Claims Analytics, Reserving and Fraud Detection

Introduction and Context

Insurance claims are the point at which an insurer’s promise becomes an operational, financial and reputational obligation. Claims data therefore contain information about loss frequency, severity, customer behaviour, repair networks, legal outcomes, recovery opportunities and emerging risks. Predictive analytics converts this information into forecasts and decisions: which claims require urgent handling, how much a claim may ultimately cost, whether litigation is likely, whether recovery is possible, and whether a pattern is consistent with fraud.

Claims analytics is not simply a modelling exercise. It is a decision system operating under uncertainty, time pressure, legal duties, unequal information and limited investigation capacity. A highly accurate model may still destroy value if it delays vulnerable customers, systematically refers legitimate claims for investigation, or produces reserves that are financially misleading. The central managerial task is to connect model outputs to proportionate actions and to monitor the consequences.

This chapter examines predictive methods for claims triage, severity estimation, settlement, litigation, recovery, reserving, catastrophe response and fraud detection. It also addresses rare-event learning, false-positive costs, explainability, governance and the distinction between prediction and decision-making.

Learning Objectives

By the end of this chapter, the learner should be able to:

1. Explain how claims data support operational and financial decisions.
2. Design triage, severity, litigation, settlement and recovery models.
3. Distinguish reported, paid, incurred and ultimate claim amounts.
4. Apply chain-ladder, expected-loss and stochastic reserving concepts.
5. Design anomaly, supervised and network-based fraud detection systems.
6. Evaluate models using business costs, calibration, lift and capacity constraints.
7. Identify ethical, regulatory, privacy and governance requirements.
8. Critically assess a claims analytics programme in an insurance organisation.

1. The Claims Analytics Decision Architecture

A claim develops through a sequence of events rather than appearing as a single observation. Typical stages are notification, coverage verification, initial assessment, adjustment, repair or medical management, settlement, payment, recovery and closure. At each stage, information is added and the prediction target changes.

Important targets include:

• Triage priority: the urgency and complexity of handling.
• Claim severity: the eventual gross or net cost.
• Time to settlement: duration until final settlement or closure.
• Litigation probability: likelihood of a legal dispute or court process.
• Recovery probability: likelihood and expected value of salvage, subrogation or contribution.
• Reserve adequacy: the amount required for unpaid obligations.
• Fraud or irregularity risk: the probability that a claim contains intentional misrepresentation or an unusual pattern requiring review.

A useful architecture separates four layers:

The distinction between prediction and decision is fundamental. If P(Fraud\mid x) is the predicted probability of fraud given information x, an investigation decision should not automatically follow whenever the probability exceeds 0.5. The correct threshold depends on investigation cost, expected prevented loss, customer harm and available capacity.

2. Data Foundations and Feature Design

Claims analytics commonly integrates:

• policy and coverage records;
• first-notification-of-loss (FNOL) information;
• claim transactions and reserve movements;
• payment, supplier and repair data;
• customer, broker and agent information;
• medical, legal and social-inflation indicators where lawful;
• geospatial and weather data;
• vehicle, property or sensor information; and
• prior claims and external fraud intelligence.

Data must be indexed by time. A feature is usable only if it was available when the decision was made. Using a final settlement amount, a later reserve movement or a post-investigation flag to predict an FNOL outcome is target leakage. Leakage creates impressive validation results but fails in production.

Feature design should reflect the claim process. Examples include elapsed time since FNOL, number of reserve changes, claimant–provider relationship, repair estimate variance, policy age, prior claim count within a defined look-back period, distance between incident and repairer, and weather intensity at the reported location. Features should have documented definitions, units, permissible use, source systems and missing-value treatment.

Missingness is often informative. A missing repair estimate may indicate that inspection has not occurred, rather than a random data defect. However, treating every missing value as suspicious can penalise customers with less digital access. Missingness must therefore be investigated substantively and tested for disparate impact.

Data splitting should be temporal where possible: train on earlier claims, validate on a later period and test on a still later period. Random splitting can leak common provider behaviour or expose the model to future conditions. Catastrophes, inflation changes and regulatory reforms require additional stress testing.

3. Claims Triage and Operational Routing

Triage assigns claims to service pathways. A low-complexity claim may be handled through straight-through processing; a high-severity or vulnerable-customer claim may require a specialist adjuster. Triage is not equivalent to fraud detection. Complexity, vulnerability and urgency are legitimate reasons for escalation even when fraud risk is low.

A triage model may estimate expected handling cost, expected severity, probability of litigation, probability of vulnerability, or expected delay. A simple severity-based segmentation can use predicted ultimate cost:

• low predicted cost and low complexity: automated or fast-track route;
• moderate cost or uncertainty: standard adjuster route;
• high cost, injury, vulnerability or litigation risk: specialist route.

A decision rule should include capacity. If only a fixed number of specialist cases can be handled daily, ranking by a priority score is more appropriate than applying an arbitrary probability threshold. Priority scores may combine predicted harm, urgency, expected savings and service obligations, but the weights must be governed and reviewed.

Performance should include cycle time, customer outcomes, leakage, adjustment cost, reserve accuracy and complaints—not merely model discrimination. A model that routes too many claims to specialists may improve accuracy while creating queues and worsening service.

4. Severity, Settlement and Litigation Models

4.1 Severity estimation

Claim costs are typically right-skewed: many small claims coexist with a few very large losses. A single ordinary least-squares model may be poorly suited to this distribution. Common approaches include:

• a two-part model for claim occurrence and positive severity;
• generalized linear models (GLMs) with Gamma or inverse Gaussian severity distributions;
• log-normal models for positive costs;
• quantile regression for conditional percentiles;
• gradient-boosted trees and other machine-learning models; and
• hierarchical or credibility models where providers, regions or portfolios differ.

For a positive claim amount Y, a log-normal model assumes ln(Y) is normally distributed conditional on covariates. If ln(Y)\sim N(\mu,\sigma²), then the conditional mean on the original cost scale is:

Here μ(x) is the conditional mean of log-cost and σ² is the conditional log-scale variance. The term σ²/2 is necessary when converting a log-scale prediction to an arithmetic mean; simply exponentiating μ(x) estimates the conditional median, not the mean. This formula assumes the log-normal specification is appropriate and that the variance is correctly interpreted.

A two-part expected-cost model is:

The first component models the probability of a positive loss or payment; the second models its conditional size. This is useful where the portfolio contains many zero-payment observations, although claims datasets must distinguish genuinely zero claims from open claims not yet paid.

For reserving and settlement, point predictions are insufficient. Prediction intervals or quantiles communicate uncertainty. A high percentile may be appropriate for a large-loss reserve or authority limit, while the mean may be appropriate for portfolio budgeting.

4.2 Settlement recommendations

Settlement analytics can estimate a fair or economically rational settlement range, but the recommendation must respect policy terms, evidence, claims-handling standards and negotiation authority. A model may predict expected ultimate cost, probability of additional documentation, or likelihood that a settlement offer will be accepted. It should not be designed merely to minimise payments without considering contractual rights and good-faith obligations.

Decision analysis can compare immediate settlement with continued handling. If C_s is the cost of settling now and C_c is the expected cost of continuing, a simplified net-benefit rule is to settle when:

where V_c represents the value of avoiding delay, legal exposure, handling expense and customer dissatisfaction. The terms must be defined consistently; this is a managerial decision rule, not a universal accounting formula.

4.3 Litigation prediction

Litigation models may estimate the probability of legal escalation or expected legal cost. Predictors can include injury type, attorney involvement, jurisdiction, elapsed time, prior communication, disputed liability and claim complexity. These variables can encode structural inequality or historic claims-handling practices. Legal involvement should not be treated as proof of dishonesty.

Evaluation should consider calibration by jurisdiction and claimant group. A model that predicts litigation accurately in one region may fail after a legislative change. Decisions influenced by the model should remain reviewable by qualified claims professionals.

5. Recoveries and Catastrophe Response

Recoveries include salvage, subrogation, contribution and reinsurance recoveries. A useful framework estimates both probability and amount:

where R is the recovery amount. The model may be used to prioritise evidence collection, identify responsible third parties, or decide whether pursuit is economically justified. Recovery models should account for administrative and legal costs, time value of money and the probability that pursuit damages a customer or commercial relationship.

Catastrophe response requires rapid adaptation. After a flood, cyclone, earthquake or wildfire, normal claims patterns may break down. Useful methods include geospatial exposure overlays, remote sensing, event severity indices, triage queues and surge-capacity forecasting. Models should be stress-tested for damaged communications, duplicate notifications, displaced customers and incomplete addresses.

A catastrophe operating model should:

1. identify affected policies and likely loss zones;
2. prioritise life safety, vulnerability and urgent mitigation;
3. separate event-related claims from ordinary baseline claims;
4. deploy mobile adjusters, approved suppliers and reserve authority;
5. monitor settlement speed, leakage and emerging fraud patterns; and
6. update forecasts as late reports and reopened claims arrive.

Automation is particularly risky immediately after a catastrophe because data quality and customer circumstances are unstable. Human escalation and clear exception rules are essential.

6. Reserving Analytics

6.1 Claims development concepts

A claim may be reported late, paid in several instalments and revised repeatedly. Key quantities are:

• Reported incurred: paid amount plus case outstanding reserve.
• Paid loss: cash paid to date.
• Case reserve: adjuster’s current estimate for a known claim.
• IBNR: incurred but not reported, commonly used for losses not yet reported or not fully recognised under a particular reserving framework.
• Ultimate loss: the estimated final cost after all development.
• Unpaid claim liability: the amount expected to be paid in the future, subject to the applicable accounting and actuarial basis.

Terminology and presentation differ by jurisdiction and reporting framework. Predictive analytics supports reserving but does not replace actuarial judgement, governance or applicable standards.

6.2 Chain-ladder method

The chain-ladder method uses development triangles. Let C_i,j be cumulative claims for accident year or underwriting year i observed at development age j. An age-to-age development factor is:

The summation includes the selected rows with observations at both development ages. The projected ultimate for a cell observed at age k is:

where J is the selected ultimate development age. The implied reserve is U_i,k−C_i,k, using the same cumulative basis for C and U. The method assumes that historical development patterns are informative about future development, and it can be distorted by changing claims practices, inflation, mix, catastrophes or sparse data.

Worked example. Suppose cumulative paid losses for two origin periods at development ages 12 and 24 months are:

The 12-to-24-month factor is:

If origin period C has paid losses of 700 at 12 months, its projected 24-month cumulative paid amount is 700×1.5=1,050. The projected additional payment is 1,050−700=350. This is only a one-step projection; a full reserve requires later development factors, an ultimate age and consideration of other reserving evidence.

6.3 Expected-loss and Bornhuetter–Ferguson approaches

Where early claims experience is volatile, an a priori expected ultimate loss may be more stable. The expected-loss method begins with exposure and an expected loss ratio:

For example, exposure may be earned premium, payroll, vehicle-years or insured value, depending on the line. The expected loss ratio must be defined on a compatible basis, including whether losses are gross, net, paid or incurred.

The Bornhuetter–Ferguson approach combines an a priori expected ultimate with the proportion of ultimate loss expected to emerge. If E is expected ultimate loss and p is the expected proportion reported or paid by the valuation age, then:

The formula assumes the reported amount and emergence proportion are measured on compatible bases. It places less weight on immature observed experience than chain-ladder, but it is sensitive to the quality of E and p.

6.4 Predictive and stochastic reserving

Modern reserving may use claim-level models, survival models for closure, GLMs, gradient boosting, Bayesian hierarchical models and bootstrap or simulation techniques. Stochastic methods estimate a distribution rather than only a central estimate. For aggregate unpaid claims S, useful outputs include the mean, standard deviation and selected percentiles. A reserve decision may require a central estimate, risk margin or capital view depending on the reporting purpose.

Model validation should include back-testing: compare historical forecasts with subsequent development, analyse calendar-year deterioration, test reserve error by segment and examine sensitivity to inflation, large losses and claim-handling changes. Reserve monitoring is a feedback system, not a one-time model approval.

7. Fraud Detection: Models and Operating Design

Fraud is a rare-event and adversarial problem. The observed label is often incomplete: investigated claims may be more likely to receive a fraud label, while undetected fraud remains unobserved. A “fraud” label may also combine confirmed fraud, repudiation, suspicious circumstance and administrative referral. These categories should not be treated as identical.

7.1 Supervised models

Where reliable labels exist, supervised classification can estimate P(Y=1\mid x), with Y=1 representing a defined fraud outcome. Logistic regression is interpretable:

Here p is predicted probability, β₀ is the intercept, β_k are coefficients and x_k are features. The model assumes a linear relationship between predictors and log-odds unless transformations or interactions are added. Coefficients should not be interpreted causally merely because they are statistically significant.

Tree ensembles, boosted models and neural networks can capture nonlinear interactions, but they require stronger controls for leakage, calibration, drift and explainability. Class weights, sampling strategies and calibrated probabilities may help with imbalance, but synthetic oversampling does not create genuine fraud examples and can distort relationships.

7.2 Anomaly detection

Anomaly methods identify observations unlike a reference population. Examples include isolation forests, robust distance methods, autoencoders and peer-group comparisons. An anomaly is not necessarily fraud: it may represent a catastrophe, a new repair technology, a legitimate high-value customer or a data error. Anomaly scores should therefore trigger contextual review rather than automatic denial.

7.3 Network and graph analytics

Fraud frequently involves relationships among claimants, addresses, vehicles, bank accounts, repairers, medical providers, lawyers and witnesses. A graph represents entities as nodes and relationships as edges. Repeated shared identifiers, unusually dense clusters, reciprocal referrals or coordinated timing can reveal organised activity.

Network models must address legitimate concentration. A specialist repairer may serve many customers in a small region; a family may share an address; a hospital may appear in many injury claims. Graph signals should be combined with claim-level evidence, and relationship data require careful privacy and access controls.

7.4 Rare events and investigation capacity

Accuracy is a poor metric when fraud prevalence is low. With prevalence π, sensitivity s and specificity c, the positive predictive value is:

This follows from Bayes’ rule. It shows why even a specific model can generate many false positives when fraud is rare.

Worked example. Suppose fraud prevalence is 1%, sensitivity is 80% and specificity is 95%. Among 10,000 claims, expected true fraud cases are 100, of which 80 are detected. Non-fraud cases total 9,900, and 5%—495—are false positives. Thus:

PPV = 80 ÷ (80+495) ≈ 13.9%.

Most referrals would therefore not be confirmed fraud. The operational response should be proportionate: a low-cost information check may be suitable for a broad alert, while intrusive investigation should require stronger evidence.

If the cost of investigating a referral is C_I, the expected cost of a false positive is C_FP, and the expected benefit of correctly detecting fraud is B_TP, the threshold should be selected using expected utility rather than accuracy alone. In practice, benefit also includes deterrence and recovered amounts, while harm includes delay, distress, complaints and reputational damage.

A capacity-constrained organisation may investigate only the top K ranked claims. Evaluation should then use precision at K, recall at K, confirmed value per investigation and investigator workload. Thresholds should be recalibrated when staffing or fraud patterns change.

8. Evaluation, Monitoring and Model Risk

Discrimination measures include area under the ROC curve (AUC), precision–recall curves and lift. AUC measures ranking performance across thresholds; it does not establish that a particular operational threshold is useful. Precision–recall analysis is often more informative for rare events.

Calibration examines whether predicted probabilities correspond to observed frequencies. A group of claims assigned probability 0.20 should experience the outcome approximately 20% of the time over a sufficiently large and comparable population. Calibration can deteriorate after inflation, regulatory change or new fraud tactics.

Monitoring should cover:

• input drift and missingness;
• prediction distribution and calibration;
• outcome and label delay;
• subgroup performance and fairness;
• referral volumes and investigator capacity;
• customer complaints and appeal outcomes;
• reserve error and claims leakage;
• cyber and access controls; and
• changes in supplier, legal or catastrophe conditions.

Champion–challenger testing, periodic back-testing and independent validation reduce model risk. Every production model should have an owner, purpose statement, approved features, limitations, version history, override policy and retirement criteria.

9. Governance, Ethics and Regulation

Claims analytics affects access to benefits, speed of payment and the treatment of vulnerable people. Governance should include:

1. Lawful purpose and minimisation: collect and use only data necessary for a defined purpose.
2. Transparency: communicate relevant reasons for delay, referral or adverse action where required.
3. Human oversight: provide meaningful review, not merely a nominal approval of automated output.
4. Fairness testing: examine disparate error rates and outcomes across relevant groups, subject to lawful data availability.
5. Contestability: allow correction of inaccurate data and reconsideration of decisions.
6. Security: control access to sensitive medical, financial and identity data.
7. Evidence preservation: retain records supporting important claim decisions.
8. Third-party governance: audit vendors, data brokers, repair networks and model providers.

Fairness is not solved by removing protected attributes. Proxy variables can reproduce them, and removing them may prevent measurement of disparate impact. Organisations should distinguish justified risk segmentation from prohibited discrimination and obtain legal and actuarial advice in the relevant jurisdiction.

Fraud analytics also creates a risk of circularity. If a model refers a group more often, the group generates more investigations and therefore more labels, apparently validating the original model. Label quality audits, random samples of non-referred claims and independent investigations help reduce this feedback loop.

10. Illustrative Integrated Case: Meridian Motor Insurance

Background

Meridian Motor Insurance, a fictional regional insurer, receives 45,000 motor claims annually. Average settlement time is 24 days, specialist adjusters are overloaded, and suspected organised fraud is concentrated around a small number of repairers and bank accounts. Reserve deterioration is significant for bodily-injury claims.

Problem

Management wants one “claims score.” The analytics team rejects this proposal because triage, severity, fraud, litigation and reserving have different targets, labels, time horizons and ethical risks.

Analysis

The team maps the claims process and finds three data problems: post-settlement variables had entered the fraud model, claim labels mixed confirmed fraud with mere referral, and historical specialist routing created selection bias in severity data. It creates a temporal training sample, separates confirmed outcomes from suspicion flags, and documents data availability at FNOL.

Approach and solution

The insurer deploys four linked but distinct models:

• a triage model estimating complexity and vulnerability-related urgency;
• a severity model using a two-part approach and a high-cost quantile;
• a fraud-ranking model combining supervised features, anomaly signals and a provider–claimant network;
• a claim-level reserving model reconciled to actuarial portfolio estimates.

The fraud model is used to rank the top 400 cases each month, matching investigation capacity. Low-confidence alerts request documents without using accusatory language. High-confidence cases receive specialist review. No claim is denied solely because of a model score.

A claims committee examines subgroup calibration, investigation yield, complaints, and reserve back-testing monthly. An independent validation team reviews drift quarterly. Catastrophe overrides suspend normal thresholds when an event is declared, preventing legitimate concentration around a disaster from being misclassified as fraud.

Outcome

In the first six months, average settlement time falls for low-complexity claims, specialist capacity is redirected to severe and vulnerable cases, and confirmed fraud value per investigation increases. The fraud referral count does not increase, because ranking replaces indiscriminate thresholding. Reserve forecasts become more stable, although bodily-injury uncertainty remains high. Management concludes that the principal benefit came from redesigning decisions and governance, not from selecting the most complex algorithm.

Learning points

The case illustrates that models must be aligned to decisions, labels must represent defined outcomes, and capacity constraints belong in model design. It also shows why catastrophe controls, human review, monitoring and actuarial reconciliation are essential.

11. Best-Practice Implementation Roadmap

A practical implementation sequence is:

1. define the business decision and harm of error;
2. specify the prediction target, observation window and outcome window;
3. map available-at-decision-time data and prevent leakage;
4. establish a transparent baseline model;
5. compare advanced models using temporal validation;
6. calibrate outputs and select action thresholds or ranking capacity;
7. test fairness, stability, stress and operational workload;
8. pilot with human review and measure outcomes;
9. document governance, overrides and appeal processes; and
10. monitor continuously and retire or revise the model when conditions change.

The most effective programmes combine actuaries, claims professionals, fraud investigators, data scientists, legal specialists, customer representatives and information-security staff. Model performance is only one component of organisational performance.

Challenges and Contemporary Developments

Claims data are increasingly enriched by telematics, connected homes, satellite imagery, natural-language processing and generative-AI tools. These technologies can improve document extraction, summarise claim files and identify inconsistencies, but they introduce new risks: fabricated summaries, opaque vendor models, sensitive surveillance, cybersecurity exposure and automation bias. Generative systems should not independently determine coverage or fraud outcomes without controlled evidence retrieval, human review and audit logging.

Climate change challenges historical reserving and catastrophe models because frequency, severity and spatial dependence may be non-stationary. Inflation and social inflation can invalidate historical severity relationships. Synthetic data may support testing but cannot substitute for representative real outcomes. Federated or privacy-enhancing analytics may allow cross-organisation learning, but governance and legal compatibility remain necessary.

Chapter Summary

Claims analytics links data, prediction, decision and governance across the claim life cycle. Triage models should route claims according to complexity, urgency and capacity rather than fraud risk alone. Severity models must address skewness, zeros and uncertainty; settlement and litigation predictions require legal and customer safeguards. Recovery analytics estimates both the chance and value of salvage or subrogation, while catastrophe analytics requires geospatial intelligence and flexible operational controls.

Reserving uses cumulative development, expected-loss assumptions and claim-level predictive methods. Chain-ladder extrapolates historical development factors, whereas Bornhuetter–Ferguson combines observed emergence with an a priori expected ultimate. Fraud detection may combine supervised classification, anomaly detection and network analysis, but rare-event mathematics demonstrates why false positives can dominate referrals. Thresholds should reflect investigation capacity and the costs of error.

Reliable implementation requires temporal validation, leakage prevention, calibration, subgroup testing, back-testing, human oversight and documented governance. Predictive analytics improves insurance only when its outputs are connected to fair, proportionate and measurable decisions.

Key Terms

Claims triage: Routing claims to handling pathways according to urgency, complexity or expected needs.

Severity: The monetary size or ultimate cost of a claim.

FNOL: First notification of loss, the initial report of an insured event.

Target leakage: Use of information unavailable at the time a prediction would have been made.

Ultimate loss: The estimated final cost of a claim or cohort after development is complete.

Case reserve: An outstanding estimate assigned to a known claim.

IBNR: Incurred but not reported or not fully recognised losses, depending on the reserving context.

Development triangle: A table showing claims by origin period and development age.

Chain-ladder: A reserving method that projects cumulative claims using development factors.

Bornhuetter–Ferguson: A reserving method combining observed emergence with an a priori expected ultimate.

Recovery: Amount obtained through salvage, subrogation, contribution or reinsurance.

Rare-event classification: Prediction where the outcome of interest occurs infrequently.

Sensitivity: The proportion of actual positive cases correctly identified.

Specificity: The proportion of actual negative cases correctly classified as negative.

Positive predictive value: The proportion of positive alerts that are truly positive.

Anomaly detection: Identification of observations unlike a reference population.

Network analytics: Analysis of relationships among entities represented as nodes and edges.

Calibration: Agreement between predicted probabilities and observed frequencies.

Concept drift: Change in the relationship between predictors and outcomes over time.

Model governance: Processes controlling model purpose, validation, use, monitoring and accountability.

Review Questions

Short Answer Questions

1. Why should claims triage and fraud detection be treated as different decision problems?
2. Define reported incurred, case reserve and ultimate loss.
3. What is target leakage, and why is temporal validation useful?
4. State two applications of network analytics in fraud detection.
5. Why can accuracy be misleading for rare-event fraud models?

Long Answer Questions

1. Explain how a two-part severity model can be used in claims analytics. State its assumptions and limitations.
2. Describe the chain-ladder method, including the meaning of development factors and reserve estimates.
3. Compare supervised, anomaly-based and network-based fraud detection.
4. Discuss how investigation capacity and false-positive costs should influence fraud-model thresholds.
5. Evaluate the governance requirements for an automated claims settlement or triage system.

Analytical / Application Questions

1. A fraud model is applied to 20,000 claims. Fraud prevalence is 2%, sensitivity is 75% and specificity is 97%. Calculate the expected number of true positives, false positives and positive predictive value. Interpret the operational implications.
2. An insurer has cumulative paid losses of 500 and 750 at development ages 12 and 24 months for one origin period, and 900 and 1,350 for another. Calculate the aggregate 12-to-24-month development factor. If a third origin period has 800 at 12 months, estimate its 24-month cumulative amount and additional payment.
3. Design an integrated claims analytics pilot for a catastrophe-prone motor insurer. Specify targets, data controls, models, capacity constraints, governance measures and outcome metrics.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from decision in claims analytics?

A. A predicted probability above 0.5 must always trigger investigation.

B. A decision threshold should reflect costs, benefits, capacity and customer impact.

C. Prediction and decision are identical when a model is calibrated.

D. Decisions should use only the model’s most important feature.

Correct Answer: B. A decision threshold should reflect costs, benefits, capacity and customer impact.

Question 2

In a two-part severity model, expected claim cost is calculated as:

A. Probability of a positive claim plus conditional severity

B. Probability of a positive claim divided by conditional severity

C. Probability of a positive claim multiplied by conditional positive severity

D. Conditional severity minus probability of a positive claim

Correct Answer: C. Probability of a positive claim multiplied by conditional positive severity

Question 3

Using two cumulative development pairs, the age-to-age chain-ladder factor is calculated as:

A. The average of individual claim counts

B. The ratio of summed later cumulative amounts to summed earlier cumulative amounts

C. The difference between later and earlier cumulative amounts

D. The ratio of exposure to premium

Correct Answer: B. The ratio of summed later cumulative amounts to summed earlier cumulative amounts

Question 4

Why is target leakage especially harmful in claims modelling?

A. It reduces the number of available variables.

B. It makes the model appear accurate using information unavailable at decision time.

C. It always causes underfitting.

D. It prevents temporal validation.

Correct Answer: B. It makes the model appear accurate using information unavailable at decision time.

Question 5

Suppose fraud prevalence is 1%, sensitivity is 80% and specificity is 95%. Among 10,000 claims, approximately how many false positives are expected?

A. 80

B. 100

C. 495

D. 9,405

Correct Answer: C. 495

Question 6

Which measure is particularly useful when an insurer can investigate only the top K ranked claims?

A. Precision at K

B. Mean claim age only

C. Overall accuracy without a threshold

D. Number of model features

Correct Answer: A. Precision at K

Question 7

An anomaly score should generally be interpreted as:

A. Proof that a claim is fraudulent

B. A reason for automatic denial

C. Evidence that a claim differs from a reference population and may need context

D. A replacement for policy wording

Correct Answer: C. Evidence that a claim differs from a reference population and may need context

Question 8

What is the principal purpose of calibration testing?

A. To determine whether predicted probabilities correspond to observed frequencies

B. To maximise the number of variables

C. To remove all missing values

D. To guarantee causal interpretation of coefficients

Correct Answer: A. To determine whether predicted probabilities correspond to observed frequencies

Question 9

In the chain-ladder worked example, if the aggregate 12-to-24-month factor is 1.5 and a new origin period has 700 at 12 months, the projected 24-month cumulative amount is:

A. 350

B. 700

C. 1,050

D. 1,500

Correct Answer: C. 1,050

Question 10

Which governance practice helps detect circularity in fraud labels?

A. Using only previously referred claims for validation

B. Randomly sampling non-referred claims for independent review

C. Removing all human investigation

D. Increasing the threshold until no claims are referred

Correct Answer: B. Randomly sampling non-referred claims for independent review