does not establish that category codes have meaningful distances or should share a single linear coefficient.
Official syllabus reference: [2]
16. Sparse-level pooling trades detail for stability
Rare categorical levels provide limited information for estimating separate effects, especially when exposure or claim counts are small. Pooling substantively similar levels can reduce variance, but arbitrary grouping can conceal real differences. Base pooling on business meaning, exposure support and training evidence, then assess predictive consequences. Define an operational rule for unseen levels. If response-driven grouping is used, learn it within training partitions to avoid leakage. A broad 'other' category may be heterogeneous, so its interpretation and future composition require particular care.
Worked example: Several vehicle-use codes each have fewer than ten vehicle-years, while two describe similar occasional commercial use. Pooling those two has a business rationale. Combining every rare code, including taxis and storage-only vehicles, could create a misleading average risk.
Common mistake: A shared low record count does not make rare categories substantively similar or guarantee that a pooled effect will transfer.
Official syllabus reference: [2]
17. Polynomial terms represent curved predictor effects
A polynomial expands a predictor into powers such as age and age squared, allowing curvature on the model's linear-predictor scale. In a log-link model, this curvature occurs in log expected cost rather than directly in dollars. Centering can improve conditioning and make the linear term refer to a meaningful age. Retaining lower-order terms usually supports coherent hierarchical interpretation. Higher degree is not automatically better: flexible polynomials can behave poorly near boundaries and outside the observed range. Validate both predictive improvement and the shape's business plausibility.
Worked example: A frequency model uses centered age a = driver age minus 40 and includes a and a squared. This can represent elevated risk at younger and older ages. The fitted curve is inspected across observed ages before extrapolating to unusually old drivers.
Common mistake: A significant squared term does not justify interpreting the linear coefficient as a constant age effect across the entire range.
Official syllabus reference: [2]
18. Interactions encode effects that depend on another variable
An interaction allows one predictor's effect to vary with another predictor. In a log-link model containing age, commercial use and their product, the age slope differs by use category. Main effects become conditional on the interacting variable's reference or zero value. Centering can make that reference meaningful. Examine joint exposure support because combinations with few observations yield uncertain effects. Adding interaction terms without checking available combinations can produce unstable or extrapolated predictions. Interpret the combined terms rather than reading each coefficient as an unconditional portfolio effect.
Worked example: A model includes centered age a, commercial indicator c and a×c. If age coefficients are −0.01 for a and 0.015 for a×c, the log-rate age slope is −0.01 for personal use and 0.005 for commercial use.
Common mistake: Once an interaction is present, the main age coefficient describes the reference use category rather than every insured driver.
Official syllabus reference: [2]
19. Weighted aggregation can reverse subgroup comparisons
An aggregate insurance rate is a denominator-weighted combination of subgroup rates. Groups with different exposure mixes can therefore have an overall comparison opposite to every within-stratum comparison, a pattern known as Simpson's paradox. Examine both composition and conditional rates before attributing differences to a territory or channel. Standardization applies a common mix when that answers the intended comparison. Neither raw aggregation nor stratification automatically proves causality; the correct presentation depends on whether the question concerns current portfolio results or comparable risks under a common distribution.
Worked example: Channel A writes mostly experienced drivers, while Channel B writes mostly new drivers. A can have higher frequency within both experience groups yet lower overall frequency because its portfolio contains far more low-frequency experienced-driver exposure.
Common mistake: An aggregate rate advantage does not establish a within-risk-group advantage when the portfolios have different exposure compositions.
Official syllabus reference: [2]
20. Univariate graphics reveal structure and support
Histograms, empirical quantiles and category-frequency displays describe a variable's shape, concentration and unusual observations before modeling. A histogram depends on bin choices, so confirm apparent clusters using alternative bins or quantiles. For insurance amounts, a log-scaled horizontal axis can reveal detail obscured by a long tail, but zeros need separate handling and labels must state the scale. Category charts should include exposure when record counts poorly represent risk volume. These displays identify preparation questions; they do not independently establish a response relationship.
Worked example: A severity histogram on a dollar axis places most claims in its first bin. A second view with a logarithmic dollar axis separates ordinary claims from several large-loss clusters, prompting checks of coverage type and units before changing records.
Common mistake: A visually prominent cluster may reflect binning or axis choice, so treating it as a proven risk segment overinterprets exploration.
Official syllabus reference: [2]
21. Data suitability requires provenance and limitations
Data suitability concerns whether available information can support the intended model, not whether every record is perfect. Establish provenance, collection definitions, coverage of the target population and known changes in operational systems. Compare available data with the decision's required variables and time horizon. Identify limitations that could materially affect results and distinguish verified facts from assumptions. The supplied outline lists data-quality and modeling standards as relevant reading, but specific professional obligations require their own texts. A preparation record should explain substantive choices so others can understand their consequences.
Worked example: An insurer's claims feed begins after a system migration and excludes one legacy product. The analyst documents that gap, limits the modeled population accordingly, and avoids describing validation results as evidence for the omitted product's expected losses.
Common mistake: A large data set is not automatically representative; systematic product omissions can matter more than the total number of records.
Official syllabus reference: [2]
Model Diagnostics and Selection
22. GLMs separate distribution, predictor and link
A generalized linear model combines a response distribution, a linear predictor and a link connecting the conditional mean to that predictor. The linear predictor can contain transformed variables, indicators and interactions even though it is linear in coefficients. The family describes support and mean-variance behavior; the link determines how predictors affect the mean. These choices perform different jobs and should be justified separately. A log link can accompany Poisson or Gamma models, for example, without making their response distributions equivalent. Diagnostics assess whether the combined specification is adequate.
Worked example: A count model uses Poisson variance, a log link and a predictor containing vehicle age and age squared. The quadratic term makes the age relationship curved, but the model remains a GLM because coefficients enter the linear predictor additively.
Common mistake: Calling a model nonlinear merely because it includes squared predictors confuses nonlinearity in variables with nonlinearity in estimated coefficients.
Official syllabus reference: [2]
23. Poisson frequency models connect counts to exposure
A Poisson GLM models nonnegative claim counts with conditional variance equal to conditional mean under the specified distribution. With a log link and log-exposure offset, expected count equals exposure multiplied by an estimated claim rate. This naturally accommodates different observation lengths. The equality of conditional mean and variance is an assumption to investigate, not a description guaranteed by count data. Heterogeneity, dependence or an inadequate mean function can create excess variation. Evaluate residuals and validation performance before deciding whether the Poisson specification is sufficiently useful.
Worked example: A fitted annual claim rate is 0.08 per vehicle-year. A policy with 0.25 vehicle-years has expected count 0.02. The count can be zero in observed data even though its conditional expectation is positive and fractional.
Common mistake: An expected count need not be an integer; confusing an expectation with an observed count leads to unnecessary rounding.
Official syllabus reference: [2]
24. Binomial models require a trial denominator
A binomial response counts successes among a known number of trials, with a modeled success probability. The logit link maps probabilities to log odds and ensures inverse-link predictions lie between zero and one. For grouped data, the trial count conveys precision: a proportion based on 1,000 offers contains more information than the same proportion based on ten. Conditional independence and a suitable probability structure matter. Extra-binomial variation can arise from dependence or unmodeled heterogeneity. A binary record is the special case of one trial.
Worked example: Two branches each renew 80% of offers. One has eight renewals from ten offers and the other 800 from 1,000. Supplying successes and trial denominators lets the binomial model distinguish their information content rather than treating the proportions equally.
Common mistake: Entering renewal proportions without their denominators can discard important precision information and misrepresent grouped binomial observations.
Official syllabus reference: [2]
25. Gamma models describe positive continuous severity
A Gamma GLM is suitable as a candidate for strictly positive continuous responses whose conditional variance scales with the square of the mean. With a log link, predictor effects multiply expected severity and fitted means remain positive. Exact zero payments are outside the ordinary Gamma response support, so their treatment must follow the response definition rather than an arbitrary small replacement. Gamma suitability does not follow from skewness alone. Examine residual behavior, large-loss performance and validation results, particularly when tail observations materially influence expected cost.
Worked example: A claims file separates closed claims with no payment from positive paid claims. The analyst fits a Gamma model to positive severities and uses a separate occurrence component when estimating overall expected cost, rather than replacing zero payments with one dollar.
Common mistake: A Gamma severity model fitted only to positive claims does not directly estimate average loss across policies that include zero-claim exposure.
Official syllabus reference: [2]
26. Lognormal severity needs response-scale retransformation
A lognormal model treats the logarithm of a positive response as normally distributed. Exponentiating its fitted log mean yields the conditional median, not generally the arithmetic mean needed for expected loss. Under constant normal log-error variance sigma squared, the mean is exp(eta + sigma squared/2). Other error structures require an appropriate retransformation approach. This differs from a Gamma model with a log link, which links directly to the response mean. When comparing approaches, assess predictions on the business-relevant dollar scale.
Worked example: A lognormal fit has eta = ln(1,000) and log-error variance 0.4. Its median severity is $1,000, but mean severity is 1,000×exp(0.2), about $1,221. Using the median for expected loss would understate this model's mean.
Common mistake: Exponentiating fitted log responses without a retransformation adjustment generally estimates a different quantity from expected dollar severity.
Official syllabus reference: [2]
27. Tweedie models combine zero mass and positive losses
For variance power between one and two, a Tweedie distribution admits a compound Poisson-Gamma representation with exact zeros and positive continuous aggregate amounts. This makes it a useful candidate for insurance pure premium or aggregate loss. A log link supplies positive expected values, while variance is proportional to the mean raised to the power parameter. Exposure treatment must match whether the response is total loss or loss per exposure. The distribution does not automatically resolve development inconsistency, dependence or poor predictor structure. Validate both aggregate and subgroup predictions.
Worked example: Most policy-years have zero loss, while others have continuous positive amounts. A Tweedie candidate models this combined response without inventing small positive losses for zero years, and its exposure weights are chosen consistently with the loss-per-vehicle-year response.
Common mistake: A Tweedie model's ability to include zeros does not make any nonnegative financial response compatible with its variance assumptions.
Official syllabus reference: [2]
28. Frequency-severity multiplication estimates expected cost
A frequency-severity approach estimates claim occurrence intensity and positive claim amount separately, then combines them into expected aggregate loss under an appropriate conditional structure. When expected severity per claim is compatible with the count process, expected loss per exposure equals expected frequency times expected severity. Conditioning and weighting matter: claim-level severity data and policy-level frequency data represent different sampling units. Dependence between count and severity may require additional structure. Separate components can reveal whether a cost difference arises from more claims, larger claims or both.
Worked example: A segment has predicted frequency 0.06 claims per vehicle-year and expected severity $5,000 per claim. Their product gives $300 expected loss per vehicle-year under the stated component assumptions. A half-year policy then has $150 expected aggregate loss.
Common mistake: Multiplying unrelated portfolio averages can be misleading when frequency and severity vary together across risk groups or conditioning sets.
Official syllabus reference: [2]
29. Offsets impose a known coefficient
An offset enters the linear predictor with its coefficient fixed, usually at one. A log-exposure offset in a count model imposes proportional scaling of expected count with exposure. An ordinary predictor, by contrast, has a coefficient estimated from data. Offsets can also incorporate a justified baseline prediction when modeling residual relativities, but their units and interpretation must be consistent. An offset is not a substitute for checking exposure quality. If proportionality is inappropriate for the process or observation design, the imposed structure may misrepresent expected outcomes.
Worked example: Two otherwise identical policies have exposures of one and two vehicle-years. With a log-exposure offset, the second receives twice the expected claim count. If log exposure were an estimated predictor with coefficient 0.8, that proportional relationship would no longer be imposed.
Common mistake: Treating an offset as a freely estimated covariate misses the substantive assumption encoded by fixing its coefficient.
Official syllabus reference: [2]
30. Deviance measures discrepancy from a saturated fit
Deviance compares a fitted model's likelihood with the likelihood of a saturated model that fits each observation as fully as the family allows. Smaller training deviance indicates better in-sample fit within a comparable setting, but adding parameters generally improves that fit. Deviance therefore needs complexity control or independent assessment for model selection. Null deviance describes an intercept-only baseline, while residual deviance describes the fitted specification. Comparisons depend on compatible observations, family conventions and weighting. Deviance magnitude alone does not establish calibration, causal validity or business usefulness.
Worked example: Adding a territory interaction lowers training deviance from 1,420 to 1,390. That confirms improved training fit, but the analyst also checks parameter cost and held-out deviance because the additional detail may be fitting noise rather than transferable structure.
Common mistake: The lowest training deviance identifies the closest training fit, not automatically the model with the strongest future performance.
Official syllabus reference: [2]
31. Pearson and deviance residuals reveal different discrepancies
A raw residual is observed response minus fitted mean. Pearson residuals scale that difference by the model's variance function, making discrepancies more comparable across fitted means. Deviance residuals use each observation's contribution to deviance and retain the direction of the discrepancy. Plotting residuals against fitted values and predictors can reveal curvature, changing spread or omitted groups. Distribution-specific patterns require care, especially for sparse counts. Residuals need not look identical across all response families, and a single visually unusual point should be investigated rather than automatically removed.
Worked example: Dollar residuals widen as fitted severity rises, as expected under a Gamma variance structure. Pearson residuals still show a curved age pattern, suggesting the age mean function needs refinement even though the raw widening alone is not surprising.
Common mistake: Interpreting all widening raw residual plots as misspecification ignores the increasing variance deliberately built into many insurance GLMs.
Official syllabus reference: [2]
32. Overdispersion can reflect variance or mean problems
Overdispersion means observed conditional variation exceeds that implied by a specified distribution, often discussed for Poisson counts. A Pearson statistic divided by residual degrees of freedom is one diagnostic, interpreted with model and sample context rather than a universal cutoff. Extra variation can reflect unmodeled heterogeneity, dependence, omitted nonlinearities or a different count distribution. Investigate the mean specification before attributing everything to the family. Variance corrections may improve uncertainty estimates without changing fitted means, whereas changing the distribution or predictors can also alter predictions.
Worked example: A Poisson model has a dispersion diagnostic near 2.3 and residual differences by fleet size. Adding a justified fleet-size effect reduces the diagnostic substantially. Remaining extra variation is then evaluated for dependence and alternative variance treatment.
Common mistake: An overdispersion diagnostic does not prove one specific replacement distribution or establish that the conditional mean is correctly specified.
Official syllabus reference: [2]
33. Leverage and influence answer different questions
Leverage measures how unusual an observation's predictor configuration is relative to the fitted design. Influence concerns how much model estimates or predictions change when an observation is removed or perturbed. A high-leverage observation need not be influential if its response agrees with the model, while a large residual can become especially influential when leverage is also high. Use influence diagnostics to prioritize investigation and sensitivity analysis. Verified unusual risks may legitimately affect the model; influence alone is not evidence that a record should be excluded.
Worked example: A very large warehouse is unusual in building area and construction mix, giving high leverage. Its loss matches the fitted expectation closely. Another unusual warehouse has a large unexplained loss and materially changes coefficients when omitted, making it more influential.
Common mistake: High leverage describes predictor position, so it should not be treated as synonymous with a large residual or a data error.
Official syllabus reference: [2]
34. Cross-validation evaluates the complete fitted pipeline
Cross-validation repeatedly trains on part of the available data and evaluates predictions on held-out folds. Every learned preprocessing step, including imputation, scaling, grouping and variable selection, belongs inside each training fold. Otherwise assessment data can influence the procedure being evaluated. Use folds that reflect dependence and deployment conditions rather than assuming random rows are independent. Aggregate performance consistently with the business unit and exposure. Cross-validation supports tuning, but repeatedly optimizing against it can still overfit its results, motivating a separate final assessment where feasible.
Worked example: A severity pipeline chooses a mileage transformation and imputes missing values. In each fold, those choices are estimated using only the fold's training data before predicting held-out claims. Precomputing transformations from all claims would evaluate a different, partially informed procedure.
Common mistake: Cross-validating only the final regression while preprocessing the full data set can produce optimistic performance despite apparently proper fold separation.
Official syllabus reference: [2]
35. Grouped and temporal holdouts mimic deployment
Assessment design should reproduce the information restrictions faced during deployment. Repeated policies, fleets or households can create dependence, so group separation may be needed to assess unfamiliar entities. Future-period deployment also calls for temporal separation because trends and operational changes may matter. Some tasks require both restrictions. A random split can be appropriate for certain exchangeable prediction settings, but its convenience does not establish suitability. Clearly state whether performance applies to existing customers, new customers, future periods or a combination of these populations.
Worked example: A model will score new fleets next year. The analyst trains on earlier periods and assesses later observations from fleets absent from training. A simple random policy-year split would allow the same fleet's characteristics to inform both sides.
Common mistake: A chronological split alone does not test unfamiliar customers when earlier observations from those same customers remain in training.
Official syllabus reference: [2]
36. Bias-variance tradeoffs explain underfitting and overfitting
Underfitting occurs when a model cannot represent important structure, creating systematic errors even with sufficient data. Overfitting occurs when fitted detail captures sample-specific variation that transfers poorly. More flexibility can reduce approximation bias while increasing estimation variance. Compare training and validation behavior, residual patterns and stability across resamples to distinguish these problems. A model with weak training and validation fit may need better structure or information; excellent training fit paired with poorer validation can indicate excessive flexibility. Complexity is useful only when supported by reproducible predictive improvement.
Worked example: A linear age term misses a U-shaped claim pattern in both training and validation. A quadratic improves both. A tenth-degree polynomial nearly fits training observations exactly but worsens validation, illustrating how additional flexibility can move from correction to overfitting.
Common mistake: Adding parameters is not inherently overfitting; the issue is whether added structure learns transferable signal rather than sample-specific noise.
Official syllabus reference: [2]
37. AIC and BIC penalize likelihood differently
AIC equals negative twice maximized log likelihood plus twice the estimated parameter count. BIC replaces the latter penalty with parameter count multiplied by the logarithm of sample size under its usual formulation. Lower values are preferred among appropriately comparable models. They answer related but distinct model-selection questions and can disagree. Use the same response observations and compatible likelihood definitions; transformed responses or quasi-likelihood fits may require special care. Neither criterion replaces validation or business assessment, and counting parameters must follow the fitted likelihood rather than an informal count of variable names.
Worked example: With 1,000 observations, adding two parameters costs four AIC units but approximately 13.82 BIC units. A log-likelihood gain of four reduces negative twice log likelihood by eight, so AIC favors the expansion while BIC does not.
Common mistake: Comparing printed information criteria across different response scales or incompatible likelihood conventions can create a meaningless ranking despite precise numbers.
Official syllabus reference: [2]
38. Nested tests and Type I errors need disciplined use
A likelihood-ratio test compares nested models through twice their maximized log-likelihood difference, using an appropriate reference distribution when regularity assumptions hold. It addresses whether added terms improve fit beyond sampling variation under the null. A Type I error rejects a true null hypothesis. Repeatedly testing many candidate terms increases opportunities for false discoveries unless the search process is accounted for. Statistical significance does not establish useful effect size or future performance. Test assumptions, hierarchy, predictive validation and business relevance should jointly inform inclusion decisions.
Worked example: An analyst tests twenty unrelated candidate interactions at a nominal 5% level. Even if none has a true effect, some small p-values can arise by chance. A selected interaction therefore needs more support than its isolated discovery-stage p-value.
Common mistake: A p-value is not the probability that the null hypothesis is true or that a selected predictor will improve future predictions.
Official syllabus reference: [2]
39. Multicollinearity destabilizes coefficient allocation
Multicollinearity arises when predictors contain strongly overlapping linear information. The model may struggle to determine which variable receives an effect, producing unstable coefficients, wide uncertainty intervals or sensitivity to small data changes. Predictions within the observed joint predictor region can remain stable because the combined contribution is better identified. Examine the design structure, correlations and resampling behavior. Possible responses include combining redundant measures, removing a less suitable variable or regularizing. Pairwise correlation alone can miss dependencies involving several predictors, and no universal correlation cutoff settles the decision.
Worked example: Building area and replacement value move closely together. Across resamples, one coefficient rises while the other falls, leaving their combined prediction similar. The analyst evaluates a cost-per-area formulation and a penalized model rather than interpreting either unstable coefficient in isolation.
Common mistake: Stable predictions do not guarantee stable individual coefficient interpretations when correlated variables can exchange their estimated contributions.
Official syllabus reference: [2]
40. Ridge and lasso regularize different aspects of complexity
Regularization adds a penalty to the fitting objective to trade some training fit for reduced complexity or instability. Ridge penalizes squared coefficient magnitude and generally shrinks correlated predictors without setting them exactly to zero. Lasso penalizes absolute magnitude and can set coefficients to zero, although selection among correlated variables may be unstable. Standardize predictors when appropriate so measurement units do not dominate the penalty. Tune penalty strength using a leakage-free validation procedure. Penalized coefficients reflect the chosen penalty and should not be interpreted as ordinary unpenalized estimates.
Worked example: Two nearly redundant property-value measures both contribute to a severity model. Ridge can retain and shrink both, stabilizing predictions. Lasso may select one and discard the other, with the selected measure changing across resamples when their information is similar.
Common mistake: A lasso-selected variable is not automatically the uniquely important underlying driver, especially when several predictors carry nearly interchangeable information.
Official syllabus reference: [2]
41. Calibration and ranking measure different performance
Calibration evaluates agreement between predicted and observed outcomes on the relevant scale. Ranking evaluates whether higher predictions identify relatively higher-risk observations. A model can rank risks well while systematically overpredicting their costs, or match the portfolio mean while failing to distinguish risks. Check both overall and subgroup calibration using suitable exposure or precision weights. Ranking summaries depend on the target and metric, and they do not directly quantify financial adequacy. Performance assessment should therefore include measures that reflect both the ordering and magnitude needed for the business decision.
Worked example: A model predicts every segment's cost at twice its true expectation. Its risk ordering is unchanged, so a ranking measure may remain strong. Yet its aggregate expected loss is doubled, making calibration poor for pricing or budgeting.
Common mistake: A strong ranking metric does not establish accurate dollar predictions or show that the portfolio's total expected loss is correct.
Official syllabus reference: [2]
42. Convergence checks distinguish a fitted model from a failed fit
Numerical software returning coefficients does not guarantee a reliable optimum. Inspect convergence status, extreme estimates, fitted-value ranges and sensitivity to starting conditions or scaling. Sparse categories can cause separation in binomial models, where coefficients grow without a finite ordinary maximum-likelihood estimate. Rank-deficient designs create nonidentifiable parameters. Correcting units, simplifying unsupported terms or using an appropriate penalized approach may address particular causes. A convergence problem differs from poor predictive performance: the former questions whether the stated fitting problem was successfully solved before its predictions are evaluated.
Worked example: Every policy in a tiny category renews, and the binomial fit reports an enormous coefficient with a warning. The analyst investigates separation and sparse support, rather than presenting the coefficient as compelling evidence of near-certain future renewal.
Common mistake: A very large coefficient with a large standard error may signal estimation failure rather than an exceptionally strong, well-established effect.
Official syllabus reference: [2]
Model Interpretation and Presentation
43. Convert log-link changes to response-scale percentages
For a GLM with log link, a predictor change of Δx changes the modeled conditional mean by the ratio exp(β × Δx), holding other predictors fixed and assuming no interaction involving that predictor. The percentage change is 100 × (exp(β × Δx) − 1). The parentheses matter: subtract one from the ratio before multiplying by 100. State the predictor's units because a coefficient applies to one modeled unit. A positive coefficient implies a higher conditional mean; a negative coefficient implies a lower one. These comparisons describe modeled means rather than guaranteed individual outcomes.
Worked example: A frequency model measures mileage in units of 5,000 miles and estimates β = ln(1.08). Comparing otherwise identical policies at 20,000 and 10,000 miles gives Δx = 2. The mean ratio is 1.08² = 1.1664, corresponding to 100 × (1.1664 − 1) = 16.64% higher expected frequency.
Common mistake: Treating β itself as the exact percentage change, or calculating 100 × exp(β) − 1, misstates the response-scale effect.
Official syllabus reference: [2]
44. Logit coefficients multiply odds rather than probabilities
A logit model represents log odds, where odds equal probability divided by one minus probability. Exponentiating a coefficient gives an odds ratio for a one-unit predictor change, conditional on other terms. To obtain the new probability, multiply baseline odds by that ratio and transform back using odds divided by one plus odds. The same odds ratio produces different probability changes at different baselines. This distinction is essential when communicating renewal or claim-occurrence models because a relative change in odds is not a percentage-point change in probability.
Worked example: Baseline probability 0.20 gives odds 0.25. An odds ratio of two raises odds to 0.50 and probability to one-third. Starting instead from probability 0.60 produces probability 0.75, so the identical odds ratio has a different absolute probability effect.
Common mistake: Multiplying baseline probability by an odds ratio can produce incorrect results and can even yield values above one.
Official syllabus reference: [2]
45. Reference levels and intercepts define the comparison baseline
An intercept describes the linear predictor when numeric predictors equal zero and categorical predictors are at their reference levels, with offsets handled as specified. This baseline may be unrealistic unless numeric variables are meaningfully centered. Changing a categorical reference redistributes coefficients without changing equivalent predictions. In a log-link model, the ratio between two nonreference levels is the exponential of their coefficient difference. Clearly identify the reference and units in output. A coefficient table without these details can invite comparisons that the fitted model does not actually represent.
Worked example: Territory B and C have log coefficients 0.10 and 0.30 relative to A. The modeled C-to-B mean ratio is exp(0.20), approximately 1.22. Comparing their individual exponentiated coefficients as percentage-point differences would misstate the relationship.
Common mistake: An intercept is not necessarily the portfolio average, particularly when zero-valued predictors describe an uncommon or impossible risk.
Official syllabus reference: [2]
46. Interaction effects require combined coefficient calculations
When an interaction is included, a main effect describes a comparison only at the interacting predictor's reference value. For a log-link model with coefficients beta for a category and gamma for its interaction with age a, the category ratio is exp(beta + gamma a). Presenting exp(beta) as universal would hide how the comparison changes. Use meaningful age values or a plotted response surface, with uncertainty where practical. Check whether those combinations are supported by data, because apparently precise algebra can still describe an extrapolation.
Worked example: A commercial-use coefficient is 0.20 and its interaction with age centered at 40 is −0.01. At age 50 the commercial relativity is exp(0.20 − 0.10), about 1.105. At age 30 it is exp(0.30), about 1.35.
Common mistake: Reporting a main-effect relativity without its interaction can conceal genuine variation and imply one effect where the model estimates several.
Official syllabus reference: [2]
47. Confidence intervals and prediction intervals cover different uncertainty
An interval for a conditional mean reflects uncertainty in estimating the expected outcome for specified predictors. A prediction interval for a future observation also includes the outcome's inherent random variation, so it is generally wider under comparable assumptions. Insurance losses can be highly variable even when the expected mean is estimated precisely. State the interval type, scale and assumptions rather than labeling every range as uncertainty. Neither interval automatically captures model misspecification, distribution shift or omitted development uncertainty. Business readers need to know which source of variation the range represents.
Worked example: A segment's expected severity is estimated near $4,000 with a relatively narrow mean interval. An individual future claim can still range widely because claim amounts vary substantially. A narrow interval for the mean is not a promise about the next payment.
Common mistake: Treating a confidence interval for expected severity as the likely range of an individual claim understates outcome variability.
Official syllabus reference: [2]
48. Standardized predictions create comparable portfolio summaries
Standardization compares model predictions using a common distribution of other predictors. Score the same reference records under each comparison condition, then average response-scale predictions with appropriate weights. This separates a modeled comparison from differences caused by portfolio composition. With nonlinear links or interactions, exponentiating an average linear predictor is generally not equivalent to averaging predicted means. State the reference population because a standardized effect can change when its composition changes. Standardization is a descriptive model-based comparison and does not alone establish an intervention's causal effect.
Worked example: To compare two territories, an analyst predicts both territory assignments for the same 1,000 properties and averages expected dollar losses. This controls the reference mix of building ages and values instead of comparing the territories' different observed property portfolios.
Common mistake: Averaging log-scale predictions and exponentiating generally does not reproduce the arithmetic average expected dollar loss required for a portfolio comparison.
Official syllabus reference: [2]
49. Aggregate financial predictions on compatible units
Portfolio expected loss is the sum of record-level expected amounts. When a model predicts loss per exposure, multiply each prediction by its exposure before summing. A simple average of rates weights records equally and may differ substantially from the exposure-weighted portfolio rate. Keep counts, rates, severities and aggregate dollars distinct throughout presentation. Expected-value aggregation does not require outcomes to be independent, although portfolio uncertainty calculations generally depend on their covariance structure. Reconciliation of modeled totals helps identify unit mismatches and duplicated exposures before business results are communicated.
Worked example: One policy has rate $200 per vehicle-year and exposure 0.25; another has rate $400 and exposure two. Expected losses are $50 and $800, totaling $850. The portfolio rate is $850 divided by 2.25, about $377.78.
Common mistake: Averaging predicted rates without exposure weights can misstate the portfolio rate even when every individual prediction is correctly calculated.
Official syllabus reference: [2]
50. Coefficient plots show effect magnitude and precision
A coefficient plot places estimated effects on a common axis and displays uncertainty intervals, making magnitude and precision easier to compare than a dense table. For log-link effects, exponentiated coefficients use a ratio scale with a no-difference reference at one. Label the reference categories and whether estimates are adjusted for other predictors. Intervals crossing one indicate compatibility with no ratio difference under the stated inferential assumptions, not proof of equivalence. Correlated estimates and interactions may require joint interpretation rather than ranking isolated point estimates.
Worked example: Three construction relativities are 1.10, 1.25 and 1.40, but the last has a much wider interval because exposure is sparse. A coefficient plot makes that uncertainty visible, discouraging management from treating the largest point estimate as the most reliable effect.
Common mistake: Ranking point estimates alone can confuse a large but imprecise effect with a smaller effect supported by substantially more information.
Official syllabus reference: [2]
51. Calibration plots need denominators and comparable scales
A calibration display compares observed outcomes with predicted means for groups such as risk bands, periods or business segments. For claim rates, both observed and predicted quantities must use compatible exposure denominators. Show exposure or uncertainty so readers can distinguish stable discrepancies from sparse-group noise. Groups formed by predicted risk answer a different question from groups formed by a business factor; both can be useful. Aggregate agreement can conceal local errors. Calibration assessed on training data is usually less persuasive than agreement on a suitably independent assessment population.
Worked example: The highest-risk band has observed frequency 0.30 and predicted frequency 0.20, but only ten vehicle-years. The adjacent band has a smaller discrepancy over 20,000 vehicle-years. Showing exposure prevents visual gap size alone from determining the investigation priority.
Common mistake: Equal-sized plotted markers can imply equal evidence even when the underlying exposure differs by several orders of magnitude.
Official syllabus reference: [2]
52. Lift charts communicate ranking relative to a baseline
A lift display sorts observations by predicted risk, groups them and compares observed outcomes across the resulting bands or against a portfolio baseline. It shows whether the model separates higher-risk from lower-risk business. Define whether groups contain equal records, equal exposure or another quantity, since that affects interpretation. Observed outcomes should come from an assessment population where possible. Strong lift does not establish calibration or profitability: a model can order risks correctly while predicting the wrong absolute amounts. Clearly distinguish model-based grouping from an independently observed outcome comparison.
Worked example: A held-out portfolio's top predicted-risk quintile has twice the portfolio claim rate, while the bottom quintile has half. This supports risk separation. The analyst still checks predicted-versus-observed rates because these lift results do not show whether dollar estimates are accurate.
Common mistake: A steep lift chart demonstrates separation under its construction, not automatic rate adequacy, causal interpretation or financial benefit.
Official syllabus reference: [2]
53. Axes and visual encodings control perceived differences
A chart's axis range, scale and encoding affect how readers perceive comparisons. Bar length conventionally represents magnitude, so truncating its baseline can exaggerate differences. Position-based displays such as scatterplots may appropriately use a narrower range when clearly labeled. Logarithmic axes represent ratios rather than equal absolute increments. Avoid decorative area or volume encodings that distort numerical comparisons. Select the display for the analytical question and label units directly. A visually dramatic chart can still be misleading even when every plotted number is correct.
Worked example: Loss ratios of 68% and 70% look radically different as bars starting at 67%. A clearly labeled dot plot or a zero-based bar chart presents the two-percentage-point gap more faithfully, depending on whether the audience needs detail or magnitude.
Common mistake: Accurate data labels do not fully correct a visual encoding that encourages readers to perceive a much larger difference.
Official syllabus reference: [2]
54. Tables preserve exact values and denominator context
Tables are useful when readers need exact values, several related measures or auditable calculation inputs. Use explicit units, consistent precision and a meaningful row order. Pair rates with their exposure or trial denominators when precision differs materially across rows. Distinguish percentages from percentage-point changes and label whether amounts are per claim, per exposure or total dollars. Totals should reconcile under the stated weighting rules. Excess decimal places suggest false precision, while excessive rounding can obscure reconciliation. A concise explanatory title should describe the comparison without overstating the conclusion.
Worked example: A territory table includes vehicle-years, claims, frequency and predicted frequency. Reporting 0.083 rather than 0.0834729 improves readability, while retaining claim and exposure totals lets technical readers verify the rate and assess whether a territory is sparsely supported.
Common mistake: A rate column without units or denominators can invite incorrect comparisons even when all displayed arithmetic is correct.
Official syllabus reference: [2]
55. Technical explanations connect choices to consequences
A technical model account should explain the target, data scope, transformations, family, link, exposure treatment and assessment design, then connect these choices to diagnostic evidence. Reporting only final coefficients hides how the model was refined and what alternatives were considered. Explain why a variable was included or excluded using availability, stability, predictive contribution and business relevance. Distinguish empirical findings from assumptions. The supplied outline supports these communication competencies; it does not make any single modeling recipe universally correct. Useful technical writing enables another analyst to understand and challenge the reasoning.
Worked example: Instead of saying 'the Gamma model fit best,' an analyst states that positive severity and approximately quadratic variance motivated Gamma, residual curvature prompted an age term, and held-out deviance improved without material subgroup miscalibration. The explanation links each decision to evidence.
Common mistake: A list of software commands or final coefficients does not explain why the modeling decisions address the business question.
Official syllabus reference: [2]
56. Expected business value combines predictions with consequences
A model predicts outcomes; a decision rule combines those outcomes with their financial consequences. Expected contribution may depend on retention probability, premium, conditional losses and variable expenses rather than predictive accuracy alone. Define the objective and avoid mixing per-offer values with per-retained-account values. Include only the cash flows relevant under the stated scenario, and recognize assumptions about customer response or risk selection. A more accurate model can still support a poor decision if the objective is misdefined. Communicate the recommendation alongside its key economic assumptions.
Worked example: Offer A retains 800 accounts with $200 margin each, yielding $160,000 expected contribution. Offer B retains 680 with $250 margin each, yielding $170,000. The lower-retention offer wins this objective, although it may not maximize premium or customer count.
Common mistake: Optimizing a convenient prediction metric or gross premium can select a different action from maximizing the business's stated expected-value objective.
Official syllabus reference: [2]
57. Adjusted associations do not automatically identify causal effects
An observational predictive model estimates relationships conditional on recorded variables. Even extensive adjustment can leave unmeasured confounding, selection effects or reverse causation. A coefficient can therefore support prediction without establishing what would happen if management intervened on that variable. Causal interpretation needs a defensible identification strategy, such as randomization or a suitable quasi-experimental design with its assumptions. Communicate the estimated association accurately while separating intervention claims. This distinction matters when recommending safety equipment, marketing actions or underwriting changes based on historical customer choices.
Worked example: Homes with leak sensors have lower modeled water losses, but their owners may also maintain plumbing more carefully. The coefficient is useful descriptive evidence. A controlled installation program would better assess how much installing sensors itself changes future losses.
Common mistake: Holding observed predictors fixed does not ensure comparable unobserved characteristics or convert a regression coefficient into a causal treatment effect.
Official syllabus reference: [2]
58. Extrapolation and distribution shift limit transportability
Extrapolation applies a fitted relationship beyond the predictor combinations supported by training data. Distribution shift occurs when deployment differs in predictor mix, outcome mechanisms or data collection. A model may perform well historically yet fail after repair-cost inflation, underwriting changes or a new distribution channel. Separate a new mix of familiar risks from a changed conditional relationship, because their implications differ. Present the assessment population and known deployment differences. Algebraically valid predictions outside support can still be poorly supported empirical claims, especially with flexible interactions or polynomials.
Worked example: A property model trained mostly on homes below $1 million is asked to price $10 million estates. Its log-link formula returns positive predictions, but those values rely on extrapolated value effects and potentially different coverage and claims processes.
Common mistake: A prediction's numerical availability does not establish that the training data support its risk combination or future operating environment.
Official syllabus reference: [2]
59. Sensitivity analysis reveals decision-critical assumptions
Sensitivity analysis changes plausible assumptions or modeling choices to examine how conclusions respond. Examples include alternative loss caps, development assumptions, trend rates, family choices or treatment of influential observations. Distinguish this from estimating sampling uncertainty: both matter, but they answer different questions. Focus on whether the recommended action changes, not merely whether a coefficient moves. Scenarios should be explicit and substantively plausible rather than selected to favor a preferred result. A recommendation that remains stable across credible alternatives is more defensible than one dependent on a single fragile input.
Worked example: Two pricing offers differ by $10,000 in expected contribution. Varying retention assumptions within a plausible range reverses their ranking, while reasonable loss-trend changes do not. Management learns that customer response, rather than trend precision, is the decision-critical uncertainty.
Common mistake: A narrow statistical interval does not guarantee a robust recommendation when untested structural or economic assumptions can reverse the decision.
Official syllabus reference: [2]
60. Audience-specific summaries preserve the same quantitative claim
Technical and nontechnical presentations should communicate the same finding at different levels of detail. A business summary states the decision, expected consequence, main evidence and material limitation in familiar units. Supporting technical material explains the model, assessment population and assumptions. Simplification should remove unnecessary machinery without converting association into causation or uncertainty into certainty. Keep percentages, dollar effects and population scope consistent across displays and prose. A persuasive recommendation connects the model to the business objective while making clear which conditions would change the conclusion.
Worked example: A technical report gives a log-link coefficient, validation calibration and sensitivity results. The management summary states that the selected offer raises expected contribution by $10,000 under the supplied retention assumptions, then identifies retention uncertainty as the main reason the ranking could change.
Common mistake: Removing caveats that could change the recommendation is not effective simplification; it changes the substance of the quantitative claim.
Official syllabus reference: [2]