Underwriting and Claims Analytics and Performance Management

Introduction

Insurance performance emerges from connected underwriting and claims decisions. Underwriters determine which risks are accepted, how they are classified, and what premium and terms apply. Claims functions validate coverage, estimate liabilities, settle legitimate claims, control leakage and support customers. Analytics connects these activities by converting policy, exposure, premium, claim, reserve and customer data into evidence for management action.

A strong performance system does more than display totals. It must distinguish changes in claim frequency from changes in severity, compare segments on a consistent exposure basis, evaluate reserve adequacy, identify operational bottlenecks and reveal trade-offs between cost control and customer outcomes. Metrics must be defined consistently, models must be validated, and decisions must be communicated through concise management reports.

Figure: The closed-loop insurance analytics and performance management cycle.

Learning Objectives

After studying this chapter, students should be able to:

1. Calculate and interpret frequency, severity, loss ratios and reserve adequacy measures.
2. Design dashboards with valid definitions, targets, segmentation and escalation rules.
3. Evaluate claim duration, leakage, productivity and customer outcomes quantitatively.
4. Explain how predictive models support underwriting and claims decisions.
5. conduct scenario and sensitivity analysis without confusing predictions with certainty.
6. Assess model, data, ethical and regulatory risks.
7. Communicate findings through decision-oriented management reports.

1. Measurement Foundations

1.1 From data to management action

A performance measure is useful only when it is linked to a decision. A complete metric specification should identify:

• its business purpose;
• numerator and denominator;
• unit of measurement;
• exposure or cohort basis;
• valuation date and development maturity;
• data owner and source;
• target and escalation threshold; and
• permitted exclusions or adjustments.

For example, “claim frequency increased” is incomplete unless the report specifies the claim definition, earned exposure, period, product segment and treatment of reopened claims.

Three dates are particularly important:

1. Occurrence date: when the insured event happened.
2. Report date: when the insurer was notified.
3. Payment or closure date: when money was paid or the claim was closed.

Accident-period analysis groups claims by occurrence date and is useful for risk performance. Report-period analysis supports workload monitoring, while calendar-period analysis reflects current accounting activity. Mixing these bases can create misleading trends.

1.2 Exposure, premium and loss bases

Exposure must correspond to the insured risk. Common units include policy-years, vehicle-years, employee-years, property-years and insured payroll. Partial periods should normally be converted into proportional units where risk is approximately uniform over time.

Written premium is recorded when business is written. Earned premium is the portion attributable to expired coverage. Loss ratios should ordinarily use earned premium so that premium and losses relate to the same coverage period.

The word incurred requires precise qualification. At an operational accounting level, reported incurred losses at valuation date t are often defined as:

This quantity does not include incurred-but-not-reported claims unless the data definition expressly includes them. By contrast, an actuarial estimate of ultimate loss is:

Here, IBNR represents claims that have occurred but have not yet been reported, while IBNER represents future development beyond current estimates on reported claims. Some organizations use “IBNR” broadly to include IBNER. Reports must disclose the convention. Ultimate loss is an estimate and may change as experience develops.

2. Core Underwriting and Claims Metrics

2.1 Claim frequency

Claim frequency measures claim incidence relative to exposure:

where N is the number of claims and E is earned exposure in the corresponding unit. If 240 claims arise from 12,000 vehicle-years, frequency is 240 ÷ 12,000 = 0.02, or 2 claims per 100 vehicle-years.

Frequency comparisons require consistent claim-count rules. One event producing several coverage payments may be counted as one claim or several coverage claims. Neither convention is universally correct, but consistency is essential.

2.2 Claim severity

Average severity is:

where L is the selected loss amount in currency units and N is the corresponding claim count. If ultimate losses are CU3.6 million for 240 claims, average ultimate severity is CU15,000 per claim.

Paid, reported incurred and ultimate severity answer different questions. Paid severity can understate costs for immature periods. Closed-claim severity may also be biased because simple claims close earlier than complex claims.

2.3 Pure premium and loss ratio

Under compatible definitions, the expected loss cost per exposure unit is:

This identity assumes the count and loss definitions are aligned. Using the previous figures gives 0.02 × CU15,000 = CU300 per vehicle-year.

The loss ratio is:

The numerator must be labelled. A paid loss ratio uses paid losses; a reported incurred ratio uses paid plus case reserves; an ultimate loss ratio uses estimated ultimate losses, including the selected IBNR and IBNER provision. If estimated ultimate losses are CU78 million and earned premium is CU120 million, the ultimate loss ratio is 65%.

The combined ratio extends the analysis:

This additive form applies when both ratios use the same premium denominator. A combined ratio below 100% indicates an underwriting profit before investment income and financing effects.

2.4 Rate and mix interpretation

An improving aggregate loss ratio may result from higher rates rather than better risk selection. Conversely, growth in a high-risk segment can worsen the aggregate even if each segment is stable. Analysis should therefore separate:

• exposure growth;
• premium rate change;
• risk-mix change;
• claim frequency change;
• severity inflation; and
• claim-development change.

Segment-level results can move differently from portfolio totals, an effect related to Simpson’s paradox. Management should inspect both aggregate and appropriately credible segment results.

3. Reserve Adequacy and Loss Development

3.1 Purpose and concepts

A claim reserve is a liability estimate for unpaid claim obligations. Reserve adequacy affects solvency, profitability, pricing and management credibility. Relevant components include case reserves, IBNR, IBNER and claim-handling expenses where required by the applicable accounting or regulatory basis.

A simple reserve adequacy ratio is:

Both quantities must use the same scope, valuation date, discounting basis and expense treatment. A ratio below 100% indicates that the held amount is below the selected indication; it does not prove the indication is correct.

3.2 Chain-ladder development

For cumulative losses C_i,j for origin period i at development age j, a volume-weighted age-to-age factor is:

The sum includes only origin periods observed at both ages. A cumulative development factor from age j to ultimate is the product of subsequent selected factors, including any explicit tail factor:

The projected ultimate and unpaid reserve are:

For a paid-loss triangle, the latest cumulative value is paid loss. Chain-ladder assumes that historical development patterns remain sufficiently applicable and that changes in settlement, coverage, inflation and case management are understood.

3.3 Worked reserve example with an explicit tail

Consider cumulative paid losses in CU thousands:

The volume-weighted 12-to-24-month factor is:

Only 2021 is observed at both 24 and 36 months, so:

Losses are not assumed to be fully developed at 36 months. Based on longer historical data and actuarial judgment, suppose a 36-month-to-ultimate tail factor of 1.05 is selected. Therefore:

The resulting estimates are:

All monetary values are CU thousands. The 12-month factor of 1.75 is a genuine 12-month-to-ultimate factor because it explicitly includes the 36-month-to-ultimate tail. Without the tail, 1.6667 would represent only development from 12 to 36 months.

This small triangle is illustrative rather than sufficient for professional reserving. The 24-to-36 factor relies on one observation, and the tail is judgmental. A real analysis would examine larger datasets, diagnostics, alternative methods, uncertainty and changes in claims practice.

4. Operational and Customer Performance

4.1 Claim duration

Claim duration may be measured from report to closure, occurrence to report, report to first payment or report to settlement. The definition must identify whether reopened claims and open claims are included.

For n closed claims with durations d_i in days, mean duration is:

Suppose eight closed claims have durations of 6, 8, 10, 12, 14, 18, 24 and 28 days. Their mean is 120 ÷ 8 = 15 days, while the median is (12 + 14) ÷ 2 = 13 days. Five of eight claims close within 15 days, producing a service-level achievement rate of 62.5%. Against a target of 80%, the gap is 17.5 percentage points.

Closed-claim statistics omit claims still open and are subject to right-censoring. Survival analysis, including Kaplan–Meier estimates or duration models, is preferable when open claims are material.

4.2 Claims leakage

Leakage is avoidable claim cost arising from errors, weak controls or missed opportunities rather than the legitimate cost of covered losses. Sources include duplicate payments, incorrect deductibles, missed recoveries, poor vendor control, coverage errors and preventable overpayments.

An audit-based leakage rate is:

Suppose a representative audit identifies CU24,000 of preventable payments in CU800,000 of indemnity. Estimated leakage is 3%. If the comparable annual portfolio has CU12 million of indemnity, a simple ratio projection gives CU360,000 of leakage. A 2% target corresponds to CU240,000, so the estimated excess over target is CU120,000.

Projection is defensible only if the audit sample represents the portfolio. Confidence intervals, stratified sampling and separate treatment of recoverable amounts improve reliability. Investigators must not label legitimate high claim costs as leakage merely because they are undesirable.

4.3 Productivity

Raw closure counts can reward premature settlement and disadvantage staff handling complex claims. A basic measure is:

A six-person team working 22 days provides 132 FTE-days. If it closes 660 claims, productivity is 5 closures per FTE-day. A target of 5.5 requires 726 closures, leaving a gap of 66 claims, assuming unchanged staffing and case complexity.

A complexity-adjusted measure assigns validated weights. Suppose the 660 closures comprise 300 simple claims weighted 1, 240 standard claims weighted 2 and 120 complex claims weighted 4. Total output is 300 × 1 + 240 × 2 + 120 × 4 = 1,260 complexity points. At 7.5 hours per FTE-day, available time is 990 hours, giving 1.273 points per hour. This exceeds a target of 1.2 points per hour. Weights should be empirically reviewed because arbitrary weights can distort incentives.

4.4 Customer outcomes

Useful customer measures include satisfaction, complaint rates, first-contact resolution, communication timeliness, fair-treatment indicators and Net Promoter Score. A satisfaction rate is:

If 134 of 168 respondents select the defined satisfied categories, satisfaction is 79.8%. An 82% target requires at least 138 satisfied respondents at the same sample size because 0.82 × 168 = 137.76 and people cannot be fractional. Four additional satisfied responses would be required.

Response bias must be considered: survey respondents may differ systematically from nonrespondents. Cost reductions should also be assessed for their effect on vulnerable customers, complaints and settlement fairness.

5. Dashboard Design and Segmentation

5.1 Dashboard architecture

Dashboards should operate at three levels:

Each metric tile should show the current result, target, prior period, trend, scope and valuation date. Traffic-light status should be based on approved thresholds rather than subjective colouring. For example, a duration target might be green at no more than 14 days, amber above 14 through 17, and red above 17 days.

Targets may be derived from pricing assumptions, budgets, risk appetite, service commitments, regulation or statistically stable historical performance. A target without an owner or response protocol is merely descriptive.

5.2 Segmentation

Segmentation reveals heterogeneous risk and operational performance. Common dimensions include:

• product and coverage;
• territory;
• distribution channel;
• customer type;
• sum insured or policy limit;
• peril or cause of loss;
• underwriting score band;
• claim complexity;
• repairer, provider or vendor; and
• customer vulnerability indicator.

Fine segmentation creates volatility when sample sizes are small. Credibility methods, confidence intervals and minimum-volume rules can prevent overreaction. Segments should also be tested for indirect discrimination and inappropriate use of protected or sensitive characteristics.

5.3 Control and comparison

A performance change should be assessed against expected random variation. For a claim frequency estimate based on count N and exposure E, a Poisson approximation may be appropriate when events are independent and the expected rate is stable. Insurance data often exhibit overdispersion, clustering and catastrophe dependence, so the Poisson assumption must be tested rather than presumed.

Comparisons should distinguish:

• actual versus target;
• actual versus prior period;
• actual versus model expectation;
• observed versus development-adjusted result; and
• statistically unusual versus commercially important variation.

6. Predictive Models for Underwriting and Claims

6.1 Frequency and severity models

Claim counts are commonly modelled using Poisson or negative-binomial generalized linear models. A Poisson model with exposure offset may be written as:

Here, N_i is the claim count, e_i is exposure, x_ij are predictor values, and β_j are estimated coefficients. The Poisson model assumes conditional equality of count mean and variance. A negative-binomial model is often preferred when conditional variance exceeds the mean.

Positive claim severity may be modelled with gamma or lognormal distributions. Separate frequency and severity models improve interpretability, although aggregate-loss and machine-learning models may also be useful.

6.2 Claims applications

Predictive claims applications include:

• triaging claims by expected complexity;
• estimating settlement duration;
• predicting litigation or escalation;
• detecting anomalous or potentially fraudulent claims;
• estimating recovery potential;
• supporting case-reserve review; and
• identifying customers requiring proactive communication.

Survival models are appropriate for time-to-event outcomes because they account for censoring. Classification models used for fraud do not establish guilt; they prioritize review. Human investigation and procedural fairness remain necessary.

6.3 Model evaluation

Model quality has several dimensions:

• Discrimination: ability to rank higher-risk observations above lower-risk ones.
• Calibration: agreement between predicted and observed outcomes.
• Stability: consistency across time and populations.
• Lift: improvement in a targeted group relative to a baseline.
• Economic value: benefit after intervention costs and unintended consequences.
• Fairness: absence of unjustified discriminatory outcomes.

Validation should use data not employed in model fitting, preferably including out-of-time testing. Data leakage occurs when predictors contain information unavailable at the decision date. Model drift occurs when relationships change after deployment.

Predictions are not causal effects. A model may identify claims associated with long duration without proving that changing a predictor will shorten settlement time. Intervention decisions require operational testing or credible causal evidence.

7. Scenario and Sensitivity Analysis

Scenario analysis estimates outcomes under coherent sets of assumptions. Sensitivity analysis changes one variable at a time to reveal which assumptions matter most. Neither method assigns a probability unless probabilities are separately estimated.

If frequency changes by proportion g_f and severity changes by g_s, expected losses change by:

This multiplicative relationship follows from losses being frequency times severity, assuming exposure is unchanged and the scenario changes are applicable to the same portfolio.

For base ultimate losses of CU78 million, an 8% frequency increase and a 12% severity increase produce:

With unchanged earned premium of CU120 million, the projected loss ratio is:

The loss increase is 20.96%, not 20%, because the frequency and severity effects compound. More advanced scenarios may incorporate exposure growth, rate changes, reinsurance, inflation, catastrophes, behavioural response and reserve development.

Stress tests should be severe yet plausible and include reverse stress testing: identifying conditions under which a limit, capital threshold or profitability objective would be breached.

8. Illustrative Case: Meridian General Insurance

Background

Meridian General Insurance is a fictional motor insurer. Management observes deteriorating profitability despite stable policy growth. Customer complaints about slow settlements have also increased.

Problem

For the latest accident year, earned premium is CU120 million and selected ultimate losses are CU78 million, giving a 65% ultimate loss ratio against a 62% target. Paid losses are CU50 million. The company holds a claim reserve of CU26 million, while the actuarially indicated reserve is CU78m − CU50m = CU28m. Median settlement duration is 18 days against a 14-day target, audited leakage is 3% against 2%, and satisfaction is 79.8% against 82%.

Analysis

Reserve adequacy is:

The reserve is therefore CU2 million below the selected indication. Segment analysis shows that commercial delivery vehicles have both higher frequency and greater repair-cost inflation than private vehicles. Vendor analysis identifies missed salvage recoveries and duplicate repair items. Workflow data show a queue at technical assessment, while customer surveys associate poor satisfaction with communication delays rather than settlement amount alone.

A stress scenario combining an 8% frequency increase and 12% severity increase produces CU94.3488 million of ultimate losses and a 78.624% loss ratio if premium remains CU120 million.

Approach and Solution

Meridian adopts a coordinated response:

1. Strengthen rates and underwriting rules for the poorly performing delivery segment, subject to regulatory and fairness review.
2. Increase the held reserve to reflect the approved actuarial indication, while documenting uncertainty and tail assumptions.
3. Introduce automated duplicate-item checks and recovery prompts, with human review of exceptions.
4. Reallocate technical assessors according to queue volume and complexity-adjusted workload.
5. Trigger customer updates when claims pass defined inactivity thresholds.
6. Monitor segment calibration, leakage, duration, complaints and satisfaction together to prevent narrow cost targets from harming customers.

Outcome

After two quarters, assume leakage falls from 3.0% to 2.1%, median duration falls from 18 to 14 days, and satisfaction rises from 79.8% to 83.0%. These results would be treated as encouraging but not automatically causal. Management should compare exposed and unexposed groups, adjust for claim mix and examine whether the improvement persists.

Learning Points

The case demonstrates that profitability, reserving, operations and customer experience cannot be managed independently. Segment analysis identifies where deterioration originates, scenario analysis quantifies downside risk, and balanced targets reduce incentives to delay, deny or underpay valid claims.

9. Management Reporting

A management report should answer five questions:

1. What happened?
2. Why did it happen?
3. What may happen next?
4. What decision is required?
5. Who will act, and by when?

A concise report normally contains:

• an executive summary;
• key performance indicators with targets and trends;
• frequency, severity, rate and mix decomposition;
• reserve-development commentary;
• segment and exception analysis;
• scenario results and uncertainty;
• customer and conduct impacts;
• recommended actions, owners and deadlines; and
• an appendix containing definitions and methodology.

Reports should distinguish facts, estimates, assumptions and recommendations. False precision should be avoided. Reserve estimates, model predictions and scenario outputs should include ranges or sensitivity commentary where material.

10. Governance, Ethics and Regulation

Performance analytics relies on personal, financial and sometimes health-related data. Access should follow purpose limitation, data minimization, security and retention requirements. Regulatory obligations vary by jurisdiction but generally require accurate reporting, adequate claims liabilities, fair customer treatment and defensible automated decisions.

Important governance controls include:

• approved data definitions and lineage;
• reconciliation to financial and policy systems;
• independent reserve and model review;
• version control and reproducibility;
• documented overrides;
• bias and fairness testing;
• role-based access controls;
• incident and model-drift monitoring; and
• board-level accountability for material models.

Metrics can create harmful incentives. Closure targets may encourage premature settlements, low average severity targets may encourage underpayment, and fraud scores may disproportionately burden certain groups. Balanced scorecards should therefore combine financial, operational, conduct and customer measures.

Contemporary systems increasingly use artificial intelligence, natural-language processing and image analytics to summarize documents, detect anomalies and estimate damage. These tools can improve consistency and speed, but hallucination, opacity, privacy risk and automation bias require human oversight. Material claim denials or adverse underwriting decisions should not rely on unverified model output.

11. Implementation Challenges and Best Practices

Common challenges include inconsistent claim identifiers, missing exposure data, immature loss periods, changing case-reserving practices, inflation, small segment volumes and weak integration between underwriting and claims systems.

Effective implementation follows several principles:

1. Begin with business decisions rather than available charts.
2. Maintain a controlled metric dictionary.
3. Reconcile dashboard totals to authoritative systems.
4. Display valuation dates and development maturity.
5. Separate paid, reported incurred and ultimate results.
6. Include explicit tail assumptions in reserve development.
7. Combine financial and customer measures.
8. Validate predictive models before and after deployment.
9. Use scenario ranges rather than a single unsupported forecast.
10. Assign every corrective action an owner, deadline and follow-up measure.

Chapter Summary

Underwriting and claims analytics converts exposure, premium, loss, reserve, operational and customer data into coordinated management decisions. Frequency and severity explain expected loss cost, while loss ratios connect claims outcomes to earned premium. Paid, reported incurred and ultimate losses must be clearly distinguished, particularly regarding IBNR and IBNER.

Reserve analysis requires development factors, explicit tail assumptions and recognition of estimation uncertainty. Operational performance should be assessed through duration, leakage and complexity-adjusted productivity, while customer satisfaction and conduct measures guard against harmful cost-control incentives. Segmentation, predictive models and scenarios deepen analysis but introduce credibility, drift, fairness and governance risks. Effective dashboards and management reports connect defined metrics to targets, explanations, decisions, owners and deadlines.

Key Terms

Claim Frequency: Number of claims relative to earned exposure.

Claim Severity: Average loss amount per claim under a specified loss basis.

Pure Premium: Expected claim cost per exposure unit.

Earned Premium: Premium attributable to the expired portion of insurance coverage.

Loss Ratio: Losses divided by earned premium on a specified loss basis.

Reported Incurred Loss: Paid loss plus case reserves at a valuation date.

Ultimate Loss: Estimated total cost after all claim development, including IBNR and IBNER as defined.

IBNR: Provision for claims incurred but not yet reported; usage may sometimes include broader development provisions.

IBNER: Provision for future development beyond current estimates on reported claims.

Tail Factor: Development factor from the final observed maturity to ultimate settlement.

Reserve Adequacy: Comparison of held reserves with an indicated reserve estimate on a consistent basis.

Claims Leakage: Avoidable claim cost caused by error, weak control or missed recovery.

Right-Censoring: A condition in which the final duration of an open claim is not yet observed.

Segmentation: Division of a portfolio into analytically meaningful groups.

Calibration: Agreement between predicted and observed outcomes.

Model Drift: Deterioration caused by changes in data or relationships after deployment.

Scenario Analysis: Evaluation of outcomes under a coherent set of changed assumptions.

Sensitivity Analysis: Evaluation of the effect of changing one assumption at a time.

Data Lineage: Documented movement and transformation of data from source to reported result.

Balanced Scorecard: Performance system combining financial, operational, customer and control measures.

Review Questions

Short Answer Questions

1. Distinguish claim frequency from claim severity.
2. What is the difference between reported incurred loss and estimated ultimate loss?
3. Why must a tail factor be considered in loss-development analysis?
4. Define claims leakage and identify two possible sources.
5. Why can closed-claim duration statistics be biased?

Long Answer Questions

1. Explain how underwriting and claims metrics should be integrated in a performance dashboard.
2. Discuss the assumptions, applications and limitations of the chain-ladder method.
3. Evaluate the benefits and risks of predictive models in claims triage and fraud detection.
4. Explain how segmentation can improve management decisions while creating credibility and fairness risks.
5. Design the structure of a monthly management report for a motor insurance portfolio.

Analytical / Application Questions

1. A portfolio has 18,000 policy-years, 450 claims, ultimate losses of CU9 million and earned premium of CU12 million. Calculate frequency, average severity, pure premium and ultimate loss ratio, and interpret the results.
2. A claims team closes 840 claims using 140 FTE-days. Its target is 6.5 closures per FTE-day. Calculate actual productivity and the additional closures required to meet the target.
3. Base losses are CU50 million. Construct scenarios for a 5% frequency increase, a 10% severity increase, and both changes together. Explain why the combined effect is not 15%.

Multiple Choice Questions with Answers

Question 1

A motor portfolio records 240 claims from 12,000 vehicle-years. What is claim frequency?

A. 0.2%

B. 2.0%

C. 5.0%

D. 20.0%

Correct Answer: B. 2.0%

Question 2

Ultimate losses are CU3.6 million for 240 claims. What is average ultimate severity?

A. CU1,500

B. CU8,640

C. CU15,000

D. CU24,000

Correct Answer: C. CU15,000

Question 3

Which amount is normally included in estimated ultimate loss but excluded from reported incurred loss defined as paid loss plus case reserves?

A. Earned premium

B. IBNR and IBNER

C. Written exposure

D. Acquisition expense

Correct Answer: B. IBNR and IBNER

Question 4

Paid losses at 24 months are CU1,080,000. The 24-to-36 factor is 1.10 and the 36-month-to-ultimate tail factor is 1.05. What is the indicated ultimate loss?

A. CU1,134,000

B. CU1,188,000

C. CU1,247,400

D. CU1,296,000

Correct Answer: C. CU1,247,400

Question 5

Eight claim durations are 6, 8, 10, 12, 14, 18, 24 and 28 days. What proportion closes within 15 days?

A. 50.0%

B. 62.5%

C. 75.0%

D. 80.0%

Correct Answer: B. 62.5%

Question 6

An audit finds CU24,000 of preventable payments in CU800,000 of audited indemnity. What is the estimated leakage rate?

A. 2%

B. 3%

C. 8%

D. 30%

Correct Answer: B. 3%

Question 7

A team closes 660 claims using 132 FTE-days. What is its productivity?

A. 4 closures per FTE-day

B. 5 closures per FTE-day

C. 6 closures per FTE-day

D. 7 closures per FTE-day

Correct Answer: B. 5 closures per FTE-day

Question 8

Base losses are CU78 million. Frequency increases by 8% and severity by 12%, with exposure unchanged. What are projected losses?

A. CU93.60 million

B. CU94.35 million

C. CU95.68 million

D. CU98.00 million

Correct Answer: B. CU94.35 million

Question 9

Which model-evaluation property measures agreement between predicted probabilities and observed outcomes?

A. Calibration

B. Segmentation

C. Censoring

D. Development

Correct Answer: A. Calibration

Question 10

Which dashboard practice best reduces the risk of misleading reserve reporting?

A. Displaying only paid losses

B. Treating 36-month losses as ultimate in every portfolio

C. Stating valuation dates, development maturity and explicit tail assumptions

D. Removing actuarial estimates from management reports

Correct Answer: C. Stating valuation dates, development maturity and explicit tail assumptions