RESEARCH DOSSIER
Exploratory Data Analysis and Statistical Standardization in Insurance
Abstract. Exploratory data analysis (EDA) in insurance establishes the statistical identity of the observations from which premiums, claims, reserves, risk indicators, and predictive assessments are later derived. Its task extends from visualization of distributions to verification of denominators, contractual payment transformations, temporal alignment, administrative reporting mechanisms, portfolio segmentation, and comparability across businesses and jurisdictions. This monograph develops an integrated analytical account of insurance EDA and statistical standardization while preserving their distinct inferential roles. It examines policy, coverage, insured object, person, risk interval, event, claim, payment, and development record as noninterchangeable observational units; derives exposure-adjusted frequency, severity, and loss-cost measures; considers heavy tails, concentration, censoring, deductibles, and limits; designs diagnostic displays; evaluates outliers, missingness, mixture effects, and changing composition; and specifies exploratory controls before statistical modeling. Synthetic numerical examples demonstrate how aggregate claim rates can reverse segment-level improvements, how a small number of large claims dominate loss expenditure, and why claim frequency and severity require separate denominators. Comparative reporting examples use institutional methodological disclosures rather than inferred equivalence between national indicators. The methodological conclusion is that an insurance statistic becomes interpretable when its unit, event definition, observation window, exposure basis, coverage boundary, sampling mechanism, currency basis, segment, and revision state are jointly specified. This is an analytical synthesis of established statistical and actuarial methods. It introduces no new named statistical theory, model, doctrine, or proprietary data standard.
Keywords: Exploratory data analysis; insurance statistics; actuarial data quality; exposure standardization; claim frequency; claim severity; loss cost; portfolio composition; Simpson's paradox; heavy-tailed losses; concentration; censoring; truncation; outliers; reporting delay; statistical metadata; reproducibility.
Introduction
Insurance creates records through contracts, risk-bearing periods, insured events, claims handling, payments, and regulatory reporting. These records become statistical observations only after the analyst specifies what each row represents and which processes make a row observable. The difference is consequential. Ten payment entries may concern one claim; one claim may involve several insured objects; a contract may cover several risks; a single catastrophe may cause thousands of claims; a portfolio may gain volume while the risk rate of every existing segment falls. A sum or mean computed before these relationships are understood is arithmetically valid and scientifically ambiguous.
John W. Tukey's Exploratory Data Analysis (1977) established a tradition of using resistant summaries, graphical inspection, and iterative questioning to examine the structure of observations. Heckert and Filliben's NIST exploratory methods chapter explicitly places the discovery of structure, unusual observations, and violations of assumptions among the purposes of EDA (2003). Insurance adds institutional complexity: policies and claims are partly generated by contractual rules, reserving practices, reporting thresholds, and market boundaries. The open actuarial Loss Data Analytics treatment explains why claim counts, positively skewed amounts, deductibles, and policy limits require specialized handling.
This dossier treats statistical standardization as the documented alignment of units, denominators, definitions, clocks, and reporting categories for a stated comparison. Standardization may involve converting raw observations into comparable rates, fixing weights to distinguish risk changes from compositional changes, encoding metadata, or expressing currency and dates consistently. It should preserve economically important heterogeneity. Equal numerical scales alone do not establish comparable risks, and identical column names do not guarantee identical definitions.
The scientific questions are: Which observational units support a stated statistic? How should exposure be constructed and checked? Which distributional summaries convey both typical losses and material tail events? How should visualizations reveal structure without obscuring rare events? When do missingness, outliers, and portfolio composition invalidate a pooled comparison? Which evidence is sufficient to proceed from EDA to formal estimation, pricing, reserving, or risk modeling?
The analytical method combines derivation from established definitions, source criticism, and intentionally synthetic examples whose arithmetic is reproducible from the tables printed here. The monograph does not estimate an empirical insurer's loss distribution, access personal claims records, or assert performance for a specific insurance market. Institutional sources are used to establish the scope of official reporting and professional guidance. Specific applications of the U.S. ASOP No. 23 Data Quality depend on the professional and jurisdictional circumstances defined in that standard; it is an important example of professional guidance, not a universal law for every insurer.
The authorial contribution is the systematic interpretation of how observational identity, exposure, payment mechanisms, portfolio structure, data provenance, and visualization jointly condition the meaning of an insurance statistic. All mathematical models invoked are established statistical methods. The analysis develops interpretive decisions from their assumptions and the documented nature of insurance records.
Statistical Units and Summary Tables
Observational identity and the insurance data-generating process
An insurance dataset can encode at least eight analytically distinct units. A policy record identifies a contractual agreement. A coverage record identifies a line, peril, limit, deductible, or insured object within an agreement. An insured person or asset identifies a risk-bearing subject. A risk interval describes time in force. An occurrence record identifies an event. A claim identifies a request or obligation attached to an event under a contract. A claim transaction identifies a reserve movement or cash payment. A financial statement row identifies a reporting aggregate under an accounting or supervisory perimeter. Each has a legitimate role, and each supports a different denominator and aggregation rule.
A transaction count is not a claim count; a claim count is not automatically an event count; an event count is not automatically a count of independently exposed policies. Consider a hypothetical household claim settled in three payments and reopened once. If a transactional export contains four rows, the analyst must retain a stable claim key and distinguish claim creation, reopening, payment, reserve changes, and recoveries. Deduplication at the wrong level can erase real payments; failure to deduplicate at the appropriate level can inflate reported frequency.
The minimum identity fields for a defensible claims analysis include a source-system identifier, stable policy or coverage key, exposure interval, claim identifier, occurrence or accident date where available, report date, status, settlement/payment transaction date, currency, development or valuation date, claim amount definition, and product or risk classification. The relevance of individual fields depends on whether the question concerns reported claims, ultimate liabilities, payment timing, or insured events. A unique key should be evaluated at its declared grain, and record multiplicity should be checked rather than suppressed as a presumed error.
Claims emerge after selection into the insured pool, occurrence of loss, awareness, reporting, coverage adjudication, possible investigation, estimation, payment, and closure. Each stage changes observable data. A policyholder may sustain an uninsured loss, a covered loss below a deductible, an insured event with several separate claims, or a claim whose reported amount changes over many valuation dates. Consequently, raw claim files are samples of administrative processes as well as samples of risk events. In the actuarial treatment of severity, loss suffered by the insured and payment by the insurer are explicitly differentiated.
Statistical population and observation window
A study population is an operational specification, not a label such as “motor claims.” A complete description identifies geographical territory, insured product and underwriting status, direct or assumed business, primary or excess layer, gross or net of reinsurance, inception and termination rules, effective dates, accident and report dates, and measurement units. The observation window should differentiate calendar period, accident year, underwriting year, policy year, report year, and development age. These lenses answer different questions.
For example, claims paid in 2026 may concern incidents from several earlier accident years. A “2026 claim average” computed by calendar payment date differs from a 2026 accident-year incurred average at a stated valuation date. The mean of a mature accident year is not directly comparable to that of an immature year when incurred development varies. The work of Mack (1993) on chain-ladder reserve uncertainty concerns reserving uncertainty under an explicit development structure; an EDA table can reveal development patterns, but is not itself an estimate of ultimate reserves.
The observational population also reflects selection and exit. Policies can cancel, renew, transfer, or change limits midyear. Claims can close, reopen, or migrate between systems. Portfolio acquisitions may alter the business perimeter. Changes in record retention or thresholds can create apparent temporal breaks unrelated to changes in underlying risk. An appropriate EDA inventory therefore captures both financial data and the metadata of how those data were collected.
Summary tables as executable definitions
A useful one-page data inventory reports the unit and key, row count, unique entity count, coverage and date span, amount and exposure units, missingness, zero and negative values, hierarchy of classifications, and revision timestamp. A more detailed diagnostic cube can cross-tabulate product, geography, coverage, calendar period, claim development state, and data source. Cells should retain both numerators and denominators. Displaying only rates masks small exposure; displaying only counts masks changing portfolio size.
Each field should have a declared type, domain, and interpretation. Dates include timezone and whether they represent local dates or timestamps. Monetary amounts identify currency, price basis, gross/net status, sign convention, paid or incurred basis, inclusion of allocated expenses and recoveries, and valuation date. Categoricals preserve effective coding versions. Exposure variables identify their physical and time units; “one policy” is not automatically “one policy-year.” For derived variables the inventory should record formulas and source columns.
In the U.S., ASOP No. 23's review and disclosure requirements emphasize suitability for intended use, internal consistency, external information where available, sufficiency, material limitations, and the role of sampling (Actuarial Standards Board, 2016). The standard also distinguishes review from formal audit. These propositions support a disciplined EDA checklist while allowing the analyst to state explicitly which facts have been inspected and which remain dependent on a supplier's representations.
Cross-system reconciliation
A coherent exploratory table reconciles its aggregates to controls established at the corresponding business grain. For a claims file, sum of paid transactions by accounting period can be compared with an authorized finance extract after adjusting for recoveries, expense components, foreign exchange, and cutoffs. A claims count need not reconcile to a financial ledger because the two use different units. Policy counts may reconcile only after identifying duplicate versions, multiple coverages, amendments, and cancellation timing. A difference is explained by a mapping and its residual; silently forcing totals to equality destroys the evidence of differing definitions.
Reconciling earned exposure to policies requires time-in-force aggregation. An annual contract canceled after three months contributes approximately one quarter of a policy-year under a simple day-based convention, subject to leap years, exact inclusivity, overlap, and the agreed exposure basis. A system that records exposure at issuance may overstate earned risk if cancellation adjustments have not yet been posted. A change from monthly billing to annual billing can change transaction counts without changing the physical exposure.
Currency reconciliation requires distinguishing nominal amounts, original transaction currency, reporting currency, and exchange-rate policy. Converting claim amounts at current exchange rates and dividing by historical earned exposures is not an inflation adjustment. A consistent monetary comparison needs a date-specific conversion or index policy and disclosure of whether it aims at accounting comparability, purchasing-power comparability, or cost-of-repair comparability. Geographic aggregation needs the same discipline for coverage scope and insured risk location.
Aggregation invariants and failure tests
Certain identities are useful as exploratory controls: segment exposures sum to total exposure when the partition is mutually exclusive and exhaustive; accident-year claim counts aggregate to the specified parent total when one claim has one accident-year assignment; total paid amount equals the signed sum of component transactions after the selected reversals and recoveries are included. Other identities require assumptions. The product of overall frequency and overall mean severity equals aggregate paid loss divided by exposure only when frequency counts and severity amounts refer to the same included claims, valuation window, and payment basis.
A table should flag structural zeros, unavailable observations, deliberately suppressed cells, out-of-range codes, negative corrections, and unknown exposure. These states must be distinguished because they induce different estimands. Zero paid loss may reflect a covered claim awaiting payment; missing amount may reflect reporting failure; negative transactions may reflect legitimate recovery or reversal. No general-purpose replacement rule resolves their meanings.
Exposure-Standardized Rates
The exposure denominator
Insurance exposure measures risk-bearing opportunity. In a motor portfolio, an earned vehicle-year may describe a year of one covered vehicle; in homeowners, an earned house-year; in workers' compensation, a payroll base; in liability, payroll, sales, or another measurable business activity. The Loss Data Analytics exposure table documents this dependence on line of business. Different bases are meaningful when each tracks the risk process under study. A rate “per policy” across portfolios with sharply different policy durations or multiple insured vehicles has no stable interpretation.
For contract or insured object i, define the earned time fraction over observation interval [a,b) as e_i = days covered within [a,b) divided by the agreed number of days in the exposure year. If insured units vary within the period, count their covered time separately. The total E = sum_i e_i is in earned unit-years. A different product may require a volume base such as payroll in standardized currency units. Exposures based on business size demand particular scrutiny when policy limits, inflation, or economic activity change the measurement scale.
Exposure construction itself is an analytical decision. Specify whether intervals are half-open, whether the first and last dates are covered, how leap days enter the denominator, how cancellations and endorsements alter units, and whether multiple simultaneous coverages create multiple exposure units. A vehicle switching insured ownership midway through a year may remain one vehicle-year at a portfolio level, while generating two policy risk intervals. The choice follows the question.
Frequency, severity, and loss cost
Let N be the count of defined claims associated with the same exposure E, and let X_j be the chosen amount for claim j. The empirical frequency is f = N/E, in claims per earned unit-year. The empirical severity is s = (sum_j X_j)/N, in monetary units per included claim when N > 0. The empirical loss cost or pure premium is p = (sum_j X_j)/E. By arithmetic, p = f s whenever the numerator and claim population coincide. The equality does not imply statistical independence of frequency and severity; it is a ratio identity for observed totals.
The conditional expected-loss relation is more subtle. For aggregate S = sum_{j=1}^N X_j, E[S|E] = E[N|E] E[X|E] under appropriate conditional distributional assumptions about claim sizes and counts, including a suitable common conditional mean of X_j. When severity depends on the claim count, catastrophe clustering, or other latent conditions, the simple product of unconditional means can fail. A formal model must investigate those dependencies rather than importing the empirical identity as a probabilistic axiom.
Consider a synthetic, fully specified one-year portfolio:
| Segment | Earned vehicle-years | Recorded claims | Paid loss, USD | Claims per vehicle-year | USD per claim | USD per vehicle-year |
|---|---|---|---|---|---|---|
| A | 1,000 | 120 | 720,000 | 0.120 | 6,000 | 720.00 |
| B | 800 | 40 | 640,000 | 0.050 | 16,000 | 800.00 |
| Total | 1,800 | 160 | 1,360,000 | 0.08889 | 8,500 | 755.56 |
Segment A has higher observed frequency; segment B has higher severity and higher loss cost. An analysis based solely on count rate would rank the segments differently from one based on paid loss per earned vehicle-year. The total severity of USD 8,500 is a claim-count-weighted average of segment severities; total loss cost is exposure-weighted. Swapping the weights without justification creates a false comparison.
A zero-claim cell requires explicit treatment. Severity is undefined if N = 0; it should not be set to zero and averaged with positive-claim cell severities. A Poisson illustration with zero claims over 100 earned units has an exact one-sided 95% upper bound for the rate of approximately -ln(0.05)/100 = 0.02996 claims per unit, under independent Poisson counting and a fixed exposure. This calculation explains why an observed zero is evidence of a limited sample rather than proof that the underlying risk is zero. Other count processes, overdispersion, and correlated losses require alternative uncertainty treatment.
Ratio of sums and sum of ratios
Suppose segment g has losses L_g and exposure E_g. The aggregate empirical loss cost is (sum_g L_g)/(sum_g E_g), equivalently sum_g w_g(L_g/E_g), where w_g = E_g/(sum_h E_h). A simple arithmetic mean of segment loss costs gives every segment equal weight even if one has 2 earned units and another 200,000. The weighted aggregate answers the portfolio question; equal-weight averaging may answer a different question about a typical segment. Neither is valid without stating the estimand.
For frequencies, the same rule applies. For severity, relevant weights are claim counts when severities use identical inclusions and amounts. For insurance penetration or other market ratios, the appropriate aggregate weighting depends on numerator and denominator definitions. The analyst must distinguish the ratio of aggregate quantities from the mean of country-specific ratios. In the OECD 2026 methodological notes, national business perimeters, calculation methods, and country exceptions are expressly disclosed; replacing them with a universal ratio-of-sums assumption is not warranted.
Direct standardization and composition
Direct standardization fixes a reference composition to compare rates under the same segment weights. Let r_gt denote the claim rate for group g in period t and w_g^* a fixed reference weight with nonnegative entries summing to one. Define standardized rate R_t^* = sum_g w_g^* r_gt. The crude rate R_t = sum_g w_gt r_gt uses contemporaneous weights w_gt. The distinction separates within-group rate changes from changing exposure composition for the defined groups. Standardization is a descriptive reweighting, not evidence that all relevant confounding has been removed.
The following deliberately synthetic two-period example exhibits a reversal:
| Segment | Period 1 claims | Period 1 exposure | Period 1 rate | Period 2 claims | Period 2 exposure | Period 2 rate |
|---|---|---|---|---|---|---|
| Higher-risk | 10 | 100 | 0.100 | 160 | 2,000 | 0.080 |
| Lower-risk | 9 | 900 | 0.010 | 4 | 500 | 0.008 |
| Total | 19 | 1,000 | 0.019 | 164 | 2,500 | 0.0656 |
Each segment's observed rate decreases by 20%, while the crude rate increases from 1.9 to 6.56 claims per 100 exposure-years. The higher-risk segment's exposure share increases from 10% to 80%. Standardizing both periods to period 1 weights (0.1, 0.9) yields 0.019 and 0.0152 claims per exposure-year, respectively. The standardized comparison records improvement; the crude comparison records a changed insured portfolio. This is an example of mixture effects associated with the general aggregation problem illustrated in Simpson's 1951 analysis of interaction. It does not assume Simpson's particular contingency-table construction is identical to exposure-rate decomposition.
For exact additive decomposition across two periods, one may write R_2 - R_1 = sum_g w_g1(r_g2-r_g1) + sum_g (w_g2-w_g1)r_g2. The first term measures change at initial weights; the second captures the residual change in weights evaluated at final rates. Reversing the reference period changes the numerical allocation of interaction. A symmetric alternative splits the interaction between the rate and weight terms. No decomposition uniquely defines an economic cause; it organizes the observed arithmetic and motivates a search for underwriting, geography, coverage, claims reporting, and portfolio acquisition explanations.
Scale alignment, time alignment, and currency
Standardization across time requires distinguishing exposure year, accident date, report delay, policy price changes, insured value changes, and claims cost inflation. An increase in average nominal paid severity may result from changes in the mix of repair types, medical treatments, litigation, high-severity losses, currency rates, limits, and inflation. A single general CPI deflator may be a poor proxy for vehicle repair or rebuilding costs, depending on the analysis. If a deflator is used, its selection, base period, and jurisdiction should be documented; real-valued comparisons are not automatically risk-adjusted comparisons.
Across borders, exposure bases may differ fundamentally: earned vehicle-years, policies, payroll, insured sums, and premium volumes have different economic meanings. The OECD Global Insurance Market Trends report documents market aggregates for 2025 and notes that the reporting scope depends on jurisdiction. EIOPA's Solvency II insurance statistics use particular reporting templates and distinguish solo from group aggregates. Nationally standardized data may be internally coherent while remaining unsuitable for a pooled micro-level claim frequency calculation.
The limits of standardized rates
Standardized rates remain sensitive to specification. Groups may be too broad, exposure quality may be uneven, rare outcomes may be volatile, new products may fall outside a common support, and a nominally fixed category may change definition across periods. Comparisons must therefore display crude and standardized estimates, selected weights, cell exposures, uncertainty, and any excluded cells. A comparison based on a fixed set of groups should clearly indicate how emerging products, missing categories, or discontinued contracts were handled. Standardization can make a descriptive comparison legible only where there is enough common meaning in the source variables to support it.
Loss Quantiles and Concentration
Why the mean is an incomplete description
Insurance loss amounts are often nonnegative, strongly asymmetric, and influenced by infrequent large claims. Some portfolios combine many small property repairs with comparatively rare total losses; others include long-tailed liability or catastrophe claims. The same sample mean may describe distinct combinations of typical payments and extreme observations. The actuarial severity literature treats skewness, long tails, and contractual modification as defining properties of the observational problem. Consequently, the mean should usually be reported with count, median, selected quantiles, aggregate amount, and appropriate measures of tail concentration.
For ordered observations x_(1) ≤ ... ≤ x_(n), an empirical quantile at probability u can be defined by the inverse empirical distribution function, Q_n(u) = inf{x : F_n(x) ≥ u}. With this nearest-rank convention, Q_n(u) = x_(ceil(nu)) for 0 < u ≤ 1. Other sample quantile conventions interpolate between order statistics. A reproducible report states the convention, especially when small segments or regulatory thresholds make quantile boundaries consequential. The estimated 99.5th percentile of 30 claims depends substantially on how the tail is handled; displaying a precise decimal does not create data that were absent.
Consider seven synthetic recorded payments in USD: 1,000; 1,500; 2,000; 2,500; 3,000; 10,000; and 60,000. Their total is USD 80,000, the mean is USD 11,428.57, and the median is USD 2,500. The largest payment accounts for 75% of the total and the two largest account for 87.5%. Using the nearest-rank convention, the 90th percentile is USD 60,000; interpolating conventions can yield a materially different estimate. This small example is not offered as an empirical severity distribution. It demonstrates why a table of averages without quantile conventions and concentration displays can be misleading.
The median reflects the center of the observed claim distribution but usually does not preserve the portfolio's monetary total. The mean has a natural additive relation to aggregate losses but is sensitive to very large observations. Both quantities are informative. The correct choice is determined by the question: typical paid claim, total expected payment, operational claims workload, deductible design, or tail capital stress. Using a trimmed average to represent total financial loss discards information that may be central to the economic object.
Distribution functions, thresholds, and tails
For a nonnegative loss variable X, the distribution function F(x) = P(X ≤ x) and survival function S(x) = P(X > x) provide complementary views. A log-scaled display of positive claim amounts can reveal multiplicative spread and tail shape, while the untransformed scale preserves currency interpretations. A zero payment or negative recovery is not suitable for an ordinary logarithm and must be identified before transformation. A histogram with equal-width currency bins may allocate nearly all observations to the first few bins; variable-width bins or log-amount displays may be preferable if their construction is disclosed.
A useful empirical tail diagnostic is the exceedance count n_u = sum_j 1(X_j > u) and, when n_u > 0, the mean excess e_n(u) = (sum_{j:X_j>u}(X_j-u))/n_u. Plots of exceedance frequencies and mean excess by threshold can indicate whether large values behave differently from the body. In small portfolios, these curves are noisy and threshold-dependent. Their primary exploratory purpose is to locate where a formal tail model might be warranted. A high quantile cannot be extrapolated reliably from a few observations without distributional assumptions.
Insurance benefits change what is observed. With an underlying nonnegative loss X, deductible d, and a payment cap L after deductible, a simplified claim payment is Y = min{max(X-d, 0), L}. This maps all losses at or below the deductible to zero paid amounts and all sufficiently large losses to the payment cap. If the dataset contains only positive paid claims, subdeductible losses are absent and the observed positive distribution is conditionally truncated. If the dataset retains all insured events with payment Y = 0, the process includes a censored-and-shifted payment measure. If the only recorded value for a large loss is capped at L, the underlying loss is only known to have exceeded the cap threshold, and the observation carries right-censoring information. The Loss Data Analytics severity chapter develops these differences carefully.
It follows that the “99th percentile of claims” is not a self-contained quantity. The phrase can refer to loss suffered, reported incurred loss, gross adjusted loss, insured payment before reinsurance, payment per transaction, or payment per positive claim. Each has a different distribution. A change in deductible, limit, reinsurance cession, settlement practice, or salvage accounting can change an observed percentile even if the underlying economic loss distribution remains identical.
Truncation, censoring, and statistical selection
Truncation means some observations do not enter the dataset because they do not meet an inclusion condition; censoring means an observation enters but its exact underlying value is partly unknown. In a loss sample observed only above threshold d, ordinary sample averages estimate a conditional mean E[X | X > d], not the unconditional mean E[X]. Under additional assumptions, a likelihood for left-truncated observations includes a normalization by the survival probability S(d). For a right-censored value known only to exceed L, the likelihood contribution is S(L), rather than a density value at L. The actuarial model-selection discussion gives the formal treatment in an insurance setting.
The distinction also applies to event timing. Claim closure data restricted to closed claims oversample rapidly settling matters and may omit large, complex open claims. A claim that remains open at the extract date is right-censored for closure time if the opening date and last follow-up are known. The classical Kaplan–Meier product-limit estimator, introduced by Kaplan and Meier (1958), offers an approach for survival estimation under its censoring assumptions. It is not a universal solution to administrative missingness: censoring related to unobserved severity or claim handling can violate simple noninformative-censoring conditions.
Loss reporting has additional selection. Small claims may go unreported, claims can be denied, incidents can be consolidated, and reserve practices can change before payment is complete. An EDA comparison must identify the selection mechanism before inferring the latent distribution of all insured losses. If that mechanism cannot be estimated, the analysis should name the actual estimand, such as the paid amount distribution among claims reported by a valuation date, instead of asserting the distribution of all economic losses.
Concentration metrics and their denominators
To describe loss concentration, order nonnegative observed payments x_(1) ≤ ... ≤ x_(n) and calculate the share of the total held by the largest k, T_k = (sum_{j=n-k+1}^n x_(j))/(sum_j x_j), where the denominator is strictly positive. A top-1% share or top-10-claim share should specify whether the count refers to claims, occurrences, policies, or adjusted payments. Claim-level aggregation is necessary when one large occurrence generated several split transaction records.
A Lorenz-style cumulative curve can place the proportion of ordered claims on the horizontal axis and their cumulative proportion of aggregate paid loss on the vertical axis. The associated Gini index is sensitive to the chosen claim population and the distinction between zero, negative, and positive amounts. A portfolio containing many reported zero-payment claims can show a different Gini index from a positive-paid-claims-only extract while having the same positive payments. Comparisons require common inclusion rules and should not collapse a tail-driven financial issue into a single normalized concentration number.
The synthetic seven-claim example makes the point starkly. The top claim contributes 75% of paid loss; replacing that legitimate amount with USD 10,000 would reduce the calculated total from USD 80,000 to USD 30,000, a 62.5% reduction. Such winsorization may improve the stability of a diagnostic fit to the central body, but it changes the portfolio total and tail representation. If winsorization is used, report original and altered aggregates, threshold, rationale, and sensitivity; do not treat an altered sample as the unmodified insurance experience.
Geographic and catastrophe dependence
A large portfolio need not provide independent exposure units. Several policies can be affected by one storm, wildfire, earthquake, cyber event, or supply-chain failure. A claim-by-claim display may create an impression of thousands of independent losses when an occurrence-by-occurrence chart reveals a small number of shared events. The statistical unit for a frequency model may be claim count, but the dependence structure of aggregate loss requires event aggregation or another approach to cross-policy correlation.
Explore both individual claim distribution and event total distribution, keeping the occurrence-to-claim link. If the available data cannot reliably identify common causes, describe the limitation and avoid equating transaction dispersion with independent risk diversification. For catastrophe portfolios, geographic heatmaps should be based on location of insured risk and matched exposure, not solely claim count or headquarters addresses. Dense insured regions naturally produce more claims even at identical per-exposure hazard.
An analyst should examine changes in coverage density, valuations, hazard maps, reporting systems, building characteristics, and catastrophe definitions before attributing a geographical rate shift to climatic or physical hazard. Exploratory figures can identify potential clusters but cannot independently identify causal mechanisms.
Quantile uncertainty and comparability
Quantile uncertainty is greatest where data are sparse and the target probability is extreme. Distribution-free intervals for a sample quantile can be constructed from order statistics under independent, identically distributed sampling, but clustered claims, delayed reporting, and dependence require care. Bootstrap intervals can be instructive for the central body; rare catastrophes often need event-level resampling or explicit tail modeling, because naive claim-level resampling destroys cluster dependence.
Comparisons should align inflation and currency treatment, amounts before or after policy terms, loss development age, benefit type, risk class, and inclusion threshold. A lower 95th percentile of net paid claims could reflect more restrictive policy limits or reinsurance retentions rather than a safer underlying insured population. A complete quantile table should report n, payment definition, threshold, calculation convention, valuation date, and uncertainty; very small cells may be suppressed or aggregated for privacy but their scientific absence should be visible.
Data Visualization
Visualization as a series of statistical questions
A good insurance exploratory graphic is constructed from a specified question, analytical unit, and quantity. Cleveland's Visualizing Data (1993) exemplifies the broader statistical literature on graphical exploration; its bibliographic identification here does not assert that its full text was independently reviewed. Claim count by accident month addresses administrative and risk volume; count divided by earned exposure addresses a rate; total paid amount addresses financial magnitude; average severity addresses the conditional size of included claims. These must not share an unqualified vertical axis label. The NIST handbook's graphical EDA approach supports using multiple complementary displays to reveal assumptions and structure rather than relying on a single preferred chart.
An initial visual inventory can include a time-line of exposure, claim occurrence, reporting and payments; an exposure-by-category bar chart; a histogram or empirical cumulative distribution of claim amounts; a count-versus-exposure scatterplot; and a claims development triangle where relevant. Every display should disclose whether units with no claims remain in the denominator. Visual inspection is not a substitute for reconciliation but often exposes a denominator error earlier than a downstream model does.
Frequency and exposure plots
Raw claim counts by segment are useful measures of operating workload. Plotting count against exposure, or a rate with its denominator, helps distinguish large segments from high-risk segments. For a Poisson-like count process, uncertainty tends to be greater for small counts, so rate plots without error information can overemphasize unstable small cells. A useful display combines point estimates with exposure or an interval, and may use an alternate panel for the volume of insured units.
If risk varies seasonally, align earned exposure by month or week and compare like periods. Month length differs, weather exposure can be seasonal, and product renewal cycles can change composition. A peak in claims in January is difficult to interpret without a time-appropriate measure of risk opportunity and the status of reporting lags. In an acquisition or system migration, use event markers and separate pre- and post-change data rather than smoothing a structural discontinuity into an apparent trend.
A plot of observed frequency by tenure, age band, vehicle type, territory, or underwriting class should show cell volumes and missing category shares. These categorical variables may be linked: tenure and product can vary together, territory mix can differ by insurer, and changes in underwriting eligibility can move policyholders into or out of a category. The plotting objective is to identify candidates for stratification and formal adjustment, not to infer causation from a steep line.
Severity displays
Because positive severity is often skewed, a complementary set of panels can show the untransformed monetary distribution, log-scale positive paid amount, empirical distribution or survival function, selected quantiles, and the aggregate amount by severity band. The first describes total money; the second reveals the typical multiplicative range; the tail panel highlights exceedances; and the amount-by-band chart shows which observations dominate cost. A histogram can be misleading when bin limits are not disclosed, particularly with a few extreme claims.
Boxplots are compact comparisons for moderately sized groups. Their conventional whiskers and flagged observations rely on conventions that should be stated. In heavy-tailed portfolios, a whisker beyond 1.5 times the interquartile range may flag many economically legitimate losses. The NIST discussion of scale explains why median absolute deviation and interquartile range react differently to the tail than variance and standard deviation. Display these measures as alternatives with different properties, rather than declaring one universally superior.
A logarithmic display requires special handling of zeros, negative recoveries, and open liabilities. Adding an arbitrary constant to every value shifts small amounts disproportionately and can contaminate cross-group comparisons if the constant differs. Log1p is a computational transformation, not an economically standardized unit for insurance loss; where used it should be labeled clearly and interpreted on its transformed scale.
Development and delay maps
For claims handling and reserving questions, heatmaps or triangles indexed by accident period and development lag distinguish the calendar axis from the age of the claim. Rows may represent accident years, columns development months, and cell entries cumulative paid loss, incremental paid loss, incurred estimates, open counts, or reporting counts. These are different objects and require separate panels or precise legends.
The most recent accident years have less observed development at a fixed valuation date. A heatmap in which blank future cells are colored as zero introduces systematic downward bias. Developmental missingness should be marked as structurally unavailable. A comparison of paid-to-incurred ratios can suggest operational changes, but shifting case-reserving policies, settlement inflation, legal processes, and benefit reforms affect its interpretation. Mack's reserve variability analysis provides a formal example of statistical reserving subsequent to, and conceptually distinct from, this descriptive inspection.
Lag plots should report the distribution of reporting delay (report minus occurrence), settlement delay (closure minus report), and payment delay where defined. Negative lags can signal date errors, legitimate retrospective corrections, or multiple date conventions. Long lag can reflect liability exposure, regulation, organizational changes, or exceptional claims, and should not be attributed automatically to administrative inefficiency.
Geographical comparisons and denominators
Maps should associate claims and exposure at a compatible geographical level. A choropleth of claims per 1,000 earned risks is interpretable only where exposures are available and sufficiently reliable; a map of raw paid amount should be labeled as cost, not risk probability. Urban areas with greater insured asset density often dominate raw claim counts. Risk locations, policyholder residences, and insurer offices are not interchangeable geographical units.
Geographical classification can change when administrative boundaries or insurer territories are revised. Standard geographic code tables and a documented version are needed for longitudinal comparability. A small-area map can inadvertently reveal identifiable claims when one event dominates the cell. Geographic aggregation and disclosure thresholds therefore form part of both statistical validity and privacy governance.
Visual auditing of differences between sources
Paired displays are especially effective when comparing policy administration, claims and finance systems. A cumulative-amount bridge can identify specific components of reconciliation: gross to net, paid to incurred, original to reporting currency, claims-only to claims-plus-expenses. Comparing histograms of records before and after a migration can reveal coding changes. A time-series of missingness by source system can locate an ingestion fault even when the overall missing proportion looks modest.
Where official institutional series are used, the analyst should mark source revision dates and methodological breaks. EIOPA states in its insurance statistics release notes that preliminary figures may be revised and that historic releases change with corrected or resubmitted company information. A smooth graph from the latest dataset is not proof that earlier vintage values were identical. The data vintage is part of the graphic's evidence.
Accessibility and presentation control
A scientific graphic requires axis units, observation period, population, denominator, uncertainty treatment, data source, transformation, and missingness note. Color alone should not distinguish essential categories. Faceting should keep a legible common scale when the purpose is comparison; free scales are appropriate for exploring small within-group details if clearly indicated. Value labels must be consistent with units and numeric precision. Dual-axis charts deserve particular scrutiny because visual alignment between unrelated quantities can suggest unsupported relationships.
When visualizing rare catastrophic losses, a logarithmic scale may make disparate values legible, but a separate absolute-amount display should preserve the magnitude of aggregate monetary exposure. The two panels communicate different aspects of the same data; neither should silently replace the other. A figure should be independently understandable from its title, legend and notes, without requiring readers to reverse-engineer the underlying data pipeline.
Outliers and Business Mix
Unexpected values as questions about origin
An outlier is a value that differs markedly from an assumed reference pattern. In insurance, the observed anomaly may be a transcription error, a legitimate catastrophic payment, a change of product, a recovery reversal, an exceptional court award, or a previously unseen accumulation event. Its scientific status follows from investigation of its provenance and economic meaning. The NIST guidance on outlier detection distinguishes labeling, statistical accommodation, and identification. A large claim should be flagged for inspection; deleting it on the basis of magnitude alone can remove precisely the losses the insurance mechanism is designed to finance.
A disciplined anomaly workflow begins by retaining the raw original value and its lineage. Next, check record grain: did a claim-level total repeat on every transaction row, or does the large value belong to one coherent event? Then validate currency, unit scaling, sign conventions, expense inclusion, duplicated coverage, valuation date and claim status. Compare against source documents where permitted. Finally, classify the observation as confirmed error, legitimate extreme, ambiguous source artifact, or data outside the declared scope. Each classification supports a different action.
A corrected error should be traceable to its source and correction rule. An extreme but legitimate claim may motivate tail analysis, robust exploratory summaries, or explicit stress testing. An ambiguous value may require a sensitivity analysis retaining and excluding it, with both results disclosed. An out-of-scope observation should be excluded by a predeclared population rule, not a retroactive rule chosen to improve a preferred result.
Robust summaries and their proper objects
Common robust descriptive tools include medians, interquartile ranges, quantile-based plots, and median absolute deviation. The median absolute deviation is MAD = median_i |X_i - median_j X_j|. For approximately normal data, it is sometimes scaled by a consistency factor, but neither the raw nor scaled MAD automatically defines an adequate insurance severity scale. A heavy-tailed portfolio can contain several genuine regimes; a single resistant spread statistic can conceal them.
Winsorizing replaces observations beyond selected thresholds by those thresholds; trimming excludes selected extremes; capping at contractual limits represents a separate economic payment process. These operations must be distinguished. Statistical winsorization is an analysis choice; contractual payment limitation is part of what the insurer owes; data correction concerns the accuracy of a recorded amount. Conflating them creates a misleading picture of loss experience. A report should retain the unmodified total, then state the effect of any analytical transformation on mean, quantiles, selected weights, and subsequent model results.
A “99th percentile threshold” estimated from a short tail sample is itself uncertain. In catastrophe and high-limit liability lines, sensitivity to thresholds should be examined across several economically meaningful cutoffs. Where events are correlated, the number of independent clusters is as important as the number of claim rows. If only three catastrophe events are present, thousands of associated payments do not provide thousands of independent tail-event realizations.
Portfolio mixture, underwriting, and exposure shifts
Insurance portfolios are mixtures of exposure classes, claim processes, reporting channels, and contract terms. Aggregating them can create patterns absent from each constituent segment. An insurer that increases writing in high-value properties can see nominal average claim severity rise even if the conditional severity distribution within each insured-value band remains stable. A shift toward higher deductibles can reduce the number of positive paid claims and increase conditional payment severity, even where underlying incident frequency is unchanged.
Underwriting selection can change observed loss experience. Price revisions, distribution channels, renewal rules, risk mitigation, geographic expansion, customer behavior, and reinsurance strategy affect who remains insured. A simple comparison of old and new portfolios may therefore mix within-risk effects, changes in population, and measurement changes. The synthesis in Predictive Modeling Applications in Actuarial Science, edited by Frees, Derrig, and Meyers (2014), situates predictive techniques within insurance applications; EDA supplies essential information about the variables and target definitions to which such techniques may be applied.
Stratification should start with materially relevant dimensions: product, coverage terms, insured amount, geography, customer group where lawful and relevant, accident period, reporting system, channel, and status. Segment definitions must be consistent enough to compare across periods. Very fine cells can reduce hidden heterogeneity while producing unstable estimates. The analyst should show cell volume and use principled aggregation or formal partial pooling when appropriate to a later model, rather than hiding sparse cells behind false precision.
Simpson reversals and direct standardization revisited
The synthetic two-period example in the exposure chapter establishes that every displayed segment rate can decrease while the overall rate increases. Its conclusion is precise: crude portfolio rate changes cannot be equated with changes in within-segment risk without addressing composition. The reference weights can be selected from period 1, period 2, a fixed external standard population, or a pooled portfolio. Different choices answer different questions. Report both the crude rates and standardized rates and publish the weights alongside the result.
The approach has limits. If the set of segments changes substantially, if category membership is measured inconsistently, or if a risk has no exposure in a reference stratum, a standardized rate can require unsupported extrapolation or arbitrary category assignment. A fixed reference mix also need not represent the economically relevant future business. It remains useful as a transparent accounting of composition, rather than an automatic replacement for the observed portfolio total.
The general lesson is similar to the interaction problem treated by Simpson (1951): marginal patterns can depend on grouping. Insurance provides a direct monetary reason to retain both views. The observed portfolio aggregate determines actual financial exposure over the period; the standardized comparison examines the relative movement of defined segment rates. A high-quality report uses both without interpreting either as a causal estimate by default.
Missingness, delayed reporting, and data selection
Missing data include unavailable claims amounts, unknown policy characteristics, absent insured locations, incomplete accident dates, unreported losses, and structurally unobservable future development. They must be separated from zeros, valid “not applicable” values, and deliberately withheld cells. Rubin's Inference and Missing Data (1976) clarifies that when the missingness process can be ignored for certain inferences depends on specific assumptions, including missing at random and appropriate parameter conditions. A high proportion of complete records is insufficient to establish such assumptions.
In insurance, missingness often depends on claim size, product, report delay, adjuster practice, source system, or administrative status. For example, severe liability claims may have incomplete early financial estimates precisely because they take longer to settle. Treating their missing paid amount as zero can produce downward bias; excluding them entirely can overrepresent rapidly settled claims. A more defensible EDA table shows missingness by accident year, development age, claim status, product, and amount proxy where available. It should state whether a value is pending, unknown, invalid, withheld, or outside scope.
Multiple imputation and likelihood approaches may be justified under explicitly assessed missing-data models in downstream analysis. EDA cannot by itself verify a missing-at-random assumption. Sensitivity ranges for missing-not-at-random mechanisms can be more informative than one confidently completed dataset. A report should preserve a record-level flag of originally missing values even after imputation; otherwise later analysts may treat synthetic imputations as original observations.
Reinsurance, coverage limits, and business boundaries
A gross-to-net conversion can alter claim distributions nonlinearly. Under a per-risk retention, the insurer's net amount may equal the gross payment up to a threshold and differ above it; proportional reinsurance scales covered layers, while aggregate protections can depend on portfolio accumulation. The mathematical form and effective dates of the treaty are relevant. An empirical distribution of net paid claims should not be compared uncritically with gross original claim severities.
A business may change limits or deductibles by underwriting cohort. Observed positive-payment frequency can decrease as deductibles rise because a greater share of small underlying losses produces no insurer payment. Conditional severity among positive payments can move in the opposite direction as the sample selection changes. EDA should produce parallel tables by cohort, original loss where observable, deductible, limit, gross payment, and net payment. Where original uninsured losses are unobserved, state that the data identify a payment distribution, not the distribution of all loss events.
International accounting and supervisory reporting introduce additional layers. EIOPA's statistical scope and QRT references are specific to Solvency II reporting. The OECD's country methodological exceptions show that business perimeters and loss-ratio definitions differ among contributors. A cross-country panel requires a definitions table before cross-country trends are interpreted. Apparent volatility can reflect revisions or a change in the statistical population.
Confidentiality and analytical granularity
Insurance records can involve personal information, sensitive incidents, geographical addresses, and health-related claims. Statistical usefulness must be considered alongside the disclosure risk created by rare claims and narrow geographical cells. In its anonymisation and pseudonymisation guidance, the European Data Protection Board distinguishes pseudonymization, which retains possible linkability, from anonymization, which aims to make individuals unlinkable; the 2026 anonymization guidelines were identified as consultation material, not silently elevated here to a final legislative standard.
Removing direct identifiers does not itself guarantee that an extreme claim in a sparsely populated region cannot be associated with an individual or business. Aggregation and suppression may reduce disclosure risks but change the distribution available for study. Statistical metadata should therefore document privacy-driven censorship, coarsening, or suppression; an external report should avoid implying that withheld tail cells contain genuine zeros. Any claim about legal permissibility requires the relevant jurisdiction and legal basis.
Exploratory Checks before Modeling
From description to a defined estimand
EDA reaches a scientifically useful endpoint when a modeling question can be stated without ambiguity about observations, predictors, target, exposure, time, and population. A frequency model requires a definition of claims and exposure; a severity model requires a definition of amount and the claim inclusion process; a pure-premium model requires a consistent monetary target and risk-bearing denominator; a reserving model requires accident, valuation, and development structure. The same dataset can support multiple questions, but each needs its own documented estimand.
For claim frequency, a common candidate is a Poisson generalized linear model with log mean log(E[N_i|x_i]) = log(e_i) + x_i^T beta, where e_i is earned exposure and log(e_i) is an offset whose coefficient is fixed at one. This formulation presumes strictly positive exposure in fitted rows and proportionality of expected count to exposure conditional on predictors. If the intensity varies over time or e_i is poorly constructed, the offset can be misspecified. Count variance should be examined for overdispersion, underdispersion, excessive zeros, clustering, and dependence. The CAS GLM insurance rating monograph, second edition with 2025 revision, provides an applied source for insurance GLM development (Goldburd et al., 2025).
For severity, the analyst must define whether positive amounts are gross or net, paid or incurred, capped or uncapped, and selected above a deductible. For compound loss or aggregate severity models, additional checks include tail behavior, dependence between count and amounts, catastrophe clusters, and data-development maturity. A lognormal, gamma, Pareto, Tweedie, or other model is a hypothesis about a defined quantity; it is never a substitute for investigating the source process.
Diagnostic sequence
A practical premodel examination starts with identity and integration: verify a stable entity grain, versioning and key uniqueness at that grain. Next, verify coverage and time: reconstruct in-force periods, accident and report dates, and time windows. Third, reconcile money: signed amounts, gross/net definitions, expense types, currency, transaction-to-claim and accounting controls. Fourth, analyze distribution: zeros, tails, quantiles, concentration, truncation and development. Fifth, analyze composition: product, geography, policy limits, insured sums, and cohort changes. Sixth, analyze missingness and availability: which variables are known at the intended decision time? Seventh, test temporal stability and leakage risk under out-of-time validation.
The order matters because a data error in policy duration can contaminate every subsequent frequency graph, and a claims-deduplication error can change both empirical mean severity and claim frequency while leaving aggregate paid amount unchanged. A model fitted to such records may exhibit seemingly excellent statistical performance because it reproduces an administrative artifact. The initial check is the reliability of the statistical object.
Model development should then challenge basic exposure proportionality and distributional assumptions, inspect fitted residual and calibration patterns by time and subgroup, examine interactions relevant to business composition, and compare appropriate out-of-sample metrics. These are later modeling tasks, but their diagnostic prerequisites are often visible during EDA. The actuarial predictive modeling text emphasizes prediction from historical associations; temporal and institutional changes limit the durability of such associations.
Data leakage and development bias
A variable may be legitimate for historical descriptive analysis yet unavailable when the insurance decision is made. The final claim payment is a known target after settlement but cannot serve as a predictor for initial underwriting. Claim closure status, later adjuster notes, and postevent geographic classifications may contain information from after the outcome. Predictor availability timestamps should therefore be recorded for each use case. A random train/test split can conceal leakage when claims or policies recur across periods or when development information after the prediction date enters the features.
An out-of-time validation split should respect the decision clock and the reporting maturity of outcomes. If claims take years to develop, a recent validation cohort may have systematically incomplete observed severity. The relevant solution depends on the target: evaluate paid-to-date prediction at a fixed horizon, estimate ultimate outcomes with a justified development method, or restrict to mature cohorts. These alternatives answer different questions. EDA must make the choice visible before a headline prediction accuracy is reported.
Repeated claims, renewals, family policies, location clusters, or shared catastrophe events can create dependence between train and test sets. A grouped or event-level split may be needed where the intended generalization is across distinct entities or catastrophes. A seemingly high score can reflect memorized policyholder or event characteristics rather than useful prediction for a new insured population.
Statistical metadata and reproducible transformations
A documented dataset can be described by a compact contract including entity grain, population, field definitions, data types, unit/currency conventions, coverage and loss bases, calendar and timezone, revision timestamp, suppression codes, transformation rules, provenance, and reference population for standardization. These are requirements for reproducibility rather than a new proprietary standard. The official SDMX 3.1 specifications provide a mature framework for exchanging statistical data and metadata, including structural definitions, dimensions, and associated formats (SDMX, 2025). They do not, without further insurance-specific agreement, determine the financial meaning of a company's claim-severity variable.
For dates and timestamps, ISO 8601-1:2019 specifies representations for information interchange, including Gregorian dates and times. A compliant string format can prevent many parsing ambiguities, but analysts still need to specify policy-effective inclusion rules, local business calendars, time zones, daylight-saving transitions, and event attribution. A uniformly encoded claim date is comparable only when its underlying meaning matches across sources.
Every derived table should be reproducible from a preserved input snapshot and transformation procedure. Record source extract identifier, checksum where permitted, extraction time, analysis code revision, mapping dictionaries, excluded and retained record counts, imputation flags, and the final data vintage. The output should include reconciliation totals and diagnostics that would detect changes on a rerun. Display precision should not exceed the information content of small samples or uncertain valuations.
International standardization without false equivalence
EIOPA's published insurance statistics series include balance sheet, premiums, claims, expenses, and other measures by regulatory template, entity level, and time period. The OECD's 2026 methodological annex documents general and country-specific qualifications, missing and confidential notation, and the possibility of divergence from preliminary publications. These sources establish how institutional metadata makes international comparisons possible and also where comparability is qualified.
It would be a category error to infer that the same term “claims” identifies the same claim-level event process in EIOPA aggregates, OECD country returns, and a commercial insurer's internal extract. The former sources report on defined regulatory or national aggregates, while the latter may consist of transactions, incurred case estimates, or settled events. A valid international descriptive table therefore identifies the observation level and reporting perimeter for each entry and separates market-level indicators from individual-policy rates.
An example of a defensible comparative table includes: country; report year; source and release date; direct/assumed business; life/nonlife class; gross/net basis; written/earned premiums; paid/incurred claims; accounting standard; conversion currency; number of observed entities; missing/suppressed fields; revision status; and definitions link. A multinational dashboard that omits these distinctions may look standardized while juxtaposing unlike measurements.
Uncertainty and adequacy of evidence
For a Poisson frequency rate N/E with fixed E, an approximate standard error is sqrt(N)/E at nontrivial N. This calculation is sensitive to overdispersion and independence assumptions. Tail severity quantiles require larger information content near the tail than a point estimate often suggests. As exposure becomes sparse, confidence intervals may dominate apparent differences between groups. EDA should show the raw denominators and sample sizes alongside effect sizes, avoiding binary claims of improvement where uncertainty is high.
Data quality is purpose-specific. The ASOP No. 23 professional standard requires the relevant actuary to consider whether data are appropriate and sufficient and to disclose material limitations. For research, this motivates a distinct question: is the available dataset adequate for the claim being made? Data can support a descriptive rate for a well-defined historical book yet fail to identify a causal pricing effect or a future catastrophe tail. The proper response is an appropriately bounded conclusion, not a universal grade of “good data.”
The planned use of EDA also changes the risk of bias. A dashboard used to monitor reported claim workload needs timely information, even if immature. A long-horizon risk study may need matured incurred losses and stable definitions. A fairness comparison may require careful assessment of selection, group membership, and legally permitted variables. A cross-border market study may rely on aggregate official statistics, with explicit limits on individual-policy inference. One descriptive data product should not be repurposed across these questions without a new estimand review.
Analytical synthesis and research implications
The previous chapters support several conclusions that arise from combining established statistical logic with the operational construction of insurance records. First, a denominator is a substantive description of risk opportunity, not a cosmetic normalization. Second, contractual terms and reporting mechanics transform the observed severity distribution, so exploratory tail statistics need a payment-basis label. Third, pooled portfolio trends are functions of both within-group experience and exposure composition. Fourth, an outlier is a prompt for source investigation before a treatment decision. Fifth, comparable-looking official indicators may have distinct supervisory and national definitions. Sixth, reproducible metadata is part of the research evidence for every plotted or tabulated number.
These are interpretive findings, not a new named framework. Their practical consequence is that exploratory analysis should be archived with its definitions and transformations, not merely with its graphics. A subsequent modeling program can then explicitly adopt, revise, or reject the observational assumptions underlying the initial summaries.
The unresolved research agenda includes methods for explaining portfolio-mix changes under evolving product definitions; event-aware uncertainty in scarce catastrophe samples; diagnostics for delayed and selective claim reporting; privacy-preserving characterization of extreme losses; cross-jurisdictional comparability maps; and the reproducibility of claims development under revisions to administrative systems. Each requires a separately specified data source and research design. None is asserted here to be settled by the numerical illustrations in this monograph.
Methodological Appendices
Appendix A. Minimum statistical data dictionary
The following fields form an adaptable inventory of research definitions, not a universal reporting mandate.
| Domain | Core fields and definition decisions | Principal diagnostic |
|---|---|---|
| Observation | unit/grain; stable key; parent/child keys; version | multiplicity and join cardinality |
| Contract | policy and coverage IDs; product; deductible; limit; inception/expiry | mismatched coverage and exposure intervals |
| Risk | insured unit; location; hazard class; risk interval | duplicate physical exposure; geography gaps |
| Event | occurrence ID; accident date; peril classification | cluster integrity; delayed occurrence coding |
| Claim | claim ID; report date; status; reopened indicator | report lag; repeated records |
| Financial | payment/incurred/reserve/recovery; currency; gross/net; valuation | signed bridge to ledger; unit and FX errors |
| Exposure | earned unit-time or other basis; cancellation/endorsement treatment | negative/zero exposure; duration mismatch |
| Reporting | accident year; report year; development age; snapshot | unavailable future triangle cells |
| Metadata | schema version; mappings; source; revision; exclusions | consistency over data vintage |
| Privacy | suppression, aggregation, pseudonymization or coarsening flags | reidentification and analytical loss of detail |
Appendix B. Reproduction of synthetic calculations
All quantities below are invented to test definitions and arithmetic, and are not observed insurer or national statistics. No simulation claim is made; these are deterministic worked examples.
- Two segments: A frequency 120/1,000 = 0.12, severity 720,000/120 = USD 6,000, and cost 720,000/1,000 = USD 720; B frequency 40/800 = 0.05, severity 640,000/40 = USD 16,000, and cost 640,000/800 = USD 800. Aggregate frequency 160/1,800 = 0.088888..., severity 1,360,000/160 = USD 8,500, and cost 1,360,000/1,800 = USD 755.555.... Multiplying aggregate frequency by aggregate severity reproduces aggregate cost.
- Two periods, higher-risk and lower-risk groups: period 1 frequencies 10/100 = 0.100 and 9/900 = 0.010. Period 2 frequencies 160/2,000 = 0.080 and 4/500 = 0.008. Crude rates 19/1,000 = 0.019 and 164/2,500 = 0.0656. At fixed period 1 weights 0.1 and 0.9, period 2 standardized rate = 0.1(0.08) + 0.9(0.008) = 0.0152. No observational causal claim follows.
- Seven positive payment values total USD 80,000; median USD 2,500; mean USD 80,000/7 = 11,428.5714; largest payment share = 60,000/80,000 = 0.75; two largest share = 70,000/80,000 = 0.875. Artificially capping the USD 60,000 claim at USD 10,000 yields USD 30,000; reduction in total = (80,000 - 30,000)/80,000 = 0.625.
- For a fixed 100 unit-exposure and zero Poisson claims, P(N=0|lambda) = exp(-100 lambda). Solving exp(-100 lambda_U) = 0.05 gives lambda_U = -ln(0.05)/100 = 0.0299573 for a one-sided 95% confidence upper bound under that model.
Appendix C. Claim-to-source and editorial reproducibility
For substantive propositions, this dossier uses direct links to source-specific official or publisher pages in the corresponding sentence and full bibliographic records below. The accompanying EVIDENCE.json identifies the relevant verified claim and the limitation of the source. Institutional guidance is assigned to its actual jurisdictional or reporting scope; statistical formulas in the text are derived from explicit definitions and checked against deterministic examples. The manuscript uses synthetic arithmetic instead of implying access to an insurer's transactional data.
A practical research archive should retain immutable source manuscript bytes, evidence ledger, editorial decision, checked date and reviewer, source-versus-evidence checksums, and source Git commit. The release record should add a publication identifier and actual date only when public release occurs. The manuscript's target canonical address is reserved editorial metadata and does not represent a claim that a public dossier page is live.
The planned article's evidentiary scope excludes independently reproduced EIOPA or OECD microdata and any actuarial valuation of an identified insurer. Theoretical derivations, source verification, and deterministic arithmetic are within this dossier's scope. Formal external peer review is a separate editorial decision. The internal research record should not misstate the distinction between checks actually performed and independent confirmation from a third party.
Conclusion
Exploratory data analysis and statistical standardization in insurance begin with the construction of meaningful observational objects. The policy, exposure interval, insured event, claim, payment transaction, and supervisory return are distinct kinds of evidence. They generate different counts, monetary totals, and denominators. Clear statistical inference therefore begins with the definition of the observed population, contract terms, risk opportunity, time axis, claim state, monetary basis, and reporting mechanism.
Exposure-standardized statistics provide a precise vocabulary for the relationship between portfolio volume and risk. Frequency is the count of a defined event per earned risk opportunity; severity is an amount conditional on a defined claim population; loss cost is aggregate loss per earned exposure. Their ratio identities are valuable checks when the underlying records align. Fixed-weight standardization shows how aggregate changes can be driven by portfolio composition even while rates decline in each segment. The numerical illustration demonstrates this reversal by transparent arithmetic and shows why crude and standardized rates convey complementary facts.
Loss distributions require simultaneous attention to the center and the tail. Median, quantiles, mean, concentration shares, exceedance curves, and payment-basis definitions describe different properties. Deductibles, limits, claims handling, delayed reporting, and reinsurance can systematically change observed distributions. An anomalous payment becomes a valid analytical finding only after its source meaning has been examined. Robust techniques may control sensitivity, but their results should remain traceable to the unmodified data and to economically important extreme losses.
Statistical comparability also has institutional limits. Official aggregates organized under EIOPA, OECD, and other reporting frameworks can be rigorous while applying different boundaries and definitions. Their methodological notes and revision histories form part of the evidence. Statistical exchange formats and consistent date conventions improve reproducibility, but do not transform different insurance quantities into identical ones. Data lineage, versioning, field dictionaries, and qualified comparisons are essential to durable scientific interpretation.
The original analytical synthesis of this dossier is that the interpretation of every insurance statistic depends jointly on observational identity, exposure, contractual payment mechanics, portfolio structure, time, and provenance. This conclusion organizes existing methodological knowledge and its practical implications for exploratory work; it is offered as a reasoned interpretation, not as a newly instituted theory. A scientifically responsible transition to modeling occurs when these dependencies are explicit, the evidence supports the intended estimand, material uncertainties are recorded, and the transformation from original data to published result can be reproduced.
References
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality (revision effective April 30, 2017). Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/data-quality/
Cleveland, W. S. (1993). Visualizing Data. Hobart Press. https://hero.epa.gov/reference/11240/
European Data Protection Board. (2026). Anonymisation / pseudonymisation (guidance portal; includes July 2026 public-consultation material). EDPB. https://www.edpb.europa.eu/topics/ai-and-technology/anonymisation-pseudonymisation_en
European Insurance and Occupational Pensions Authority. (2026). Insurance Statistics (aggregated Solvency II statistical releases and methodological notes; October 12, 2026 update). EIOPA. https://www.eiopa.europa.eu/tools-and-data/insurance-statistics_en
Frees, E. W., Derrig, R. A., & Meyers, G. (Eds.). (2014). Predictive Modeling Applications in Actuarial Science, Volume 1: Predictive Modeling Techniques. Cambridge University Press. https://www.cambridge.org/core/books/predictive-modeling-applications-in-actuarial-science/82A7E5B9545B1BEDD7F6A7491265113A
Goldburd, M., Khare, A., Tevet, D., & Guller, D. (2025). Generalized Linear Models for Insurance Rating (2nd ed., 2025 revision; CAS Monograph No. 5). Casualty Actuarial Society. https://www.casact.org/monograph/cas-monograph-no-5
Heckert, N. A., & Filliben, J. J. (2003). NIST/SEMATECH e-Handbook of Statistical Methods: Chapter 1, Exploratory Data Analysis. National Institute of Standards and Technology. https://www.nist.gov/publications/nistsematech-e-handbook-statistical-methods-chapter-1-exploratory-data-analysis
International Organization for Standardization. (2019). ISO 8601-1:2019: Date and time—Representations for information interchange—Part 1: Basic rules. ISO. https://www.iso.org/standard/70907.html
Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. https://www.tandfonline.com/doi/abs/10.1080/01621459.1958.10501452
Loss Data Analytics Project. (n.d.). Loss Data Analytics: Chapter 3, Modeling Loss Severity. Open Actuarial Textbooks. https://openacttexts.github.io/Loss-Data-Analytics/ChapSeverity.html
Loss Data Analytics Project. (n.d.). Loss Data Analytics: Chapter 4, Model Selection and Estimation. Open Actuarial Textbooks. https://openacttexts.github.io/Loss-Data-Analytics/ChapModelSelection.html
Loss Data Analytics Project. (n.d.). Loss Data Analytics: Chapter 7, Premium Foundations. Open Actuarial Textbooks. https://openacttexts.github.io/Loss-Data-Analytics/ChapPremiumFoundations.html
Loss Data Analytics Project. (n.d.). Loss Data Analytics, Second Edition: Chapter 2, Introduction to Data Analytics. Open Actuarial Textbooks. https://openacttexts.github.io/LDAVer2/ChapDataAnalytics.html
Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213–225. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/distributionfree-calculation-of-the-standard-error-of-chain-ladder-reserve-estimates/E8D207F9A4DCE30300A76780FE510437
National Institute of Standards and Technology. (n.d.). NIST/SEMATECH e-Handbook of Statistical Methods: Measures of Scale. NIST. https://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm
National Institute of Standards and Technology. (n.d.). NIST/SEMATECH e-Handbook of Statistical Methods: Detection of Outliers. NIST. https://itl.nist.gov/div898/handbook/eda/section3/eda35h.htm
Organisation for Economic Co-operation and Development. (2026). Global Insurance Market Trends 2026. OECD Publishing, November 2, 2026. https://www.oecd.org/en/publications/global-insurance-market-trends-2026_c8be6585-en/full-report.html
Organisation for Economic Co-operation and Development. (2026). Global Insurance Market Trends 2026: Annex B, Methodological Notes. OECD Publishing, November 2, 2026. https://www.oecd.org/en/publications/global-insurance-market-trends-2026_c8be6585-en/full-report/component-11.html
Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592. https://academic.oup.com/biomet/article-abstract/63/3/581/270932
Simpson, E. H. (1951). The interpretation of interaction in contingency tables. Journal of the Royal Statistical Society: Series B (Methodological), 13(2), 238–241. https://academic.oup.com/jrsssb/article/13/2/238/7026675
Statistical Data and Metadata eXchange. (2025). SDMX 3.1 Technical Specifications (issued May 2025). SDMX. https://sdmx.org/standards-2/
Tukey, J. W. (1977). Exploratory Data Analysis. Addison-Wesley. https://books.google.com/books/about/Exploratory_Data_Analysis.html?id=ggWJ0AEACAAJ