Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Claim Valuation Error Bands and Independent Appraisal

Abstract. Individual insurance claim valuation transforms evidence about loss into a monetary assessment under contractual and legal constraints. Estimates may differ because observations are incomplete, prices change, repair scopes are disputed, experts use different methods, or legally relevant quantities have been defined differently. This monograph studies error bands, independent appraisal, statistical agreement, calibration, decision thresholds and adjudicative review at the level of an individual claim. It integrates metrological uncertainty analysis, probabilistic forecasting, paired-method comparison and selected original regulatory and adjudicative materials from international supervision, the United States, the United Kingdom and Australia. It distinguishes estimation error from uncertainty about a target, a prediction interval from an expert disagreement range, and measured loss from contractually payable benefits. Reproducible synthetic calculations demonstrate covariance-sensitive propagation, nonlinear payment limits, agreement analysis, interval scoring and the economics of additional evidence. Specific judicial and ombudsman examples illustrate how the amount, scope of loss, applicable price basis and qualifications of appraisers can change the decision. The principal analytical finding is that independent appraisal improves the evidentiary basis of valuation when it supplies competent and genuinely differentiated information to a correctly specified target; neither agreement nor a narrow interval independently establishes accuracy or lawful settlement. This is a documentary, comparative and methodological synthesis, with no private claim microdata, fitted empirical model, new doctrine or claim of previously undiscovered scientific law.

Keywords: Claims science; insurance claim valuation; measurement error; uncertainty intervals; prediction intervals; independent appraisal; loss adjustment; repair scope; expert disagreement; inter-rater agreement; calibration; proper scoring rules; claims adjudication; conflicts of interest; consumer protection; settlement; value of information.

Introduction

Scientific object

A claim valuation is a dated, evidentially supported estimate of the amount relevant to a specified insurance obligation. In indemnity insurance the amount may concern the cost of restoring property, the value of a destroyed asset, lost income or another covered financial loss. In a benefit contract the amount may instead follow a schedule or formula once eligibility is established. In either setting the identity of the monetary quantity governs the meaning of the result. A valuation of the gross damage, a revised repair quote, a determination of covered loss, an offer, a final award and a payment recorded in the insurer's ledger can all be denominated in dollars while representing different objects.

The scientific problem begins when two assessments of the same claim are difficult to reconcile. It is tempting to perceive their difference as a statistical error around a single amount. However, one expert may be estimating the work needed to restore the property, the other the price an insurer's contracted network would charge for a narrower job. A difference between these numbers cannot be interpreted until their damage scope, quantity units, valuation date, quality assumptions, procurement channel, taxes and contractual basis are compared. Statistical inference requires a common target, and legal analysis requires fidelity to the policy and applicable rules.

The parent discipline, Claims Science, situates the individual file among coverage determination, claim investigation, settlement and financial recognition. This dossier isolates uncertainty in individual valuations and the treatment of expert disagreements. Aggregate reserving uncertainty is a different estimand: a sound distribution for unpaid portfolio losses cannot establish whether one household received a sufficient settlement. The existence of a covered event is also a separate evidentiary problem, although particular decisions about the extent of damage can intersect with causation. The analysis treats payment as an institutional outcome, not automatically as the true amount of loss.

Research questions and method

The study addresses three linked questions. How can valuation intervals represent measurement, model and expert uncertainty without creating false precision? Under what circumstances can independent appraisal materially improve settlement decisions? How can investigators test calibration, agreement and the results of review at case level?

The research design is a targeted documentary and analytical review. Source selection prioritizes original methodological publications, official supervisory instruments, regulatory text, judicial opinions, and published dispute-resolution approaches. Sources were identified using searches concerning uncertainty propagation, paired measurement agreement, probabilistic calibration, insurance appraisal and cash settlement. Retrieved text was examined for the claims expressly attributed to it. Legal instruments are identified by jurisdiction, date or accessible edition and role; selected opinions are historical examples. Original study results are used within their observed domain. Methodological transfer to claims is explicitly an analytical interpretation, rather than a reported empirical finding from insurance portfolios.

The formal examples use constructed units, invented claim records and stipulated probabilities. They are intended to establish logic and make calculations reproducible. The study does not estimate market-wide appraiser bias, the causal effect of independent appraisal, or the prevalence of underpayment. A future empirical study would require controlled appraiser assignments, source evidence, a defensible outcome reference and legally comparable claims. The methodological chapter specifies how such research should distinguish actual observation from selection, adjudication and measurement effects.

Historical and disciplinary context

Insurance appraisal grew as an institutional means of settling disagreements about monetary loss. The historical discussion in State Farm Lloyds v. Johnson (Texas, 2009) traces early Texas appraisal jurisprudence and explains the long-standing separation of the valuation function from liability adjudication. Modern digital estimators introduce a more granular architecture of unit costs, repair operations and reference markets, yet the basic institutional question remains the same: who is empowered to assess what quantity, using what evidence, subject to what review?

The statistical tradition has a related but different history. Bland and Altman's 1986 work on method comparison investigates agreement even when a true measurement is unavailable. Metrological guidance distinguishes known corrections, residual uncertainty and the reporting of measurement results. Probabilistic forecast evaluation subsequently formalizes the simultaneous importance of calibration and concentration. These traditions supply tools that can be adapted to insurance only after its contractual and factual objects have been specified. They are not legal authorities for determining the amount owed.

Valuation Error Sources

Defining the target before measurement

Let V denote a gross loss quantity defined by damage scope, valuation basis, date, geography and assumptions. Let E denote the evidence available at a specified decision time, C the operative contract terms and applicable law, and A the appraisal method. An estimate v = A(E) is meaningful relative to V. A payment amount may be represented as a function g(V,C), provided the contract and legal facts determine that relationship. Estimates must preserve these distinctions even where one claims platform stores the results in the same monetary field.

A physical restoration target requires a technically sufficient work specification. A market-value target requires a relevant asset condition and market definition. Replacement cost may depend on functional equivalence, availability and whether policy conditions require actual replacement. Actual cash value may involve depreciation or another specified basis. Business interruption measurement requires a defined counterfactual revenue or profit trajectory and an applicable recovery period. A fixed sum payable under a specified insured event calls for assessment of eligibility and schedule inputs, not estimation of an open-market asset price.

Researchers should record the material scope of each target as a structured statement: event, property or benefit, unit of account, covered damage assumptions, measurement time, monetary currency, price date, contractual basis, exclusions from the estimate and intended use. In a repair claim, for example, a photograph can establish broken tiles without determining the area required for an enduring repair. A quote without a usable scope does not establish that work can be procured for the indicated total. A decision may therefore be well calculated from an inadequate specification.

Measurement, sampling and price error

Measurement error arises when dimensions, quantities, condition classifications or other factual observations disagree with a contemporaneous target. Some errors are evident from contradictions in the claim file: a roof section measured in square feet may be entered as square meters, labor quantities may be duplicated, or a pre-existing damage photograph may be assigned to the insured event. Correcting an identified mistake changes the estimate; leaving it in place and merely increasing the error band obscures an avoidable error.

Measurement uncertainty describes incomplete knowledge of the quantity after appropriate correction. It can arise from inaccessible damage, non-repeatable observations, degraded evidence, uncertain material specifications or a range of feasible repair methods. The combined uncertainty guidance of Taylor and Kuyatt (1994) distinguishes uncertainty related to correcting a known systematic effect from the error caused by that effect. Applied to claims, a demonstrated unit-conversion error calls for recalculation, while imperfect knowledge of concealed dimensions calls for conditional analysis or further evidence.

A market sample is not automatically representative. Three repair quotations obtained from firms using the same subcontractor may be more dependent than three separate letterheads imply. A vendor's rate may assume volume commitments, unrestricted scheduling, narrow service area or excluded services. Historical invoice data may understate the cost of urgent post-catastrophe work if labor and materials have become scarce. A price book can be current as a publication and stale for the specific geographic market. Currency conversion introduces a separate valuation date if components are purchased across borders.

The relevant population is defined by the valuation target. A hypothetical reasonable procurement price for standardized work differs from the amount a particular claimant can actually procure after an insurer elects cash settlement. Price variation between feasible providers may be real economic heterogeneity rather than instrument error. An estimate should record supplier availability, quote expiry, delivery assumptions, labor standards and applicable taxes. A market range should never be interpreted as a license to choose its lowest element without testing whether that element provides the required service under the relevant terms.

Scope, model and contractual uncertainty

The gross amount can often be written as the sum of quantities times unit prices, together with mobilization, overhead, taxes and other appropriate adjustments. Still, repair scope is commonly a discrete decision: repair a component or replace a larger assembly; dry an area or remove saturated material; use a comparable item or manufacture a custom replacement. Where scope alternatives have different physical prerequisites, treating them as small continuous perturbations of one estimate produces a misleading distribution.

Consider an expert who prices 20 damaged units and another who prices 50 because adjacent components must be disturbed to achieve a sound repair. The difference consists of the additional work, any changes in unit price from scale, disposal or installation costs, and potentially different contractual assumptions about what restoration requires. A percentage band around the lower estimate is an inappropriate representation of that dispute unless the larger scope has actually been incorporated as an alternative. The first task is to establish the engineering question and the evidentiary basis of both proposals. Legal eligibility of associated work is an additional issue.

Model uncertainty concerns the adequacy of the relation between inputs and the target. A linear cost formula may omit minimum contractor charges or threshold-type mobilization costs. A depreciation model based entirely on age can fail to account for condition, maintenance, technological obsolescence or the applicable valuation basis. In commercial interruption, the projected counterfactual business performance may depend on seasonal patterns and events unrelated to the insured loss. Different models may give different values even if they share inputs. Validation should assess this uncertainty separately from observation noise.

Legal uncertainty concerns the determination of an obligation under contract and law. It is not the same thing as a probability distribution of repair prices. If a disputed clause controls which works are included, compute amounts conditional on each identified interpretation and state who has authority to decide that question. Assigning subjective probabilities to legal outcomes may serve litigation-risk or settlement analysis, but those probabilities do not determine an insured's entitlement. Net payment must follow the applicable obligation once resolved. The valuation report should show the conditions of each alternative rather than hide contractual ambiguity inside an unexplained margin.

Time and evidence maturity

At first notice of loss a file may contain a short description and a few photographs. At inspection it may include measurements and preliminary repair assumptions. During dismantling, hidden conditions may become observable. At completion, invoices can establish actual expenditure, although reasonable cost, covered cost and actual expenditure still need separate assessment. Each stage changes the information set. Valuation revision can represent rational updating, not necessarily a prior mistake.

Statistical error should therefore be defined at a fixed information date. Comparing an initial estimate for visible damage with a final paid amount that includes later supplements conflates legitimate learning, valuation error, policy changes, negotiations and operational drift. An appropriate study records the evidence available to each appraiser before review, what subsequently changed and whether the original estimate included an allowance for uncertain items. It separates errors discoverable at the original date from genuinely new facts.

For a claimant, however, the consequences of an incomplete settlement can be immediate. A delayed discovery of concealed damage might be predictable as a risk even though the quantity remains unobservable. Where the claimant bears that risk through a final cash offer, the decision should consider the transfer of uncertainty and any supplemental mechanism. Valuation is therefore both a knowledge problem and an allocation problem.

Interval and Scenario Estimation

The meanings of an error band

Error bands are ranges associated with a stated valuation or decision target. The phrase alone carries no probability interpretation. A deterministic sensitivity interval varies specified inputs within defensible bounds. A scenario range compares selected distinct assumptions. A statistical confidence interval refers to a procedure for estimating a fixed parameter. A prediction interval is designed to contain a new or future observation under a stated stochastic framework. A Bayesian credible interval describes posterior uncertainty conditional on the model, prior and evidence. An expert dispersion interval summarizes the variation of appraisers' estimates and may be wide even when all experts share a systematic bias.

Conflation of these concepts produces false claims. A range based on the lowest and highest of two expert reports has no automatic 95% coverage. A 95% confidence interval for average historical appraisal error gives no direct assurance that a particular future repair amount lies within that range. A high-probability predictive band conditional on a statistical population may exclude a legally required work item because the outcome used to train it omitted that item. The appraisal document needs the interval's construction method, nominal level if any, target, assumptions and limitations.

There are two mathematically distinct problems in comparing claims. One involves unknown inputs to a present value, such as uncertain quantity and price. The other involves predicting a later outcome, such as the eventual technically and contractually normalized repair cost. The latter includes developments after the valuation date. Both can be useful, but their bands answer different questions. A narrow measurement interval should not be presented as a warranty of future expenditure.

First-order propagation and a worked estimate

A local approximation to uncertainty in a smooth function f of a vector of inputs X is Var(f(X)) approximately grad(f)' Cov(X) grad(f). In the special case of a linear function this variance relation is exact when the covariance exists. The approach is consistent with the law of uncertainty propagation described in NIST Technical Note 1297 (Taylor and Kuyatt, 1994). Covariance matters whenever the uncertain inputs move together. A large-scale scope change is not well represented by the linear approximation and should be modeled separately.

Take a constructed property repair with estimated quantity Q = 100 units, unit price P = USD 120 and other repair charges F = USD 8,000. The gross cost function is V = QP + F. Suppose standard uncertainties of the quantity, price and other charges are 5 units, USD 8 per unit and USD 600, respectively. Suppose quantity and price have correlation 0.4 while F is uncorrelated with both. All inputs are hypothetical.

At the estimated quantities, V = USD 20,000. The gradient components are 120, 100 and 1. The first-order variance in squared dollars is 120 squared times 25, plus 100 squared times 64, plus 600 squared, plus twice 120 times 100 times 0.4 times 5 times 8. The independent-input part is 1,360,000; the covariance contribution is 384,000. The total is 1,744,000, giving a standard uncertainty of USD 1320.61. Omitting the covariance would give USD 1166.19, appreciably narrower uncertainty.

If the approximate distribution around the point estimate were normal and the estimated standard uncertainty sufficiently precise, a 90% symmetric band could use multiplier 1.645. The resulting half-width is USD 2172.40, or a gross interval of approximately USD 17,828 to 22,172. This is a modeled illustration, not a claim that any actual repair cost follows that distribution. The expanded uncertainty discussion in NIST TN 1297 explains how the choice of multiplier depends on the distribution and uncertainty in the standard uncertainty itself.

The product term QP is nonlinear. In the presence of dependence between Q and P, the expectation of their product need not equal the product of their expectations. Accordingly, USD 20,000 is the plug-in estimate, not an asserted expected value. The calculation deliberately uses a first-order approximation; a fuller probabilistic analysis would specify a joint distribution, include the nonlinear product, discard impossible quantities and assess the adequacy of the approximation. For a small or highly skewed price sample, a symmetric normal band may be unsuitable.

Contract transformation and discrete scenarios

Suppose the policy admitted all the modeled work, applied a USD 1,000 deductible before a USD 20,000 payment limit, and had no other terms affecting this example. Define G(v) = min(20,000, max(0, v - 1,000)). The gross estimate produces a net value of USD 19,000. Applying the same monotone function to both ends of the gross interval gives a payment-value band of approximately USD 16,828 to USD 20,000. The upper limit binds; the transformed distribution can have positive mass at the cap. The gross band remains informative for loss adequacy even where its upper part cannot increase the stipulated payment.

Real policies may prescribe another ordering, sublimits, separate deductibles, a coinsurance formula, salvage offsets or a replacement condition. Their actual wording must be applied rather than substituting this simplified expression. Nonmonotone or multidimensional payment functions need more than endpoint transformation.

Now suppose engineering evidence leaves two alternatives: a local repair, with the USD 20,000 point estimate, and replacement of a larger assembly with an estimated gross cost of USD 30,000. Those alternatives do not become an unconditional 90% interval merely by reporting 20,000 to 30,000. The two points may concern alternative legitimate scopes rather than the same uncertain price. If both alternatives remain plausible, report the conditional estimates and the evidence required to select the scope. Assign scenario probabilities only when their basis can be defended; then a mixture distribution may be useful for specified predictive or decision-analysis purposes, but the legal obligation is still determined separately.

For a fixed benefit under a schedule, the transformation can instead be a step function. There may be little uncertainty about the scheduled amount once eligibility is established, but major uncertainty about which category applies. In such a case a continuous symmetric band is positively misleading. The same reasoning applies near total loss thresholds and policy limits. A modest change in gross damage can produce no payment change beyond a cap, while a modest eligibility change can produce a substantial discontinuity.

Reporting feasible intervals

The reporting uncertainty guidance of Taylor and Kuyatt (1994) requires explanation of the uncertainty components and probability basis for its measurement applications. A claims report can apply the underlying discipline by providing the point estimate, specified target, interval construction, uncertainty sources, numerical assumptions, time stamp, scope conditions and residual unknowns. The notation should inform a claimant about the decision rather than suggest false scientific authority.

An error band can be expanded or narrowed as evidence changes; retain previous values instead of overwriting them. A wide initial range may be scientifically appropriate when inspection is impossible. A narrow adjusted estimate may be appropriate after a detailed independent survey. Width alone is not an adequacy criterion. Where human valuation judgments form the inputs, a report should distinguish quantitative uncertainty calculations from qualitative warnings and scenario boundaries.

For a risk-bearing decision, asymmetric intervals may be useful because underestimated costs and overestimated costs have different effects on the parties. However, an asymmetric operational threshold cannot change the definition of a covered amount. A quantile chosen to manage repair uncertainty serves a distinct purpose from a contract interpretation. The record should identify who bears the consequence of either side of the band, whether supplementary claims are permitted, and whether the assumptions support an expectation, percentile or bound.

Appraisal Independence and Evidence

Functions and authority

Independent appraisal is an assessment performed by a competent evaluator whose work is sufficiently separated from interests, prior conclusions and decision incentives that could compromise the assigned valuation task. Its institutional form varies. A second adjuster may conduct internal quality control. An external surveyor may provide engineering evidence. A party-selected appraiser may participate in contractual appraisal. An umpire may resolve remaining disagreement under the prescribed process. A complaint reviewer, arbitrator or court may have authority to resolve matters beyond the monetary calculation. Their legal roles and evidentiary functions cannot be assumed interchangeable.

An expert's independence has financial, organizational, cognitive and methodological aspects. Financial separation concerns outcome-linked fees, ownership interests and repeated economic dependence. Organizational separation concerns who appoints, supervises and can remove the expert. Cognitive separation concerns exposure to prior values and instructions that anchor judgment. Methodological separation concerns genuinely different observations, sources, assumptions and analytical pathways. Expertise and independence are complementary: independent work conducted by an unqualified evaluator may introduce additional error.

The international supervisory benchmark in ICP 19.10 claims handling (IAIS, 2024) associates claims fairness with clear processes and conflict avoidance, competence and ongoing training. Guidance addresses impartial dispute resolution and insurer oversight of outsourced claims functions. The scientific implication is that referral to a separate corporate contractor cannot by itself demonstrate that the valuation is independent or adequately reviewed. Its authority, qualifications, evidence and incentives must be considered.

Appointment and engagement conditions

A valuation engagement should identify the asset, date, questions, applicable valuation basis, scope assumptions and permitted evidence. A qualified expert should disclose relevant interests, prior involvement, compensation arrangements and limitations. The assignment should allow relevant contradictory evidence to be considered. Documentation of limitations needs to identify their effect on the conclusion: unavailable access to an area is consequential when the omitted area determines restoration scope, while missing historical photographs may be less consequential if independent evidence establishes condition.

A comparative illustration of financial independence is Parrish v. State Farm (Florida, 2023). The court considered an appraisal clause requiring a disinterested appraiser. The proposed appraiser held an interest in a firm entitled to a contingency fee for its public-adjusting work. The majority rejected qualification under the disputed wording. The legal point depends on that policy language and the specific interest. It does not establish an international prohibition on all party-selected experts, contingency-fee arrangements or paid consultants. It does show why disclosure and substantive qualification must be distinguished.

The same scientific issue arises irrespective of which party retains the evaluator. An expert whose future work depends on producing systematically lower estimates may encounter subtle incentives even without an explicit contingent fee. Conversely, payment of a normal fixed professional fee does not establish bias. Actual testing requires examining the allocation of work, evidence access, documented methodology, deviations from benchmarks and communications. Independence should be assessed symmetrically across parties.

Blinding and information independence

During an initial reassessment, the expert can be provided with common original evidence and a defined assignment while being blinded to previous valuations and settlement offers. This reduces anchoring on the first monetary conclusion. Blinding should not withhold facts necessary for an accurate assessment. Relevant damage reports, measurements and safety concerns may need to be disclosed, and a second-stage reconciliation must permit comparison of assumptions. The research protocol thus distinguishes blind initial assessment from subsequent open reconciliation.

External appraisers can rely on the same pricing database, vendor network and photographs as the original evaluator. Shared inputs create correlated errors. For illustration, suppose two unbiased estimates have equal error variance sigma squared and correlation rho. The variance of their simple mean is sigma squared times (1 + rho)/2. At rho = 0 it is half the variance of either estimate. At rho = 0.8 it is 90% of the original variance. At rho = 1 there is no variance reduction. Common bias remains unchanged by averaging. These algebraic results show why nominal independence is not enough.

A second appraiser's value may lie in information rather than a new final number. Separate inspection can reveal omitted damage; an independent market quotation can test whether the stipulated repair is actionable; a review of labor operations can identify duplication; and a structural engineer can distinguish feasible repair alternatives. Where both experts use the same incomplete file and estimator, the additional report may confirm internal reproducibility without improving target validity. That result should still be recorded rather than exaggerated into independent verification.

Evidence provenance and the audit trail

Evidence must be preserved at the level needed to reconstruct a conclusion. Records should identify photographs and inspection times, locations, measurement procedures, relevant technical qualifications, source quotations and quote validity, assumptions, policy wording and dates, correspondence, original line items and subsequent amendments. A claim-level file should preserve the original monetary values and the version of software or pricing data actually used. Merely saving the final total prevents an independent analyst from identifying the origin of a disagreement.

A reconciliation document should distinguish agreement about the physical facts, agreement about the valuation basis, disputed scope or causal attribution, disagreement about inputs, and differences in contract application. Quantitative evidence should be linked to line-level calculations. Qualitative expert reasoning should be recorded where numerical models cannot settle the matter. If two reports disagree about a causal fact, further specialist evidence may be needed rather than selecting a more favorable monetary result.

The claimant should be able to understand the decisive assumptions and correct demonstrable mistakes. Disclosure arrangements must follow applicable privacy, privilege, safety and procedural requirements, but supporting the decision with a comprehensible statement of the reasons is an important part of reliable review. Where appraisal is outsourced, the record should make clear which decisions remain with the insurer and which arise from other institutions' authority.

The cost of appraisal

Independent review consumes time, money and attention. For a minor loss with readily verifiable invoices, an additional full engineering appraisal may cost more than any plausible correction. For a complex loss with scarce market evidence, safety-critical scope or a large potential shortfall, the expected benefit can be substantial. Cost alone cannot justify suppressing rights, required procedures or fair treatment. The value-of-information problem arises within the permissible decision space.

The economics differ by party. An insurer may pay appraisal fees but a claimant may bear the costs of delay, temporary housing and repeated document requests. A narrowly defined financial savings objective can favor low estimates even when they are wrong. The appropriate institutional assessment considers loss accuracy, actual settlement adequacy, timing, administrative cost and the distribution of burdens. Referral policy should not be calibrated solely against downward payment adjustments.

Inter-Rater Disagreement

Agreement, reliability and accuracy

Agreement measures how closely different evaluators assess a common target. Reliability concerns reproducibility under defined conditions. Accuracy concerns closeness to an independently justified reference or target. A pair of appraisers may agree on every value while both systematically omit taxes or concealed work; alternatively, appraisers may be far apart because one of them accesses significant new information. Each question requires a distinct design.

The original method comparison study by Bland and Altman (1986) demonstrates why ordinary correlation is unsuitable as a direct measure of interchangeability. If appraiser B always adds USD 5,000 to appraiser A's estimate, the correlation can be perfect while the dollar excess remains constant on every claim. Correlation primarily describes association across claims of different sizes. It does not state whether differences are small enough to be institutionally acceptable.

For paired assessments on a common valuation basis, let d_i = A_i - B_i and m_i = (A_i + B_i)/2. Examine the relationship between d_i and m_i, average the differences and estimate their standard deviation. Under approximate normality and stable variance, a descriptive pair of limits of agreement can be formed as mean(d) plus or minus 1.96 times its sample standard deviation. These limits summarize individual paired differences rather than confidence bounds for their mean. An agreement study should assess sampling variability of both estimated quantities. Its acceptable disagreement threshold should be specified independently of the observed distribution.

Synthetic paired example

Six constructed cases have the same scope and valuation dates for both independent estimates. Dollar values are as follows.

Case Appraiser A Appraiser B A minus B
A 10,200 10,000 200
B 14,700 15,000 -300
C 20,100 20,000 100
D 30,500 30,000 500
E 49,900 50,000 -100
F 80,000 80,000 0

The difference total is USD 400. The mean difference is USD 66.67. With denominator n minus one, the sample standard deviation is USD 273.25. The descriptive limits are USD -468.91 and USD 602.24. All six observed differences are inside these limits by construction of the illustrative dataset; such inclusion establishes no external coverage guarantee. This very small sample is not a basis for a population estimate or the validation of a normal model.

A predeclared operational tolerance of USD 250 would flag two cases (differences of minus 300 and plus 500), even though the estimates may be highly correlated across the six cases. This is a different question from whether the mean difference is statistically distinct from zero. For a USD 10,000 claim, a difference of USD 500 is 5%; for a USD 100,000 claim, it is 0.5%. A reasonable analytic review considers absolute differences, proportional discrepancies and any decision threshold affected by that difference. No tolerance is established as a legal standard by this illustration.

When the spread increases with size, dollar differences can be misleading for comparative diagnostics. For positive estimates a log ratio, log(A/B), can describe multiplicative disagreement. It cannot be applied to zero amounts without a separate modeling strategy, and it does not resolve distinct scope definitions. Plotting differences against means, examining signed and absolute relative discrepancies, and reporting both dollar-weighted and claim-weighted effects can reveal different features of variation. The goal is to identify where estimates are interchangeable for a specified use, not to maximize a convenient statistical coefficient.

Multirater and repeated designs

A two-appraiser design assesses a defined pair. An institution seeking conclusions about a broader appraiser population needs multiple evaluators and a design that can separate case effects, appraiser effects and interactions. For example, a crossed design randomly assigns multiple appraisers to the same stratified set of cases, holding the evidence and valuation target constant. Repeated independent assessments by the same appraiser on cases where memory and evidence changes are controlled can estimate within-appraiser variation. A mixed-effects decomposition can represent systematic rater differences, between-case variation and residual interactions when its assumptions and allocation structure are defensible.

A reliability statistic must specify the measurement design and intended use. Absolute agreement and consistency answer different questions. A high reliability ratio can result from very large true differences between claims even when absolute discrepancies remain material. Therefore report signed differences, empirical dispersion, amount-specific tolerances and the frequency of threshold crossings alongside any summary coefficient. Appraisers assigned different sets of cases cannot be compared directly without accounting for case mix and referral criteria.

Variance between appraisers is not automatically uncertainty about the correct loss. If everyone uses a common defective pricing formula, the inter-rater variance can be close to zero. If the variation stems from evidence about different plausible legal scopes, a numerical reliability target is premature. The research design should classify discrepancies before attributing residual variation to appraisers. The institutional aim is to uncover decision-relevant disagreement and improve evidence rather than to force artificial consensus.

Decomposing differences

For line-item estimates, differences can be attributed to portions of a calculation. When both appraisers price identical work categories, one exact algebraic decomposition for a single item is Q_A P_A - Q_B P_B = (Q_A - Q_B)P_B + Q_A(P_A - P_B). This fixes B as the price base and A as the quantity for the second term. Reversing the ordering changes the allocation of the interaction term but not the total. A symmetric decomposition uses average quantity and average price, attributing the interaction evenly. Such attribution is a diagnostic identity, not a causal estimate.

Before decomposing, align item identifiers and units. An omitted line may indicate different scope, duplication elsewhere or a coding mismatch. Suppose A includes a disposal line separately while B embeds it in a unit price. A direct row-by-row difference would double count the discrepancy unless the item scopes are reconciled. Distinct cost components such as contractor margin and sales tax require comparably defined bases. Where totals are constrained by policy sublimits or settlement arrangements, perform the reconciliation on gross component values first and apply contractual rules separately.

This technique connects statistical disagreement to a repairable institutional problem. An institution can discover that most large discrepancies originate in hidden-damage allowances, vendor-market differences, inconsistent depreciation assumptions or an incomplete mapping of item codes. Corrective action should target the actual source. A blanket demand that all appraisers stay within an arbitrary percentage of an initial number can suppress important evidence and increase systematic bias.

Calibration and Decision Rules

Choice of validation target

Calibration asks whether probability statements correspond to frequencies under an explicitly defined validation process. For prediction intervals with nominal coverage 1 minus alpha, empirical coverage is the fraction of independently observed reference outcomes falling within their stated intervals. It is meaningful only when the outcome is the one the prediction purported to describe. A model predicting the final payment can be well calibrated to payments while systematically reproducing an unfavorable settlement practice. A report predicting technically sufficient market cost requires evidence of actual comparable procurement or independent cost normalization.

A reference outcome should identify the covered scope, procurement basis, valuation date, currency, repair quality and adjustments for later information. Final invoices establish actual expenditure and can be audited, but they can omit work the claimant could not afford, reflect upgrades, include unrelated damage or contain inefficient expenditures. A final appraisal award provides an institutional reference; it may not be an independently measured loss. A disputed settlement reflects bargaining and procedural constraints. A reference hierarchy can use independently validated line-item quantities and actual comparable cost where available, and separately mark outcomes whose target is only agreement or adjudicative resolution.

The study design must also account for missing labels. Some claims close without work being performed; others settle privately, are abandoned or remain under dispute. Removing them can create selection bias. Outcomes must be characterized by observation status, and the reasons for missing reference amounts should be recorded. Report the population on which calibration is assessed and avoid presenting results for accessible completed cases as though they describe all claims.

Sharpness, miss severity and interval scoring

Prediction intervals should provide useful distinctions while achieving their intended coverage on the validation population. Gneiting and Raftery's proper interval score (2007, §6.2, equation 43) provides a scoring rule for central prediction intervals. With lower and upper bounds l and u, observed outcome y, and miscoverage level alpha, the negatively oriented score is:

S = (u - l) + (2/alpha) max(l - y, 0) + (2/alpha) max(y - u, 0).

Lower scores are better. The width term penalizes uninformative breadth, and the two remaining terms penalize misses by their distance. Compare forecasts using the same nominal level, target, validation cases and units. A proper statistical score is a method for evaluating forecasts, not a settlement entitlement formula.

Here are four hypothetical observations in thousands of dollars with 90% central prediction bands.

Reference outcome Method A band A score Method B band B score
10 9 to 11 2 7 to 13 6
12 10 to 14 4 9 to 15 6
16 13 to 15 22 12 to 20 8
22 18 to 24 6 17 to 27 10

Method A covers three of four, with average width 3.5 and average score 8.5. Method B covers all four, with average width and score 7.5. In this constructed sample, B has the preferable score despite its wider bands. A single miss of one thousand dollars incurs a penalty of twenty score units at alpha = 0.10. Neither four-case coverage figure establishes 90% population calibration, and the example makes no assertion that B is universally better.

In a real study, record lower misses separately from upper misses. A systematic pattern in which outcomes exceed upper bounds can reveal underestimation, omitted work or shifts in market prices. The reverse can indicate overestimation or a change in the definition of the reference. Dollar-weighted and unweighted summaries can produce different signals. Assess subgroup calibration by peril, coverage basis, repair complexity, region, season, claim size and evidence maturity, while reporting sample sizes and avoiding overinterpretation of small groups.

Finite-sample calibration and distribution shift

A practitioner can fit a prediction model and use a separate calibration dataset to convert its error scores into prediction bands. Standard split conformal prediction uses a finite-sample order statistic to obtain a marginal coverage guarantee under exchangeability. In the tutorial by Angelopoulos and Bates (2021), the calibration rank takes account of the future test observation. For a regression point model, a simple choice is absolute residuals between observed and predicted values. With n calibration residuals, use the k-th ordered residual, where k is the ceiling of (n + 1) times (1 - alpha), provided k does not exceed n. When k exceeds n, a finite distribution-free symmetric bound cannot be produced by that rule without further assumptions; a conservative convention is an unbounded interval.

For example, if n = 99 and target marginal coverage is 90%, k = 90. A band centered on the predicted amount with half-width equal to the 90th smallest absolute residual is one possible method. Residuals could be normalized by a pre-fitted scale to improve adaptation to heterogeneous claim sizes if the scoring procedure is properly specified before calibration. The theoretical guarantee concerns a fresh observation from the exchangeable population, not every conditional risk group or individually disputed claim. Ties and discrete outcomes can make coverage conservative.

Claim populations change after catastrophes, pricing updates, statutory changes and modifications to the handling process. Sequential claims clustered within the same storm or vendor network may violate the assumptions of naive independent splitting. Randomly allocating nearly identical files from one event into both calibration and test sets can exaggerate performance. A study should use time-respecting and event-aware partitions, with separate validation for materially changing conditions. A conformal label cannot repair a misdefined target, biased training labels, a broken evidence process or a shifted legal environment.

Intervals should be recalibrated under documented conditions. Historical predictions, calibration sample versions and model changes must remain recoverable. If a new event differs too strongly from the validation population, annotate the result as outside the supported domain and obtain further evidence. A modeled interval is not a reason to deny a claim merely because the claimant's documented cost lies outside it. Such a case may identify a tail, a scope error or a model failure.

Referral as a value-of-information decision

A referral rule identifies cases in which another appraisal or investigation may change a material decision. Useful triggers include disputed damage scope, a difference large relative to the insured's financial position, unstable price evidence, inconsistent technical measurements, a potential conflict of interest, a policy limit or eligibility threshold, suspected model drift and documented inability to procure work. Some triggers arise from legal or contractual procedural duties and must be addressed independently of an economic calculation.

Within the discretionary space, a simple expected-value comparison can be illustrated. Assume there is a stipulated 0.35 probability that an initial estimate contains a material correctable error. Conditional on such error, assume that preventing it avoids an average social or decision loss of USD 6,000, and that independent appraisal has probability 0.70 of identifying it in time to prevent that loss. Under these simplifying assumptions the expected gross benefit is 0.35 times 0.70 times 6,000, or USD 1,470. If appraisal costs USD 700 and expected additional delay imposes USD 250 in burden, the stylized net benefit is USD 520.

The calculation ignores false positives, additional evidence costs, errors introduced by the review, variation in harm and interactions with payment terms. All probabilities, loss values and costs are invented. It is not a fitted model and does not set a mandatory referral threshold. If a potential correction concerns a legal obligation, a necessary safety assessment or an applicable review right, those constraints take precedence over an economic screening rule. The model is valuable for explaining why additional assessment can be economically worthwhile, not for monetizing away a valid claimant entitlement.

A threshold should be assessed relative to urgency and consequences. A small absolute difference can prevent a low-liquidity claimant from initiating repair; a large numerical difference above an unconditional payment cap may not change that payment but may affect related benefits or future dispute rights. A rule based only on percentage thresholds can miss these cases. The analysis should identify the specific downstream decision before testing whether disagreement is material.

Adjudicative Review and Replicability

Distinct institutional routes

Claims review may occur within an insurer, through a contractual appraisal procedure, through an external complaint mechanism or by litigation or arbitration. Different institutions have different authority. Internal review often reconsiders the investigation and reasoning. A contractual appraisal can determine an amount under policy language. An ombudsman assesses complaints within its jurisdiction and remedial powers. A court determines disputed legal rights under procedural and substantive law. Technical assessments can inform each process, but their conclusions do not automatically travel between institutions without regard to their assigned authority.

The distinction matters most when disputes combine physical scope and policy interpretation. In Johnson (2009), the insurer and insured disagreed about hail damage to a roof and the extent of replacement. The Texas Supreme Court treated the disputed number of damaged shingles as within the amount-of-loss inquiry under the policy in issue while distinguishing liability and policy interpretation. It addressed the causation boundary in context, rather than making every causal question exclusively monetary. This opinion illustrates why appraisers may need to address component-level damage attribution without assuming authority over all legal coverage disputes.

A valuation report can support separate decisions by explicitly describing the amounts under alternative assumptions: damage attributed to an event, other observed damage, costs for each repair method, gross loss, amounts affected by exclusions and the contractual transformations applied by the authorized decision-maker. A legally effective award may require a particular format in a particular jurisdiction. Scientific documentation is compatible with such requirements; it does not define them. The historical cases here are selected examples, not a current exhaustive fifty-state or international legal survey.

United States model and state examples

The Unfair Claims Settlement Practices Act (NAIC Model 900, 1997 compilation) describes standards for investigation and disposition and identifies failures of reasonable investigation and fair settlement among the model's enumerated practices. Section 1 expressly addresses the absence of a private cause of action under the model. Its binding effect depends on state enactment, amendments, applicable policy and judicial interpretation. It cannot be cited as a directly enforceable nationwide statute.

An actual regulatory example is California's automobile claims rules, title 10, section 2695.8(f). In the specified partial-loss context, a written estimate used for settlement must be made available to the claimant and sufficient for repairs meeting the stated trade standards. The provision also identifies alternatives where a claimant submits a higher estimate for necessary work, including payment of the difference, an applicable repair-shop arrangement or a reasonable adjustment of the claimant's estimate. The rule is materially more specific than a statistical comparison of two amounts; the feasibility and standard of the work are relevant. The text was examined as a published regulation reproduction, not as a comprehensive determination of the law governing a particular claim.

The Florida and Texas decisions concerning contractual appraisal demonstrate that expert qualification and scope of authority may be contested before the amount is resolved. An appraisal outcome is not scientifically reliable merely because it has contractual finality; equally, a scientifically informative range does not replace a contractually required procedure. The institutional record should preserve the distinction between technical disagreement, appraisal award, insurer coverage decision and ultimate payment.

United Kingdom

The Financial Conduct Authority's ICOBS 8.1.1R provides a general claims-handling rule in its applicable regulatory scope: timely and fair handling, guidance and information, appropriate claim acceptance practices and prompt settlement after agreement. The accessible Handbook version identifies the rule's January 2008 effective date. The rule does not prescribe a universal statistical interval or independent appraisal mechanism. A claims review that simply rejects a competing estimate because it exceeds an internal band would fail to investigate why it differs.

The Financial Ombudsman Service's home insurance settlement guidance evaluates factors including wording, evidence, circumstances and the practical availability of a proposed repair. It distinguishes a cash settlement chosen by the insurer from the circumstances in which a claimant prefers cash after a fair repair or replacement offer. These distinctions can affect which procurement cost is a relevant reference. Where additional independent evidence is required, the service may request it; the weight of that evidence remains a contextual judgment.

The published water-damage settlement case makes the issue concrete. The claimant obtained two quotations near GBP 10,000, while the insurer proposed GBP 6,000 following its own narrower assessment. The public narrative describes a disagreement about restoration of plaster and the lack of an actual repair offer from the insurer. The complaint was upheld using the more realistic repair basis in those circumstances. This is an illustrative public case, not independently authenticated private claim evidence and not a statistical estimate of how often underpayment occurs. Its relevance is qualitative: the difference in amount corresponded to work specification and procurement feasibility rather than simple random estimation error.

The comparative lesson is that an actionable quotation can provide stronger evidence than a numerical price assertion. Whether an insurer's network price or a claimant-accessible market price is appropriate depends on the actual settlement mechanism and applicable contract and law. A scientific comparison must normalize these assumptions before treating the amounts as interchangeable.

Australia and international supervision

The Australian Financial Complaints Authority's general insurance claims approach (2025 file edition, pp. 7–8) examines the policy, work scope, quotations, contingency and applicable benefits when assessing cash-settlement fairness. It discusses the risk transferred to the complainant when future work costs increase. The contingency considered in that context has an institutional function distinct from a statistically calibrated probability band. Even an amount calculated as a percentile does not guarantee sufficiency if the scope or valuation basis is wrong. Conversely, a negotiated contingency percentage may serve a practical remedial purpose without a defined probability coverage.

The document is an adjudicative approach, not a statutory rule requiring a uniform percentage and not an empirical study of cost overruns. Its applicability depends on the complaint and governing arrangements. Together with the UK example, it shows how procedural review can connect valuation with restoration feasibility and the allocation of risks associated with settlement mode.

At the international level IAIS expects supervisors to establish a framework for timely, fair and transparent claims handling and supports objective review. It is a common supervisory reference rather than uniform domestic law. The research uses these sources comparatively to distinguish similar functional issues under different institutional arrangements. It makes no claim that identical appraisal procedures or remedies exist in all markets.

Replicable file review

A review record should retain original estimates, their assumptions, the evidence supplied to each appraiser, the version of contract language and the controlling legal jurisdiction. Record the dates of initial and revised assessments. Where a review changes work scope or price, provide a reasoned reconciliation at item level. The reviewer should identify the monetary target and explain whether differences arise from new facts, correction of error, methodological preference or different contractual interpretations.

Review metrics should include signed and absolute valuation changes, changes in coverage or scope classification, additions and deletions of repair operations, time to resolution, reference-quality category, payment changes as a separate outcome, claimant costs and results of subsequent correction. Record cases that receive no adjustment as well as upward and downward changes. If a case is reopened later, retain the relationship to the original estimate and the reason for reopening. Do not treat a final administrative closure as definitive proof of the economic adequacy of the result.

Researchers need to distinguish disputes referred for review from the full claim population. Disputed files are selected by differences in claim complexity, claimant willingness and access to resources, insurer referral rules and potential amount. A high adjustment rate in referred files may coexist with either strong or poor original handling across the full population. A low adjustment rate may indicate reliable original work or a review process that reproduces existing assumptions. The rates alone do not identify the cause.

An empirical design for future work

A reproducible study can sample claims stratified by product, jurisdiction, valuation basis, peril, severity and evidence maturity. It should predeclare eligibility and exclusions, the outcome references, primary metrics, sample-size reasoning and the plan for inaccessible or missing evidence. It can randomize assignment to multiple competent appraisers given an identical original file and record a blind initial estimate. It can then provide later observations and conduct an open reconciliation, preserving all versions.

Where ethically and legally permissible, randomized assignment of a supplemental independent review can test its incremental effect on a predefined outcome. It must not withhold mandatory review, compromise claimants' rights or cause avoidable material harm. An alternative is a randomized order of equally permitted reviews or a staged implementation with careful accounting for time trends. Without randomization, comparison across reviewed and unreviewed claims requires explicit attention to selection and confounding. Report associations as associations; do not attribute a causal improvement to appraisal merely because cases receiving it change more often.

Reference outcome construction should itself be blinded where practical and governed by a documented protocol. Actual invoices, repair completion inspections, available market quotations and reasoned expert adjudication can be combined only after each is understood as evidence for a particular target. Outcome coders should classify whether later costs were foreseeable, whether repair scope evolved and whether the work met the defined standard. For disputed liability issues record the actual legally authoritative outcome separately. Report missingness and subgroup-specific outcome quality.

A statistical analysis should include paired bias and disagreement dispersion, empirical coverage and interval width, proper scores where probability forecasts were made, and material-error rates on auditable reference cases. Resampling or inference should respect event and appraiser clustering. Predefined equity-related analyses can consider whether outcomes or review accessibility differ among relevant groups, with lawful data use and safeguards. The design should report both improvements and costs, including claimant delay and friction. The central empirical question is whether extra review produces better decisions for the relevant target, not simply whether it reduces claim expenditure.

Integrated Analytical Synthesis

The study supports three interdependent conclusions about individual claim valuation.

First, the identity of the valuation target precedes all numerical quality assessment. A physical repair estimate, a replacement value, a claimant-accessible quotation, a legally covered amount, an insurer offer and a final payment are distinct objects. Apparent precision or agreement can be created by comparing quantities that have merely been assigned the same currency field. An error band acquired through mathematical modeling is scientifically interpretable only if the target and conditions have been specified. This is why a valuation report should show gross loss, scope alternatives and contract transformation separately.

Second, independent appraisal contributes information under specific conditions. Financial and organizational separation, technical competence, access to original evidence and methodological differentiation improve the possibility of discovering errors. An expert may identify missing data or inconsistent assumptions even without producing a definitive point value. Common source bias and shared estimator assumptions can survive multiple assessments. Review must therefore be assessed by evidence added, corrections supported and material decisions improved, as well as the dispersion of the resulting estimates.

Third, institutional review requires the attribution of disagreements to their underlying cause. Numerical deviation calls for neither automatic averaging nor automatic rejection. Distinguish factual discrepancies, legitimate price dispersion, alternative engineering scope, model specification, legal interpretation, and procedural or financial incentives. An arithmetic correction should be made directly. A price-evidence problem requires comparable market observations. A structural-scope dispute may require technical investigation. A coverage dispute requires the authority designated by the relevant policy and law. The outcome should be traceable from evidence to decision.

These conclusions can be illustrated by a cash settlement based on a highly precise estimate of a narrowly specified repair. The insurer and two experts may agree closely because all rely on the same scope and rate book. If restoration actually requires additional work or the offered price is unavailable to the claimant, agreement and precision coexist with a potentially inadequate settlement. The appropriate correction is to establish feasible scope and procurement, revise the target if necessary, and recalculate. In another case, experts may estimate different reasonable prices for the same feasible technical scope; a properly constructed range and a reasoned decision can be more informative than forcing the experts to agree on one number.

The opposite problem also arises. Extremely wide intervals may hide a process whose central estimate is unresponsive to available evidence. A wide range covers more outcomes, but its usefulness decays unless the range narrows where evidence supports narrowing. Proper scoring, paired agreement analysis and uncertainty reporting provide complementary checks. The institutional use of these statistical tools should be tied to an explicit error consequence rather than a generalized mandate for lower or higher payments.

A decision system combining these existing methods can therefore distinguish three forms of audit evidence. Technical verification establishes that computations and documented assumptions are internally consistent. Empirical validation establishes performance on a relevant, independently defined outcome population. Adjudicative verification establishes that the decision was made under applicable terms, authority and procedures. Each form answers its own question. None can establish the other two by substitution.

This synthesis is an application and critical integration of established methods and selected documented cases. It makes no claim to originate a new statistical model, legal doctrine or exclusive professional standard. Its added value lies in the case-level interpretation of uncertainty, evidence independence and procedural authority, supported by the calculations and comparative examples presented here.

Research Limitations

The research is a structured, targeted documentary synthesis rather than a registered systematic review or meta-analysis. The searches selected authoritative methodological texts, regulators, judgments and adjudicative guidance for their direct relevance to the individual claim. The comparative legal examples are limited to selected instruments and decisions and do not constitute an exhaustive survey of jurisdictions, policy forms, amendments or subsequent treatment. The FCA rule was examined on an accessible Handbook snapshot; the date-filtered main host was inaccessible. Judicial opinions and California regulation were read in published reproductions. The Australian approach and Bland–Altman paper were inspected via extracted PDF text; screenshot retrieval failed for those documents, so no independent visual full-document verification is asserted.

No proprietary claims data, authenticated policyholder case files, executed repair contracts, appraiser-level assignments or complete invoice histories were available. The insurance cases were used at the level of their published opinions and public dispute narratives, without re-estimating the actual underlying loss. Consequently the study cannot estimate the population distribution of claim valuation errors, causal effects of independent appraisal, rates of conflict of interest, claimant treatment differences or a universal optimal referral threshold. All numerical examples are synthetic and their assumptions expressly stated. Statistical techniques drawn from metrology, clinical method comparison and forecasting are applied by analogy, with scope boundaries stated.

Research access does not establish the current legal status of every source for every claim date. Actual insurance policy terms and competent jurisdiction-specific review control individual legal decisions. The purpose of this monograph is scholarly understanding and research design, not an individualized claim recommendation. As empirical claim-level evidence becomes available, the appropriate next investigation is independent reference construction and controlled multi-appraiser validation under the design outlined above.

Conclusion

Claim valuation error bands express uncertainty around a specified monetary object or within explicitly identified scenarios. Their interpretation depends on the evidentiary basis, valuation date, method, distributional assumptions and legal context. An interval can be useful even in the absence of a single exact observable amount, but the type of interval and its intended meaning must be clear. A probability statement requires an appropriate stochastic framework; an expert range or sensitivity analysis carries a different interpretation.

Independent appraisal has scientific value when it adds reliable and differentiated knowledge to the assessment. Independence includes interests, competence, evidence access and method rather than simply a separate appointment. Statistical agreement can be tested through paired differences and suitably designed multi-rater studies, and calibration can be assessed through independently defined outcomes, interval coverage, miss patterns and scoring. These tests support decisions within their observed domains; they do not establish an insured's entitlement alone.

The practical institutional conclusion is to resolve the underlying source of disagreement rather than average estimates indiscriminately. Make corrected facts explicit, compare feasible work specifications and price bases, record uncertainty without false precision, and submit genuinely legal issues to the competent authority. Preserve the distinction between the gross valuation, the contractually governed amount, the review outcome and the actual transaction. A claims system that can explain these steps can support more replicable assessments and more defensible adjudication.

References

  • Taylor, B. N., and Kuyatt, C. E. (1994). Guidelines for Evaluating and Expressing the Uncertainty of NIST Measurement Results. NIST Technical Note 1297, §5, Combined Standard Uncertainty. National Institute of Standards and Technology. Source

  • Taylor, B. N., and Kuyatt, C. E. (1994). Guidelines for Evaluating and Expressing the Uncertainty of NIST Measurement Results. NIST Technical Note 1297, §6, Expanded Uncertainty. National Institute of Standards and Technology. Source

  • Taylor, B. N., and Kuyatt, C. E. (1994). Guidelines for Evaluating and Expressing the Uncertainty of NIST Measurement Results. NIST Technical Note 1297, §7, Reporting Uncertainty. National Institute of Standards and Technology. Source

  • Bland, J. M., and Altman, D. G. (1986). Statistical methods for assessing agreement between two methods of clinical measurement. The Lancet, 327(8476), 307–310. DOI: 10.1016/S0140-6736(86)90837-8. University-hosted reproduction. Source

  • Gneiting, T., and Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477), 359–378. DOI: 10.1198/016214506000001437. Authors' university-hosted copy. Source

  • Angelopoulos, A. N., and Bates, S. (2021). A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv:2107.07511. Preprint; related published tutorial: Conformal Prediction: A Gentle Introduction (2023), Foundations and Trends in Machine Learning, 16(4), 494–591, DOI: 10.1561/2200000101. Source

  • International Association of Insurance Supervisors (IAIS). (2024). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups, revisions adopted December 5, 2024. ICP 19, Conduct of Business, §19.10 and guidance 19.10.1–19.10.12. Online consolidated tool. Source

  • National Association of Insurance Commissioners (NAIC). (1997). Unfair Claims Settlement Practices Act. Model 900, January 1997 compilation, §§1, 3–4. Adopted June 1990. Source

  • California Code of Regulations, title 10, §2695.8, Additional Standards Applicable to Automobile Insurance, particularly subsection (f). Legal Information Institute, Cornell Law School, reproduction of regulatory text. Source

  • State Farm Lloyds v. Johnson, 290 S.W.3d 886 (Tex. 2009). Supreme Court of Texas, majority opinion, No. 06-1071. Justia reproduction. Source

  • Parrish v. State Farm Florida Insurance Co., 356 So. 3d 771 (Fla. 2023). Supreme Court of Florida, No. SC21-172, February 9, 2023. Justia reproduction. Source

  • Financial Conduct Authority (FCA). Insurance: Conduct of Business Sourcebook (ICOBS), chapter 8, Claims Handling, rule 8.1.1R (effective January 6, 2008 as displayed; accessible chapter snapshot last updated December 9, 2025). FCA Handbook accessible host. Source

  • Australian Financial Complaints Authority (AFCA). (2025). The AFCA Approach to General Insurance Claims Handling. Public approach document, July 2025 file edition, pp. 7–8. Source

  • Financial Ombudsman Service (United Kingdom). (n.d.). Settling home insurance claims. Business guidance, sections on evidence, repair quality and cash settlement. Accessed November 19, 2026. Source

  • Financial Ombudsman Service (United Kingdom). (n.d.). Simon's insurer disputes his estimates for repairing water damage. Published illustrative case study. Accessed November 19, 2026. Source