RESEARCH DOSSIER
Liability Claim Frequency and Severity by Accident Year
Abstract. Liability Claim Frequency and Severity by Accident Year examines the statistical construction, actuarial interpretation, and legal conditioning of incidence and claim size in casualty insurance cohorts defined by the date of the underlying event. An accident year is analytically valuable because it assigns the origin of an insured loss to an identifiable period; meaningful comparison nevertheless requires coherent exposure denominators, compatible trigger and reporting rules, mature estimates of ultimate claim counts and costs, and adjustments for legal and economic conditions. This monograph develops the foundations of liability frequency and severity measurement, distinguishes occurrence and claims-made contract cohorts, reconstructs claim development across valuation dates, and critically examines average costs, large-loss tails, inflation, litigation, case mix, and the boundaries of international comparison. It uses verified actuarial research, first-party regulatory documentation, original algebraic derivations, and explicitly hypothetical numerical demonstrations rather than claiming access to proprietary insurer claim files. Its central analytical finding is that an observed change in accident-year average severity combines the consequences of claim composition, monetary valuation, settlement and reporting development, legal coverage, and selection. A reliable interpretation therefore makes every denominator, development assumption, valuation date, and measurement basis explicit. The work synthesizes existing knowledge and develops independently reasoned conclusions without establishing a new proprietary theory or presenting illustrative figures as market estimates.
Keywords: liability insurance; casualty claims; accident year; claim frequency; claim severity; exposure units; ultimate claim counts; loss development triangles; occurrence coverage; claims-made coverage; reporting lag; IBNR; open claims; inflation; social inflation; claims mix; large losses; chain ladder; Bornhuetter–Ferguson; GLM; negative binomial; heavy tails; stochastic reserving; calendar-year effects; cross-jurisdiction comparison; model risk.
Introduction and Research Object
Liability insurers promise defined responses to claims arising from legal obligations. Each claim begins with an exposure, an allegedly harmful event, a policy and its applicable terms, one or more reports, and a subsequent investigation, defense, adjustment, judgment, settlement, or closure. The actuarial observer usually sees only a dated, revisable representation of that process. Accident-year analysis assigns an event to the period of its occurrence so that observed claims can be compared with the activity that created the insured risk. The resulting statistics are most useful when the event date, the exposure denominator, and the ultimate count and cost are all defined consistently.
The research object is the conditional distribution of the ultimate number and cost of liability claims arising from exposures during an accident year, together with the way that distribution becomes observable through successive valuation dates. A claim count is a statement about how claim records are grouped. A monetary total is a statement about what components of loss and expense have been included. A severity statistic is meaningful only relative to those choices. The CAS ratemaking foundation by Werner and Modlin supplies the basic premium, loss, expense, and exposure setting; the detailed Friedland reserving text demonstrates how accident-year information develops over consecutive valuations.
Three connected research questions guide the argument. First, how can accident-year frequency be estimated from relevant exposures and claims while accounting for delayed reporting and changing business composition? Second, how can ultimate claim severity be separated from observation maturity, settlement practice, inflation, and concentrations of large losses? Third, under what contractual and institutional conditions can the resulting quantities support comparisons between lines, periods, insurers, and jurisdictions?
This dossier uses the parent Science, Casualty and Liability Insurance, for the institutional definition of the field. It concentrates on accident-year incidence and severity measurement rather than reproducing a general treatise on tort law, liability policy coverage litigation, or comprehensive insurer reserving. It evaluates these adjacent subjects only where they change the count, timing, interpretation, or valuation of a liability claim. The conclusions concern estimands and evidence requirements. They do not constitute actuarial certification for any insurer.
Definitions, History, and Institutional Foundations
An accident year is the calendar year, or another declared annual accounting interval, in which a defined loss-causing event occurs. The relevant event can be an accident, an injury, an occurrence, or another contractual trigger. An underwriting year ordinarily collects policy business written or incepted in a period, while a report year collects claims first reported in a period. A calendar year collects transactions or changes recognized during an accounting period. Because a claim incurred in one accident year can be reported and paid in later years, these dimensions classify different historical facts. Accident-year series are cohort series, and calendar-year financial results are mixtures of contributions from multiple cohorts.
For accident year a, define E_a as eligible earned exposure units, N_a as ultimate claim count under a declared counting rule, and L_a as ultimate undiscounted indemnity and allocated claims expense under a declared cost convention. Frequency f_a=N_a/E_a and mean severity s_a=L_a/N_a, provided N_a is positive. Pure premium, or loss cost per exposure, is p_a=L_a/E_a=f_a s_a. These identities are exact conditional on one coherent exposure population, one claim-count basis, and one monetary loss definition. They provide neither a causal decomposition nor a guarantee that separately projected count and severity models will preserve the observed product unless reconciled at a common ultimate definition.
Underwriting practice may instead quote claim count per 100 or 1,000 exposures, a paid-closed average, a reported-claim average, an occurrence limit loss, or a policy-limit loss. The label “average severity” is incomplete without its basis. Severity per reported claim at 12 months, ultimate indemnity severity, and defense-inclusive ultimate severity can move in different directions. Averages are also sensitive to a small number of very large cases, a defining concern for commercial and professional liability.
The history of modern actuarial loss-development technique contains several foundations directly relevant to these ambiguities. Bornhuetter and Ferguson (1972) formalized combining actual emergence with expected unreported losses. Berquist and Sherman (1977) examined distortion arising from changes in case reserve adequacy and settlement patterns. Mack (1993) quantified estimation uncertainty within distribution-free chain-ladder assumptions, and England and Verrall (1999) developed analytic and bootstrap prediction-error approaches under a generalized-linear-model representation. Their combined significance is methodological: emergence, development, and uncertainty must be studied before an accident-year comparison is interpreted as a change in the underlying risk.
An official reporting context is available through the U.S. National Association of Insurance Commissioners' Schedule P data description, covering losses, expenses, unpaid amounts, reported claims, and multiple insurance lines across successive years. Such aggregated regulatory data allow examination of development and loss costs but usually cannot settle questions about individual event dates, multiple claimants, limits, changes in coding, or unreported risks without compatible microdata. In Europe, the supervisory tradition of non-life development triangles similarly emphasizes gross and net claims, paid amounts, estimated claims provisions, and accident or underwriting year as specified in the reporting framework. These different reports reveal elements of the same claims process, not necessarily identical target populations.
Legal Triggers and the Accident-Year Unit
For an occurrence-based liability contract, the event triggering the insurer's obligations is commonly connected to the relevant occurrence or injury during the coverage period, subject to the policy's definition and applicable law. Under a claims-made arrangement, coverage may instead depend on the date a claim is first made, reported, or both; retroactive dates, extended reporting periods, related-claims provisions, and other conditions can determine inclusion. The NAIC's medical malpractice policy comparison explains the basic difference using an official U.S. regulatory example. The designation “accident year” therefore requires an explicit actuarial event attribution rule even when the insurer's attachment or policy-year reporting is determined by a claims-made trigger.
The word “event” also presents difficulty in latent injury, repeated exposure, professional negligence, environmental harm, and product-liability disputes. Harm may develop over an interval, be diagnosed years afterward, or be attributed to more than one policy. A single event can generate several claimants, defendants, legal proceedings, and insured claim records. An insurer must distinguish a physical loss event, a legal claim, a coverage claim, a claim feature, a claimant, and a system transaction. Each yields a different count.
A scientifically reproducible accident-year panel should retain the source event date, reported date, policy effective and expiration dates, coverage-trigger classification, policy limit and attachment, claim identification and claimant linkage, financial accounting date, and valuation snapshot. If an event period is contested, preserve the original asserted dates, the adjudicated treatment where known, the insurer's reserved allocation, and the actuarial assignment rule. Analytical treatment of uncertain dates requires either a declared allocation convention or a probabilistic sensitivity assessment. Silently choosing the earliest or most recent known date changes cohort composition.
The applicable substantive law, policy wording, reporting duty, defense obligation, damages system, and commencement or amendment dates can affect ultimate severity. An example of a jurisdiction-specific financial parameter is England and Wales' personal injury discount rate order, which changed the prescribed rate for relevant personal-injury damages calculations from January 11, 2025. Its application is territorially and legally limited; the parameter illustrates why a nominal settlement series cannot be interpreted independently of the legal method of quantifying compensation. This example does not prescribe how any other jurisdiction or type of claim must be valued.
Exposure, Claim Incidence, and Denominator Integrity
Exposure is the quantitative measure of the insured activity or unit at risk over the assigned accident period. It can be earned policy years, insured vehicle years, employee payroll, professional procedure counts, occupied bed days, transactions, revenue, units sold, products in circulation, or another measure justified by the hazard process. These proxies have different statistical properties. A policy-year denominator counts contracts rather than harmful interactions; payroll can reflect wage inflation even if worker-hours are constant; revenue can rise with prices while unit production falls. Frequency estimated on these bases will differ even with the same underlying claim events.
An eligible exposure must correspond to the insured population, geographic territory, coverage limits, and cohort definition used in the numerator. Mixing primary and excess layers in the claim count while using only primary exposure produces a spurious apparent frequency. Dividing all reported claims by the current written policy count introduces both development and temporal mismatch. Exposure earned over a calendar year may require policy-level prorating, cancellations, reinstatements, seasonal usage, and differential attachment to the underlying hazard. Denominator adjustments can be as consequential as claim projections.
The relationship between frequency and the exposure process is often interpreted through an expected count model. If N_i is the number of claims for exposure stratum i and E_i is its exposure volume, a Poisson starting point is E[N_i|X_i]=E_i exp(X_i beta), with log E_i as an offset. Repeated incidents, heterogeneity and shared events typically generate excess variation; negative-binomial or hierarchical count models may be more defensible. The CAS generalized linear modeling monograph provides an established foundation for estimating rating relationships and evaluating prediction. Statistical convenience never supersedes the requirement to construct a valid exposure and claim population.
The treatment of zero-claim exposures matters. An event ledger consisting only of claims cannot measure frequency because its denominator excludes exposures with no claim. Equally, a claims database may contain contractual defenses to allegations subsequently found groundless; counting the report as a claim and estimating indemnity separately can be appropriate for defense-inclusive liability, while counting only indemnity-positive claims yields a different incidence series. The investigator must choose a count definition before assessing changes over time. A frequency comparison without this declaration is not replicable.
Exposure comparability across accident years requires studying changes in product mix, distribution, geography, policy limits, deductibles, self-insured retentions, employer or professional activity, policy attachment, retroactive dates, and claims handling. A rising count of small reported claims can reflect easier reporting channels or lower deductibles rather than more harmful incidents. A fall in paid-claim counts can reflect slower settlements rather than lower ultimate incidence. A shift toward high-deductible arrangements may remove small claims from the insurer's recorded series while aggregate physical injury remains stable.
Where no defensible common exposure measure exists, the scientific conclusion is a narrower one: report ultimate claim counts and costs as descriptive series and state that comparable frequency per insured risk is unidentified. Monetary premium can support an empirical loss ratio under carefully specified earned premium and gross/net matching, but premium is itself influenced by rate changes, limits, and mix. It is an imperfect substitute for exposure.
Reporting Lag, Claim Identities, and Data Lineage
Claim emergence is an observation process. A harmful event may be recognized immediately, reported after weeks or years, assigned several file identifiers, settled in parts, reopened, or closed without indemnity. The ultimate count N_a cannot generally be read from reported counts at an early valuation t. Let C_{a,d} be cumulative distinct claims attributed to accident year a and observed at development age d. A reporting completion fraction q_{a,d}=E[C_{a,d}]/E[N_a] summarizes the expected proportion reported by age d for a declared population. If the completion fraction is stable and estimable, an illustrative count development estimator is N_hat_a=C_{a,d}/q_d. Its validity depends on representative development experience and on the assumption that previously unreported claims resemble the historical group in their future reporting timing.
One system's “reported claims” may mean notifications; another may mean valid files after triage. A serious liability accident involving three injured claimants may be one event but three claim records, while an action involving one claimant and multiple insured defendants may generate separate cover records. A reopening can add transactions without increasing unique ultimate claimants. Defense-only files can disappear when a database reports only indemnity claims. Data lineage is therefore a substantive scientific requirement. An authoritative claim-count dataset should preserve immutable identifiers, source-system transitions, duplicate and linked-claim decisions, event and report timestamps, and status histories.
Triangle-based observation also suffers from right truncation: recent accident-year cohorts have not been observed long enough to reveal their late notifications. Historical cohorts can be left truncated by system migrations, missing older records, acquisitions or prior reinsurance treaties. Longer development tails imply greater dependence on projections and lower confidence in a purely observed comparison. Longer reporting and settlement delays may also correlate with claim size, so reported mean severity can be systematically unrepresentative of ultimate mean severity.
A robust investigation therefore reconciles event counts, distinct claim counts, open and closed inventories, paid indemnity, paid defense expense, case outstanding, incurred, and any allocated or unallocated expense pools at every valuation. It separately documents coverage decisions, claim reopenings, new claimants on old events, large-loss tagging, deductibles, salvages, subrogation, and reinsurance recoveries. Inconsistent currency translation can create an apparent change even when underlying local-currency costs are stationary.
Development Triangles and Ultimate Frequency
A development triangle places accident years in rows and successive valuation ages in columns. Its upper portion contains observed history; its lower-right cells correspond to future emergence or payment. For claim counts, the observed cumulative counts C_{a,d} at maturity d are extrapolated using estimated development ratios or an explicit reporting delay distribution. For amounts, separate paid and reported-incurred triangles are generally constructed because settlement speed and case reserving operate differently.
A chain-ladder count development calculation selects age-to-age factors F_d from comparable historical cohorts, projects C_{a,d} to ultimate with the product of future factors, and treats the selection as an estimate subject to uncertainty. In its simplest form, N_hat_a=C_{a,d} product_{k>=d}F_k. The formula is descriptive only until the assumptions about reporting stability, cohort comparability, tail completeness, and source data hold. Mack's chain-ladder standard errors show why point projections require explicit variability assessment, especially for recent accident years.
A Bornhuetter–Ferguson style estimation can combine an independent prior expected claim count N0_a=E_a f0_a with the observed claims and an estimated fraction remaining unreported: N_hat_a=C_{a,d}+N0_a(1-q_d). This expression produces a smoother early-cohort estimate when historical development ratios are unstable, at the cost of greater sensitivity to the expected frequency basis. The same method also exists for monetary losses, and the expected counts and expected losses must be calibrated to the same coverage, development and mix definitions. The original Bornhuetter–Ferguson reserve study remains the historical foundation.
For an annual accident cohort that is only 12 months mature, a 70% reported fraction implies that 140 distinct reported claims would project to 200 ultimate claims. That projection is a constructed illustration, not an observation. If earned exposure were 10,000 comparable units, the projected ultimate frequency would be 2.00%, while the raw reported frequency would be only 1.40%. The 0.60 percentage-point gap is an emergence adjustment. It is invalid to compare the raw 12-month frequency against the ultimate frequency of a five-year-old cohort and describe the difference as an accident-year risk trend.
Selection of reporting factors requires testing their stability by line, attachment, coverage trigger, tort environment, geography, and system. The analyst should inspect development patterns by successive valuations and use repeated historical holdouts where practicable. A large newly reported claim can materially change incurred severity even while the number of claims hardly moves. Conversely, many late small claims can change frequency with limited impact on total losses. The count triangle and monetary triangle thus contain complementary and sometimes conflicting information.
Statistical Frequency Estimation and Sparse Cohorts
Frequency modeling can represent heterogeneity beyond broad triangle categories. With claim count N_{a,i} for accident year a and risk cell i, model N_{a,i}|X_{a,i} as Poisson or negative binomial with mean mu_{a,i}=E_{a,i} exp(alpha_a + X_{a,i} beta + u_i). The exposure offset must be known, the coefficients should correspond to identifiable rating or hazard covariates, and grouped observations must preserve the denominator. A flexible model can incorporate smooth time terms, line effects, territory, limits, professional activity, and cohort interactions. It is inappropriate to interpret an estimated year effect as an exogenous change in physical risk where unmeasured exposure mix and reporting changes remain.
Sparse and heterogeneous liability books invite partial pooling. A hierarchical model can share information across adjacent years or risk strata while permitting persistent differences in exposure, legal environment and development. Pooling reduces sampling volatility but may import bias when the groups have distinct claim mechanisms. The investigator must compare pooled results with independent segment estimates and document why information transfer is credible. An external benchmark is not a substitute for matching the insurer's underwriting selection and definitions.
If claims are clustered around a small number of events, Poisson independence may understate variability. A workplace event can harm multiple employees; one defective product may generate many proceedings; a professional services dispute can span years and several insured parties. Modeling event frequency and claims per event separately is often more transparent than treating all claimant records as independent events. Correlations across accident years can also arise from common inflation, litigation climate, court backlog, or claim administration changes. The statistical model must match the actual unit being counted.
The frequency estimation exercise should report at least the definition of the event, exposure unit, original numerator count, development age, reporting completion model, ultimate estimate, uncertainty interval, and material differences under alternative plausible assumptions. Without such a reporting table, a modeled frequency can look precise while concealing the very choices that determine its interpretation.
Accident-Year Severity Measurement
Claim severity is the expected or realized monetary cost per claim under a defined coverage and accounting convention. A severity measurement needs at least five coordinates: the origin cohort, the unit counted, the loss components, the valuation basis, and the development status. Ultimate severity is estimated ultimate aggregate cost divided by estimated ultimate distinct claim count on the same basis. Case incurred severity, reported incurred severity, average paid-to-date severity, average cost of closed claims, and settled-claim severity answer different questions.
Let L^I_a represent ultimate indemnity, L^D_a ultimate allocated defense and cost-containment expenses, and N_a the ultimate number of covered or tracked liability claims. The indemnity-only mean is L^I_a/N_a; an indemnity-plus-defense mean is (L^I_a+L^D_a)/N_a. Where expense is booked at policy or calendar-year pool level, allocation to a must be explicitly estimated. The contrast matters for professional liability, where investigations and defense can be substantial even when a case produces no indemnity. The NAIC's medical malpractice insurance overview describes defense-and-cost-containment intensity in that market.
An observed mean may be materially biased by closure selection. If smaller claims close sooner, the mean among closed claims initially appears low and rises as large cases settle. If catastrophic cases receive rapid interim payments, early paid costs may instead be dominated by the tail. Neither pattern supplies a stable universal correction factor. A claim-state history should record open, reopened, closed, denied and settled claims, payment timing, changes in incurred estimates, and delays conditional on claim size and legal process. To analyze accident-year severity trends, both numerator and denominator should ultimately refer to all relevant claims, not only those reaching a particular closure milestone.
An economic severity measure may differ from accounting incurred cost. Case reserves are management estimates at a valuation date, not observations of ultimate indemnity; reported incurred can be affected by systematic reserve-strengthening exercises. Currency translation, discounting, deductible erosion, retentions, paid-versus-incurred expense placement, and netting of salvage or subrogation each change the displayed result. These must be represented in the data dictionary and sensitivity analysis. The historical Berquist–Sherman reserve analysis remains a direct basis for examining the sensitivity of paid and incurred trends to changes in settlement speed and case adequacy.
Severity estimation commonly combines paid and reported incurred triangles, individual claim development, and scenario adjustments. The simplest accident-year ultimate estimate is L_hat_a=A_{a,d} multiplied by a selected development factor D_d for cumulative paid or incurred A_{a,d}. The factor is a statistical extrapolation. A separate prior-expected amount can instead be incorporated using a Bornhuetter–Ferguson calculation. Where case-reserve adequacy or inflation regimes change, stable development factors may fail. A coherent account reports the method for counts and amounts separately and reconciles resulting severity before drawing any historical conclusions.
One further distinction concerns the tail. The arithmetic mean is the appropriate decomposition of aggregate loss cost into aggregate count and monetary size, but it is insufficient for the solvency or reinsurance significance of infrequent severe cases. Median severity, trimmed means, quantiles, layer-limited severity, and expected excess losses can clarify how a tail shifts. These statistics describe different functionals of the same distribution and should not be substituted silently for the aggregate mean used in p=f s.
Large Losses, Tail Behavior, and Censoring
Commercial liability distributions often exhibit a mixture of routine claims and much larger bodily injury, product, professional or litigation losses. A sample mean can increase because of a small number of high-limit awards, even where most claims remain stable. The statistical task is to distinguish a movement in the typical claim from changes in the probability or magnitude of tail cases. Separate layers and limits are especially important: the insurer's retained claim severity is truncated or transformed by the attachment point, cap, deductible and reinsurance structure, while ground-up damages may be substantially greater.
A useful severity model can represent ordinary losses with a lognormal or gamma family and exceedances above a declared threshold with a generalized Pareto tail, provided sufficient data support threshold stability and the coverage contract permits interpretation of ground-up loss. This is a modeling option inherited from established statistical practice, not a new scientific doctrine. The investigator must account for the fact that the upper tail is observed under legal limits, claims settlement negotiation, insurer reserving, and potentially sparse large-claim sampling. A threshold fitted to post-limit retained values may conceal the distribution of actual damages.
Claim amounts may be right censored by unresolved litigation, left truncated by reporting thresholds, top coded by policy limits, or interval censored by reserve bands. A dataset of settled indemnities alone combines the severity distribution with the closure mechanism. If an insurer's large claims are slower to close, a naive fitted distribution on closed claims will disproportionately represent small claims in recent accident years. Joint observation and cost modeling, or at minimum stratification by resolution and maturity, is required.
The stochastic reserving literature distinguishes expected outstanding loss from uncertainty in the eventual outcome. The Mack (2000) comparison of reserving models demonstrates that stochastic formulations can reproduce the same chain-ladder point result while embodying different distributions and predictive interpretations. A liability tail calculation should state whether its uncertainty interval addresses process randomness, parameter uncertainty, model selection, latent tail behavior, or all of them. A single fitted scale parameter is not a complete statement of aggregate risk.
For excess-of-loss analysis, define retained payment Y=min(max(X-A,0),U) for a ground-up loss X, attachment A and layer width U. Expected layer cost depends on the whole distribution of X above A. A change in retention or attachment will alter observed count and average cost of claims entering the layer even when physical injury frequency is unchanged. If exposure is reinsured on a claims-made versus occurrence basis, delayed reports may shift layer frequency by underwriting period. The analyst should reconstruct gross, ceded, and net series with each contract and policy limit appropriately dated.
A tail study should test alternative thresholds, severity distributions, inflation trends, open-claim assumptions, and large-event definitions. Independent event and claim-level counts reveal whether concentration is generated by more events, more claimants per event, larger costs per claimant, or a higher probability of a severe award. The conclusions can support stress analysis without pretending that a sparse series identifies a unique extreme-event distribution.
Economic and Social Inflation Across Settlement Years
Accident-year origin and monetary settlement year are different dimensions. Liability claims originating in year a may be adjudicated and paid in years a+1 through a+k. A rise in the price of medical care, wages, legal services or future care may therefore affect payments associated with several historical accident years at once. Economic inflation is often tracked through relevant cost indices, while social inflation refers in industry research to changes in liability claim severity beyond specified economic drivers. Both are empirical constructs whose measurement depends on the chosen loss and exposure series.
The UK Prudential Regulation Authority's 2022 reserving and inflation review expressly cautioned that a general consumer price index is not a direct measure of the insurer's claim settlement inflation. It identified variation by line, region and exchange rates and the significance of lag before inflation materializes. This observation has an immediate consequence: mechanically deflating all historical accident-year indemnities with one headline CPI series can obscure the relevant mixture of medical costs, judicial damages, defense expenses, and jurisdiction-specific settlement processes.
A portfolio-wide inflation effect can enter through several channels: changes in the price of rehabilitation and future care; wage-related income replacement; medical treatment and life expectancy assumptions; legal fees; greater time to resolution; changes in litigation financing or bargaining; and new or altered damages law. The premium or reserve actuary can incorporate selected indices and hypotheses but must disclose what empirical part of severity growth remains unexplained. A residual is not itself proof of the causal operation of any one litigation mechanism.
Swiss Re Institute's sigma 4/2024 social-inflation analysis estimates that U.S. social inflation averaged approximately 5.4% annually in 2017–2022 while comparison economic inflation averaged about 3.7%. The report states its index is modeled after adjusting claims growth for economic activity and frequency and then comparing inferred severity growth with economic cost drivers. It also reports different circumstances across selected non-U.S. jurisdictions. These are attributed model-based findings and not directly observed universal annual accident-year severity changes. The method and its aggregate data do not identify a causal effect for an individual insurance book.
Separation of accident-year and calendar-year effects encounters an age–period–cohort identification problem. With development age d and accident period a, the observation period c is determined by c=a+d under a consistent annual indexing convention. In a log-additive model for incremental loss, log E[Y_{a,d}] might be alpha_a + beta_d + gamma_{a+d}, but unconstrained linear trends can be reallocated among these effects without changing fitted values. Assumptions, external information, constraints, or structural interventions are needed to make attribution meaningful. A recent peer-reviewed contribution by Riegel (2026) directly studies calendar-year effects and the incorporation of external prior information into a chain-ladder setting. The analysis underscores that a credible calendar-year component must rest on explicit additional evidence.
A practical inflation analysis can compare the same accident year at successive valuations, inspect payment and incurred changes by calendar-year diagonal, review claim-level economic drivers, and normalize to a declared monetary cost basis. A financial year with unusually high court settlements may reflect delayed disposal of old cases rather than a sharp increase in newly arising accidents. Conversely, a recent accident year can have materially understated ultimate costs if a new future-care or inflation regime has not yet emerged in its reported incurred triangle. The temporal interpretation must follow the evidence, not the column heading.
Where monetary amounts are compared internationally, local-currency severity, purchasing-power transformation, accounting currency translation, and claims-cost indexation should be reported separately. A change in exchange rates is a translation effect, not an automatic increase in underlying damages. The discount rate applied in an individual jurisdiction to future compensation may also change the lump sum payable without altering the underlying injury. The England and Wales discount rate decision supplies one specific example of legal valuation sensitivity, effective in January 2025 for that jurisdiction.
Case Mix and Portfolio Comparability
An unadjusted accident-year average severity is a mixture of claims of different types. Let g index risk strata or claim categories, w_{a,g}=N_{a,g}/N_a be the accident-year proportion in category g, and s_{a,g} its mean ultimate severity. Then s_a=sum_g w_{a,g}s_{a,g}. This identity separates the question of changing within-category severity from the question of changing claim composition. A change in the aggregate mean can be generated by either part and by their interaction.
A standardization can use fixed reference weights w*_g to compare s^std_a=sum_g w*g s{a,g}. Choosing historical, pooled, or policy-relevant weights changes the target population but makes the choice explicit. The difference between crude and standardized means should be interpreted as a descriptive composition effect under the selected partition, not as a causal estimate unless the strata are adequate to control relevant confounding. In liability books, useful categories may include coverage line, claimant injury type, jurisdiction, policy limit, retention, insured industry, reporting channel, claim age, and large-loss threshold.
There is a limit to statistical standardization where the legal objects themselves differ. Claims arising from a workplace injury regime, a medical malpractice policy, an automobile third-party injury policy, and a mass tort can have distinct plaintiff rights, burden-of-proof rules, damages caps, defense obligations, reporting periods, and policy limits. A common numerical average is possible after translation to a declared monetary basis, but a common causal interpretation is not established by currency conversion or inclusion in the broad label “liability.”
Market changes can confound the trend further. An insurer may move toward larger accounts, higher deductibles, specialist underwriting, litigation-prone occupations, or high-limit layers. A favorable period may reflect selection rather than an improvement in underlying industry safety. Claim coding can change after a system migration. A reserve review may strengthen open cases across many accident years simultaneously. A shift in the use of mediation may change time to payment without changing compensable damages. Each alternative has testable implications in event data, reported counts, open inventories, transactions and policy metadata.
The historical Berquist–Sherman adjustments are particularly relevant because a change in settlement rates or case reserves can mimic an apparent loss trend. The Friedland reserving exposition likewise makes repeated valuations and development ages fundamental to interpreting accident-year triangles. A scientifically adequate comparison should examine the original raw series, maturity-adjusted series, standardized mix series, and alternative method sensitivity side by side.
Joint Frequency–Severity Analysis
Under a collective-risk representation, aggregate loss S_a=sum_{j=1}^{N_a}X_{a,j}, where N_a is ultimate claim count and X_{a,j} is the ultimate cost for claim j. When conditional claim severity has a stable mean m_a and the dependence structure allows the conditional expectation of total loss given N_a to equal N_a m_a, E[S_a]=E[N_a]m_a. The relationship is an expectation identity subject to the conditioning assumptions. It does not assert that counts and costs are independent or that a model calibrated on separate historical datasets will forecast the same exposure population.
Claims count and severity are often dependent through event size, type of incident, litigation, and reporting. A catastrophe generating many claimants may also affect the severity distribution; a legal reform may increase both case filings and average settlements; a high limit may alter whether a claim enters the insurer's dataset. The common actuarial practice of estimating frequency and severity separately is convenient, but the aggregate distribution and prediction interval must acknowledge dependence where material.
A joint model may combine a count component for events, an event-to-claimant component, and severity components conditional on accident year, claim type, region, limit and event characteristics. For small claim groups, a hierarchical model can control overfitting; for large-loss layers, scenario analyses can complement sparse statistical tail data. The basic GLM techniques in Goldburd and coauthors (2025) support structured estimation, while the England–Verrall prediction-error work supports the distinction between point estimates and distributions of future claims outcomes.
For a fixed accident-year population, a growth-accounting identity is p_b/p_a=(f_b/f_a)(s_b/s_a), provided frequencies and severities refer to the same ultimate and coverage basis. With f_b/f_a=1.04 and s_b/s_a=1.20, loss cost per exposure rises by 24.8%, not 24%, because the factors multiply. The log decomposition log(p_b/p_a)=log(f_b/f_a)+log(s_b/s_a) is exact for positive quantities. It is useful for reporting movements but remains an accounting relation rather than causal attribution. Changes in definitions, claim development and composition can sit inside each factor.
Uncertainty can be simulated by generating counts and severities with the estimated dependence and the reporting process, projecting open and unreported claims, and applying calendar-year inflation and coverage conditions. The output should include tail measures, prediction distributions and alternative scenarios. A model that correctly matches the aggregate historical loss ratio can still misstate claim frequency or layer severity; validating the components independently provides stronger diagnostic evidence than relying on aggregate fit.
Claims-Made, Occurrence, and Cross-Jurisdiction Interpretation
Accident-year grouping is a useful statistical choice for understanding the origins of casualty loss, but it does not determine the applicable insurance coverage in isolation. In an occurrence-based portfolio, the covered event may fall within a historical policy period while reports and settlements follow many years later. In a claims-made portfolio, a later first claim or notification may determine whether the relevant policy responds. The NAIC explanation of coverage triggers illustrates this distinction in medical professional liability. It follows that a comparison of occurrence business with claims-made business must reconcile insured exposure and attachment periods before it can interpret accident-year frequency.
A useful data design maintains at least two independent cohort fields: economic incident year and policy coverage or report cohort. A cross-tabulation then allocates observed and expected ultimate claim amounts by event year and coverage-year treatment. This reveals whether changes in reported claims originate in insured incidents, retroactive-date exclusions, reporting obligations, policy attachment, or claim administrative timing. Reporting by event year alone can conceal coverage periods; reporting by policy year alone can conceal physical loss origin. Neither information dimension replaces the other.
Cross-border differences in liability law can change both the probability of a successful claim and the damages paid. Legal definitions of negligence, claimant standing, limitation periods, compulsory insurance, punitive damages, class or collective procedures, litigation funding, legal-cost allocation, and discounting of future financial loss affect the response to otherwise comparable injuries. The corresponding accident-year series are conditional on these institutions. Treating different damages regimes as if they generated samples from one common severity distribution would erase a substantial component of the research object.
Comparability within a jurisdiction may also fail around a legal change. A court decision or statutory amendment can apply prospectively, retrospectively, to specified claims, or to events after an effective date. The dataset must preserve relevant dates and the classification of claims under the operative law. For example, the England and Wales 2024 damages instrument relates to a defined discount-rate regime and jurisdiction; it does not establish a global damages assumption. If its effects are analyzed in an insurer's portfolio, the exposure to affected claims and actual case valuation must be tested rather than inferred from the order's existence.
An international statistical comparison can proceed by constructing like-for-like strata with explicit country, currency, coverage form, reporting threshold, damages category, legal regime and claim development. Even with these controls, unmeasured differences in care systems, reporting propensity, claims handling and litigation can remain. The defensible output is a bounded comparison of defined insurance claim outcomes, not an assertion that the societies exhibit the same underlying accident incidence.
Reinsurance, Expenses, and Accounting Boundaries
Gross ultimate liability claims measure amounts before the insurer's applicable reinsurance recoveries, whereas net ultimate claims depend on treaty terms, attachment points, cessions, recoverability and sometimes counterparty conditions. An insurer with unchanged gross injuries can display lower net frequency if a revised per-occurrence treaty absorbs a different layer of losses, or higher retained severity if the retention increases. Source data should distinguish ground-up claimant loss, direct insurer paid indemnity and defense, gross contract claim amount, ceded estimate, and net retained cash flow. The term “severity” must be paired with its layer and basis.
Defense and cost-containment expenses can be allocated to particular claims or pooled across the organization. An indemnity-severity trend may be flat while defense-inclusive severity rises because of litigation duration or investigative expenditure. Conversely, closing claims by withdrawal or defense success can leave material expense and no indemnity. A complete casualty study should separately reconcile claim counts by outcome, indemnity, allocated loss-adjustment expense, other expenses, and ultimate cash flows. The NAIC medical malpractice discussion is a concrete illustration of how legal defense can be material to the insurer's cost base.
Financial reporting estimates require a different level of classification from raw event data. IFRS Foundation's IFRS 17 insurance contracts standard describes insurance contract groups, fulfillment cash flows and the measurement of incurred obligations; the Foundation's IFRS 17 key terms distinguishes the liability for incurred claims from the liability for remaining coverage. An actuarial accident-year loss-cost model is an input to, and potential diagnostic of, financial measurement; its nominal undiscounted claim costs are not identical to the accounting carrying amount once cash-flow timing, risk adjustment and reporting requirements are considered.
Supervisory data provide an independent reporting lens. EIOPA's official S.19.01 reporting analysis describes the non-life claims-development template as collecting paid claims, claims provisions, and reported-but-not-settled estimates through successive valuation years. EIOPA's Q&A on accumulated incurred claims explicitly addresses the interpretation of cumulative paid amounts and outstanding claims by accident or underwriting year. These reports are institutional reporting structures, not a license to infer individual claim incidence from statutory monetary triangles alone.
When comparing insurer data with supervisory aggregates, the scientific procedure is to align gross versus net, paid versus incurred, discounting, expenses, accident versus underwriting year, development depth, currency, and reporting basis. Aggregate public statistics often omit the denominator exposure needed for an ultimate frequency estimate, so a published claims triangle may support reserve development analysis while leaving frequency unidentified.
Worked Numerical Demonstrations
The following examples are original, stylized calculations. Every count, monetary amount, exposure, completion fraction and claim mix in this section is hypothetical. They demonstrate algebraic implications and sensitivity; they are not observations of any insurer, national market, or legal case.
Exposure-normalized frequency and ultimate severity
Consider two mature accident years in one hypothetically stable occurrence-coverage segment. In Year A, 10,000 earned exposure units generate 125 ultimate claims with total ultimate claim cost of $2,500,000. In Year B, 12,000 comparable earned units generate 156 ultimate claims at a total ultimate cost of $3,744,000. Each count represents distinct ultimate claims, and each cost uses the same indemnity-plus-allocated-expense convention.
| Quantity | Year A | Year B |
|---|---|---|
| Earned exposure units | 10,000 | 12,000 |
| Ultimate claim count | 125 | 156 |
| Ultimate claim cost | $2,500,000 | $3,744,000 |
| Frequency per exposure | 1.250% | 1.300% |
| Ultimate mean severity | $20,000 | $24,000 |
| Ultimate loss cost per exposure | $250 | $312 |
Frequency increases by 4.0%, severity by 20.0%, and loss cost per exposure by 24.8%. Aggregate ultimate cost increases by 49.76%, which also reflects 20% greater earned exposure. A presentation of the 49.76% aggregate growth as claims inflation would misattribute exposure volume. Conversely, reporting only the 20% severity increase would miss the smaller but positive frequency contribution. These figures are verified by the identities f=N/E, s=L/N, p=L/E, and the multiplicative change formula.
The decomposition has an accounting meaning. It cannot determine whether severity rose because of medical prices, a greater proportion of high-limit claims, legal valuation, case reserve practices or random tail experience. Those mechanisms require additional stratified data and claim histories.
Reporting maturity and cohort comparability
Suppose an immature accident-year cohort has 140 distinct reported claims after 12 months, 180 after 24 months, and ultimately 200. A simplified 12-month reporting completion fraction of 0.70 yields 140/0.70=200 ultimate claims. The corresponding 24-month completion fraction of 0.90 yields 180/0.90=200. With 10,000 earned units, both estimates imply ultimate frequency of 2.00%. If historical delay patterns change, both selected completion assumptions can fail simultaneously. Agreement of two projections does not prove that the ultimate count is known.
Consider also an accident-year loss amount of $2.4 million paid at 12 months and $6.0 million ultimate, contrasted with a second cohort at $3.0 million paid at the same maturity but $10.0 million ultimate. The paid-to-date ratio is 1.25 whereas the ultimate loss ratio is 1.6667. Early payment alone does not recover the ultimate trend because payment completion differs: 40% versus 30%. This illustration reveals how apparent inflation can be damped or amplified by settlement speed and large-claim timing. It reproduces conceptually the need for the Berquist–Sherman settlement adjustments.
Composition shift and the meaning of mean severity
Consider two claim classes: routine losses and large losses. Year A contains 100 routine claims at $10,000 each and 10 large claims at $200,000 each, yielding 110 claims and $3.0 million total ultimate loss, a mean of approximately $27,272.73. Year B contains 100 routine claims at $10,500 and 20 large claims at $210,000, yielding 120 claims and $5.25 million total ultimate loss, a mean of $43,750.
The raw mean increases by approximately 60.42%. Within each claim class, severity rises by exactly 5%. Standardizing Year B to Year A's claim-class weights gives $3.15 million across a synthetic 110-claim base, or approximately $28,636.36; the standardized mean increases by 5%. The difference between the raw and standardized trends is explained arithmetically by the doubled count of large claims and the resulting composition change. The exercise is deliberately simple but exposes an essential issue: a portfolio-level average can exhibit dramatic growth when conditional severity is comparatively stable.
Class boundaries can be endogenous. Defining “large” by a nominal fixed dollar threshold will mechanically shift more inflation-adjusted claims into the category. A careful study should predefine thresholds or test stable inflation-adjusted definitions and check whether the class proportions themselves are economically meaningful.
Large-loss layer and policy terms
Let ground-up claim severity be X and an insurer's excess layer attach at $100,000 with $400,000 of capacity. For hypothetical ground-up losses of $50,000, $250,000 and $800,000, retained layer payments under Y=min(max(X-$100,000,0),$400,000) are respectively $0, $150,000 and $400,000. The layer has two positive claims and $550,000 of cost, yielding $275,000 average per claim entering the layer. Ground-up severity across three claims has a very different denominator and distribution.
If the attachment point rises, the number of positive layer claims may fall while the average conditional layer severity rises. Such a change is contractual selection, not necessarily a change in accident incidence. A comparative study must therefore report layer terms, treaty years, and the alternative ground-up series where observable.
Calendar-year effects and nonidentification
Imagine a medical-cost shock occurring in calendar year 2025 and affecting the future care expenses of unresolved cases from accident years 2021 through 2025. The shock can appear on one calendar-year diagonal of several loss development triangles even though the underlying injuries arose in different years. Conversely, a new accident-year class with unusually severe injuries can contribute to several later calendar-year diagonals. Both mechanisms can raise financial results but imply different underwriting interpretations.
In a linear model containing unrestricted accident-year, development-year and calendar-year trends with c=a+d, a simultaneous transformation of coefficients can leave fitted values unchanged. Hence separating all three effects without restrictions or external information is mathematically unidentified. This is why the Riegel (2026) calendar-year study adds external information; an attribution that ignores this dependence would overstate what a triangle alone establishes.
Validation, Uncertainty, and Alternative Explanations
A reliable accident-year dossier starts by auditing the exact population: one consistent mapping from insured exposures to harm events, claim identities, coverages and payments. Tests should identify negative or impossible date intervals, duplicate records, changes in policy and claims systems, unmatched recoveries, missing cancellations, repeated claimants, currency anomalies, reopened claims, and inconsistent incurred calculations. Large claims should be reviewed both as individual events and as part of the aggregate distribution.
A second layer of validation reconstructs historical information as it appeared at each successive valuation, rather than training models on hindsight-corrected later values alone. Claims reported after a valuation date, later-resolved case statuses, and subsequently revised legal codes must not leak into the predictors for that date. Development estimates should be back-tested on prior cohorts using a rolling origin. The key diagnostic is how the model would have performed using only information available then. Mean absolute error alone is insufficient for skewed liability losses; error by cohort, tail outcomes, coverage strata, calibration and interval coverage should also be examined.
The England and Verrall (1999) reserving study illustrates the importance of predictive rather than merely fitted uncertainty. The Mack (1993) variability method supplies a contrasting distribution-free measure under stated assumptions. Neither method by itself establishes the effects of legal reforms, severe inflation regime breaks, unknown latent claims or incompatible portfolios. Those require scenario stress and source-grounded judgment.
A comprehensive uncertainty report should distinguish sampling or process volatility; parameter estimation error; uncertainty about report, settlement and reopening delays; model specification and tail selection; coverage or event-date interpretation; data quality; and legal or macroeconomic scenario changes. Where these cannot be quantified credibly, the correct representation is an explicit bounded scenario or an identified information gap. A narrow numeric prediction interval obtained from a misspecified stable-development model does not encompass structural uncertainty merely because it is mathematically calculated.
Sensitivity exercises should compare count-development, paid development, incurred development, prior-expected approaches, and feasible claim-level estimators. They should include at least alternative inflation indices, larger and smaller late-reporting tails, claims-made exclusions, boundary changes in large-loss definitions, and the effect of major individual claims. If different approaches give materially different implied frequency or severity, the disagreement becomes a substantive finding about the available evidence, not something to conceal in a single selected figure.
International comparisons need additional checks for legal effective dates, policy wording, insurer reporting definitions and currency conventions. The Swiss Re cross-country social-inflation analysis provides comparative institutional estimates but does not eliminate the need to verify local cohort coverage and claims data. Even within one country, state-level legal conditions can produce divergent exposure to large liability judgments. A national average should not be assigned mechanically to every regional or insured subgroup.
An important alternative explanation for any rising accident-year severity series is adverse change in observed portfolio composition. Others include faster recognition of severe open claims, slower closure of complicated cases, a revised case reserving philosophy, higher policy limits, inflation in future care, changing litigation practice, reporting threshold changes and random concentration. The analysis is stronger when it states the predicted observable implication of each explanation and examines whether the claim-level evidence supports that implication. Explanations that cannot be distinguished in the available data remain observationally equivalent.
Integrative Scientific Synthesis
The most defensible accident-year comparison begins with a clear estimand: ultimate claim count per eligible exposure and ultimate cost per defined claim under common contractual, valuation and monetary rules. These identities establish a base on which uncertainty and explanatory questions can be organized. The research then asks how much of the observed movement remains after aligning development maturity and within-stratum composition. That adjusted movement is a description of differences in the target population, not a mechanically identified causal effect.
Historical claims are simultaneously occurrences, notifications, legal disputes, managed files, and accounting estimates. Their dates have different meanings, and their monetary values change as the legal and financial processes progress. A coherent analysis retains each dimension rather than selecting a single date and treating it as sufficient. Accident year organizes the origins of loss; development age organizes incomplete observation; calendar year organizes changes in costs and recorded settlements; coverage or underwriting year organizes contractual attachment. Each coordinate addresses an independent part of the empirical question.
The author's synthesis yields four practical conclusions from the verified literature and actuarial identities. First, credible frequency comparison depends at least as much on exposure integrity and claim identity as on statistical sophistication. Second, ultimate severity cannot be inferred from immature paid or closed-claim means without evidence about development and selection. Third, the interpretation of an aggregate severity trend requires identifying case mix, policy layer and inflation assumptions, with explicit limits on causal claims. Fourth, uncertainty associated with large losses, late emergence, and common calendar-year conditions is structurally related across accident years and should be modeled or scenario-tested jointly.
These conclusions do not introduce a novel actuarial theory or identify a uniquely new empirical mechanism. Their contribution is an integrated explanation of how definitions, legal attachment, actuarial development, composition and monetary measurement jointly determine what an accident-year statistic can mean. The approach connects accepted actuarial methods to their evidentiary and legal prerequisites and furnishes reproducible demonstrations of where misleading interpretations arise.
The evidence base for this work comprises published methodological studies and official institutions rather than proprietary microdata. It supports theoretical identities, established techniques, regulatory classification and transparent stylized calculations. It does not establish a new global causal estimate of liability claims inflation, a universal claim frequency, or the numerical distribution of a particular insurer's unpaid losses. Such empirical work would require a reproducible exposure-linked dataset, claim-level longitudinal development and documented jurisdiction-specific legal coding.
Conclusion
Liability claim frequency and severity by accident year are estimable quantities only after their units, exposures, insurance triggers, claim costs and observation dates have been fixed. Frequency measures ultimate claims per eligible unit of exposure; severity measures ultimate cost per claim on a declared financial and coverage basis. Their product measures aggregate loss cost per exposure. The identities are straightforward, while their valid empirical application requires careful reconstruction of delayed reporting, payment, case reserves, litigation, coverage and portfolio composition.
The actuarial literature supplies powerful methods for development and prediction, but each method carries assumptions about claim emergence, stability and dependence. Institutional reporting structures reveal important claim costs without necessarily providing a comparable exposure denominator. Economic and litigation-related inflation can influence multiple accident years through subsequent calendar periods, while changes in policy limits, mix and legal regimes can produce apparent shifts in both frequency and severity. The resulting scientific task is to distinguish observation, estimation and explanation with full traceability.
An accident-year study reaches its strongest form when the investigator preserves the longitudinal claims ledger, reports gross and net monetary definitions, tests development and tail assumptions, standardizes meaningful strata, discloses legal and geographic scope, and evaluates uncertainty using multiple models and scenarios. The original calculations in this monograph demonstrate why those requirements matter. They establish a rigorous interpretation of existing theory and evidence and identify the additional observations needed for insurer-specific or international empirical inference.
References
Berquist, James R., and Richard E. Sherman. 1977. “Loss Reserve Adequacy Testing: A Comprehensive Systematic Approach.” Proceedings of the Casualty Actuarial Society 64: 123–184. CAS original publication and abstract.
Bornhuetter, Ronald L., and Ronald E. Ferguson. 1972. “The Actuary and IBNR.” Proceedings of the Casualty Actuarial Society 59: 181–195. CAS primary bibliographic record.
England, Peter D., and Richard J. Verrall. 1999. “Analytic and Bootstrap Estimates of Prediction Errors in Claims Reserving.” Insurance: Mathematics and Economics 25 (3): 281–293. CAS abstract and bibliographic record.
European Insurance and Occupational Pensions Authority (EIOPA). 2020. EIOPA-BoS-20/754: Quantitative Reporting Templates: Analysis of S.19.01 Non-life Insurance Claims. Primary reporting analysis.
European Insurance and Occupational Pensions Authority (EIOPA). 2017. “Question 897: Interpretation of Accumulated Claims Incurred, Solvency II S.21.01.” Official regulatory Q&A.
Friedland, Jacqueline Frank. 2009. Estimating Unpaid Claims Using Basic Techniques. Casualty Actuarial Society Exam Study Note, 416 pp. CAS publication record.
Goldburd, Mark, Anand Khare, Dan Tevet, and Dmitriy Guller. 2025. Generalized Linear Models for Insurance Rating. 2nd ed., CAS Monograph No. 5, 2025 revision. Casualty Actuarial Society. CAS monograph page.
IFRS Foundation. 2017, as amended. IFRS 17 Insurance Contracts. Effective for annual reporting periods beginning on or after January 1, 2023. Issuer's official standards page.
IFRS Foundation. n.d. “IFRS 17 Insurance Contracts—Key Terms.” Official implementation glossary.
Mack, Thomas. 1993. “Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates.” ASTIN Bulletin 23 (2): 213–225. CAS primary bibliographic record.
Mack, Thomas. 2000. “A Comparison of Stochastic Models that Reproduce Chainladder Reserve Estimates.” Insurance: Mathematics and Economics 26 (1): 101–107. CAS bibliographic abstract.
National Association of Insurance Commissioners (NAIC). n.d. “Schedule P: Analysis of Losses and Loss Expenses.” Official NAIC Schedule P description.
National Association of Insurance Commissioners (NAIC). 2026. “Medical Malpractice Insurance.” Updated April 1, 2026. Official insurance topic overview.
Prudential Regulation Authority, Bank of England. 2022. Insights from PRA Thematic Review of General Insurance Reserving and Capital Modelling. Letter from Nylesh Shah, October 20, 2022. Original supervisory letter.
Riegel, Ulrich. 2026. “A Chain Ladder Method Using Prior Information on Calendar Year Effects.” ASTIN Bulletin: The Journal of the IAA, First View, published online August 10, 2026. Publisher research article.
Swiss Re Institute. 2024. Social Inflation: Litigation Costs Drive Claims Inflation. Sigma, no. 4/2024, September 7, 2024. Original institute research summary.
United Kingdom, Secretary of State / Lord Chancellor. 2024. The Damages (Personal Injury) (England and Wales) Order 2024. Statutory Instrument 2024 No. 1261, effective January 11, 2025. Official statutory instrument.
Werner, Geoffrey, and Claudine Modlin. 2010. Basic Ratemaking. Casualty Actuarial Society Exam Study Note, 416 pp. CAS study text and record.