Pricing, Underwriting and Risk Selection

Introduction and context

Insurance pricing converts uncertain future losses into premiums that are adequate, competitive, equitable and commercially viable. Underwriting determines whether a proposed risk should be accepted, modified, referred or declined. Risk selection concerns the composition of the insured portfolio: which risks enter, at what terms, and in what proportions.

Predictive analytics has transformed these activities from primarily expert-rule processes into evidence-informed decision systems. Generalized linear models, gradient boosting, survival models, credibility methods and machine-learning algorithms can identify heterogeneous risk patterns that are difficult to detect through traditional rating factors alone. Nevertheless, prediction is not identical to pricing. A model estimates expected outcomes; pricing must additionally consider expenses, capital, reinsurance, profit, competition, demand elasticity, regulation, fairness and operational constraints.

The central managerial challenge is therefore integration. A technically accurate model can still produce an unsuitable rate if the data are biased, the exposure base is unstable, the price is unaffordable, the model is unlawful, or the resulting portfolio is excessively concentrated.

Learning objectives

After completing this chapter, the reader should be able to:

1. distinguish risk classification, underwriting, technical pricing and portfolio selection;
2. construct a technical premium from frequency, severity, expenses, capital and profit considerations;
3. interpret predictive models used for pricing and underwriting triage;
4. apply actuarial credibility to combine internal and external experience;
5. evaluate competition and price elasticity in rate decisions;
6. design governance controls for fairness, explainability, validation and monitoring; and
7. assess portfolio-level consequences of individual acceptance and pricing decisions.

1. The decision architecture of insurance

A typical insurance decision has four linked layers.

The distinction between expected loss and premium is fundamental. If a policy has claim count N and claim severities X₁, X₂, ..., X_N, aggregate loss is S = ∑_i=1^N X_i. The pure premium, or loss cost per exposure, is

where E_x is the exposure measure and E[·] denotes expectation. Under a frequency–severity decomposition, assuming the exposure basis is consistent,

where λ = E[N]/E[E_x] is expected claim frequency per exposure and μ = E[X] is expected severity per claim. The decomposition is useful for diagnosis, although frequency and severity may be correlated and should not automatically be modelled independently.

A commercial premium may be represented as:

where P is premium, F is fixed expense per policy, V(P) is variable expense expressed as a function of premium, C is the cost of allocated capital, R is risk or reinsurance cost, and Π is target profit or contingency margin. If variable expenses are a constant proportion v of premium, then

This formula assumes the components are measured on the same time, policy and exposure basis, and that variable expenses are proportional to premium. It should not be used when commissions, taxes or acquisition costs have nonlinear schedules without suitable adjustment.

2. Risk classification and predictive modelling

Risk classification seeks to group policyholders with sufficiently similar expected outcomes while avoiding classifications that are prohibited, unfairly discriminatory or operationally impractical. Common predictors include exposure, vehicle or property characteristics, prior claims, geography, coverage structure, policy tenure and risk-mitigation actions. A predictor should have a defensible relationship to the insured peril, be measured reliably at the point of decision, and be subject to legal and fairness review.

Generalized linear models

For claim frequency, a Poisson or negative-binomial model is often used. A basic frequency model is:

where N_i is claims for policy i, e_i is exposure, x_ij are predictors, and β_j are estimated coefficients. The term log(e_i) is an offset: it accounts for differing exposure without estimating an ordinary coefficient for exposure. The Poisson assumption implies conditional variance approximately equals conditional mean. If variance materially exceeds the mean, a negative-binomial model or another approach may be more appropriate.

For positive claim severity, a Gamma GLM with a log link is common:

The combined indicated loss cost is the product of predicted frequency and predicted severity when the models and exposure definitions are compatible. GLMs are attractive because coefficients can be interpreted through multiplicative relativities. A one-unit increase in x_j changes the expected mean by a factor of e^β_j, holding other variables constant.

Machine learning and hybrid models

Tree ensembles, gradient boosting and neural networks may improve predictive accuracy by modelling nonlinearities and interactions. However, their value must be assessed using out-of-sample performance, stability, calibration and business consequences rather than a single accuracy statistic. A practical architecture often combines:

• a transparent actuarial model for the base rate;
• machine learning for residual risk or interaction discovery;
• constrained outputs that preserve monotonicity, regulatory limits and pricing logic; and
• human review for exceptions and high-severity cases.

The model should predict the target that is relevant to the decision. A claims model trained on reported claims may learn reporting behaviour, settlement practices or historical underwriting rules rather than underlying risk. Target leakage occurs when a variable contains information unavailable at quotation or acceptance. Random train–test splits can also overstate performance when policies from the same customer, household or accident period appear in both samples. Temporal validation is usually more realistic for pricing.

Performance measures

Discrimination measures whether higher-risk observations tend to receive higher predictions. Calibration measures whether predicted aggregate outcomes agree with observed outcomes. For a portfolio divided into groups g, a simple calibration ratio is:

A value near 1 indicates aggregate calibration, but does not prove individual accuracy. Other measures include deviance, mean absolute error, root mean square error, lift, Gini or concordance. Metrics should be calculated by product, territory, channel, renewal status and relevant protected or vulnerable groups where lawful and ethically appropriate.

3. Technical pricing and actuarial credibility

Loss development and trend

Historical claims are rarely directly comparable with future claims. Actuaries adjust for changes in exposure, claim development, inflation, legal conditions, benefit design and mix of business. The projected loss cost can be expressed conceptually as:

where the factors represent exposure, loss-cost trend and portfolio-mix adjustments. This multiplicative representation is a framework rather than a universal law; components must be estimated consistently and interactions considered.

Credibility

Internal experience may be highly relevant but statistically volatile, particularly for small books or low-frequency risks. Credibility combines an individual risk or class estimate with a broader complement:

where Z lies between 0 and 1. Under limited-fluctuation credibility, a common approximation is

where n is observed exposure or expected claims and k is the full-credibility standard. The exact value of k depends on the selected confidence level, allowable error and assumed claim distribution. Thus, the formula is not universal; it is a simplified square-root credibility rule.

Worked example: indicated premium

Suppose a commercial motor segment has 10,000 vehicle-years. Expected frequency is 0.08 claims per vehicle-year and expected severity is $6,000 per claim. The loss cost is:

L = 0.08 × $6,000 = $480 per vehicle-year.

Assume fixed expense is $75, capital and reinsurance cost is $35, target profit is $30, and variable expenses equal 18% of premium. Then:

P = ($480 + $75 + $35 + $30)/(1 − 0.18)

P = $620/0.82 = $756.10 approximately.

The result is an indicated premium before considering competitive constraints, taxes not included in the stated assumptions, policy-specific modifications or affordability effects. If the company charges $700, the premium is below this indicated level under the assumptions; management may nevertheless accept the deviation if there is credible evidence of lower future loss, strategic value or a planned corrective action.

4. Underwriting triage and acceptance decisions

Underwriting triage allocates submissions to decision paths according to expected complexity, risk and authority. A simple system may include:

1. straight-through acceptance for low-risk, complete submissions;
2. automated verification for moderate-risk submissions;
3. referral to an underwriter for unusual features, high limits or conflicting evidence; and
4. decline or restricted terms where the risk violates appetite, capacity or law.

A triage model should estimate more than expected claims. Relevant outputs may include probability of a large loss, likelihood of missing information, fraud-alert priority, expected profitability and capital consumption. The model should not silently convert a statistical association into an automatic decline rule.

A useful decision rule compares risk-adjusted value with required resources. In simplified form, accept a risk when:

This is a managerial screening criterion, not a complete solvency test. Correlation with existing portfolio losses, operational capacity, liquidity and legal obligations must also be considered.

Human oversight is particularly important where data are sparse, the risk is novel, the model is outside its training range, or the decision has substantial consequences. Override procedures should require a reason code, preserve an audit trail and be monitored for systematic bias.

5. Competition, elasticity and strategic pricing

Technical adequacy does not determine the charged price in isolation. Competitors, distribution arrangements, customer lifetime value and demand response affect the commercial decision. Price elasticity of demand is:

If demand falls by 6% following a 4% price increase, ε = −6%/4% = −1.5. The negative sign indicates the usual inverse relationship. Elasticity estimates are difficult because observed price changes may be targeted at different risks, and retention is affected by service, brand, payment terms and switching costs.

A simplified retention model might estimate:

where logistic(t) = 1/(1 + e^−t). The model estimates probability, not certainty, and requires careful treatment of confounding and temporal drift.

Strategic pricing should distinguish:

• technical price: expected cost plus required margins;
• market price: the price observable in the competitive environment;
• charged price: the actual price after discounts, fees and adjustments; and
• value price: the price justified by customer-perceived benefits.

Persistent underpricing can create adverse selection, where relatively high-risk customers are more willing to buy, and can erode capital. Persistent overpricing can produce profitable retention in the short run but invite regulatory intervention, reputational damage and selective lapse by lower-risk customers.

6. Portfolio optimization and capital

Individual policies contribute to a portfolio rather than existing independently. The aggregate loss of a portfolio is:

If risks are correlated, portfolio variance is:

The covariance term explains why concentration in the same flood zone, catastrophe peril, industry or supply chain can consume disproportionate capital. A policy that appears profitable on a standalone basis may be undesirable when it increases tail dependence.

Portfolio optimization can use constraints such as geographic limits, maximum industry concentration, catastrophe exceedance probability, reinsurance capacity and return on capital. A simplified return-on-capital measure is:

Allocated capital is method-dependent and may be based on marginal capital, covariance allocation, stress testing or a regulatory framework. ROC should therefore be used with clear methodology rather than as an apparently objective number.

Scenario analysis is essential. Examples include a severe storm, inflation shock, litigation change, cyber accumulation event or rapid change in repair costs. Predictive models trained on ordinary historical observations may not represent such stresses adequately.

7. Fairness, regulation and model governance

Insurance differentiation is not automatically unfair: risk classification is central to insurance. The challenge is distinguishing legitimate risk information from prohibited discrimination, proxy discrimination and practices that produce unjustified disparate effects. Legal requirements differ by jurisdiction and product, so governance must involve legal and compliance specialists.

Important controls include:

• documented purpose and lawful basis for each variable;
• prohibition or restriction of protected attributes where required;
• proxy analysis for variables strongly associated with protected characteristics;
• fairness testing across relevant groups;
• reasoned explanations for adverse decisions;
• human review and appeal channels;
• data minimization, security and retention controls;
• independent validation before deployment; and
• periodic monitoring for drift and outcome disparities.

Fairness metrics can conflict. For example, equal error rates, equal selection rates and group calibration cannot generally all be guaranteed when base rates differ. A governance committee must therefore specify the decision objective, legal standard and acceptable trade-offs. Removing a protected variable does not remove discrimination if correlated variables reproduce its effect.

Model risk management should cover conceptual soundness, data quality, implementation correctness, performance, stability, explainability and change control. Every production model requires an owner, version history, approval record, monitoring thresholds and a retirement or remediation process.

8. Illustrative case: commercial property renewal

Background

A fictional insurer writes small commercial property policies. Its book has grown rapidly in coastal regions. The existing rating plan uses occupancy, construction type, insured value and broad territory. Renewal retention is high, but catastrophe losses have increased and capital consumption is approaching the internal limit.

Problem

Management must decide whether to introduce a predictive catastrophe-risk score, adjust rates, refer selected risks and alter the portfolio mix. The score is correlated with coastal location and building characteristics, while some geospatial variables may act as proxies for socioeconomic conditions.

Analysis

The actuarial team first reconciles policy, exposure and claims data, removes post-loss variables, and validates geocoding. It performs temporal validation using earlier years for training and the most recent year for testing. Performance is assessed by calibration, lift and catastrophe-stress scenarios rather than accuracy alone.

The team finds that the model ranks risks effectively, but calibration deteriorates for older masonry buildings because repairs are under-recorded. It also finds that the proposed score produces materially different referral rates across some neighbourhood groups. The result is not automatically proof of unlawful discrimination, but it requires investigation, documentation and possible redesign.

Approach and solution

The insurer adopts a constrained model with transparent factors, monotonic treatment of distance to hazard where scientifically justified, and a cap on automated price movement. High-hazard risks are referred rather than automatically declined. Underwriters may request verified mitigation evidence, such as roof replacement or flood barriers.

The pricing indication combines predicted loss cost, expense and capital requirements. A separate portfolio model imposes regional concentration limits and tests a severe catastrophe scenario. Renewal communications explain the principal rating factors and provide an appeal process. The insurer monitors retention, claims, referral outcomes and group-level effects quarterly.

Outcome

The revised approach reduces concentration in the most exposed zones without eliminating all high-risk customers. Some policies receive higher prices, while verified mitigation produces lower risk classifications. Model monitoring identifies drift after a building-code change, leading to recalibration rather than silent continuation.

Learning points

First, predictive accuracy does not remove the need for underwriting judgement. Second, portfolio constraints can change the value of an individual policy. Third, fairness analysis must examine variables, outcomes and process, including access to mitigation and appeal. Finally, implementation quality and monitoring are as important as model selection.

9. Contemporary developments and best practices

Insurance organizations increasingly use real-time data, telematics, connected-home sensors, geospatial imagery and automated document processing. These technologies can improve risk measurement but introduce consent, cybersecurity, availability and explainability concerns. Real-time pricing may also intensify customer sensitivity to surveillance and create unequal access where safer behaviour is easier for affluent customers to demonstrate.

Best practice is to:

1. define the decision and unit of exposure before selecting a model;
2. separate prediction, pricing and acceptance rules;
3. use temporal, leakage-resistant validation;
4. calibrate predictions and test stability across segments;
5. combine credibility with model output where experience is sparse;
6. include expenses, capital, reinsurance and catastrophe accumulation;
7. evaluate elasticity and competitor effects before implementation;
8. apply fairness, privacy and explainability controls;
9. preserve human review for exceptional or high-impact cases; and
10. monitor outcomes continuously, with documented triggers for recalibration.

Chapter Summary

Pricing, underwriting and risk selection are related but distinct insurance decisions. Predictive models estimate frequency, severity, retention, fraud or other outcomes; they do not by themselves determine a responsible premium or acceptance decision. Technical pricing begins with expected loss and incorporates expenses, capital, reinsurance, profit and relevant uncertainty. Credibility combines volatile internal experience with a suitable complement.

Underwriting triage should allocate cases according to complexity and risk while preserving human oversight for novel, high-impact or poorly represented cases. Competitive pricing requires analysis of elasticity, retention and adverse selection. Portfolio decisions must account for correlation, concentration, catastrophe exposure and capital. Finally, fairness, privacy, regulation, explainability and model governance are integral to professional practice rather than optional additions.

Key Terms

Risk classification: Grouping risks according to relevant differences in expected insurance outcomes.

Technical price: Premium indication based on expected costs, risk margins, capital and related financial requirements.

Pure premium: Expected loss cost per unit of exposure.

Frequency: Expected number of claims per exposure unit.

Severity: Expected cost per claim.

Exposure: Measure of the quantity of risk insured, such as vehicle-years or policy-years.

Credibility: Statistical weight assigned to observed experience relative to a broader complement.

Underwriting triage: Allocation of submissions to automated, verified, referred or declined decision paths.

Risk appetite: Documented range of risks an insurer is willing and able to accept.

Adverse selection: Disproportionate purchase or retention by risks with higher expected loss.

Price elasticity: Percentage change in demand associated with a percentage change in price.

Calibration: Agreement between predicted and observed outcomes.

Discrimination: Unjustified differential treatment or impact associated with protected or proxy characteristics.

Proxy variable: Variable that indirectly represents a protected characteristic.

Target leakage: Use of information unavailable at the time the prediction would be made.

Allocated capital: Capital assigned to a policy, segment or portfolio under a stated allocation method.

Concentration risk: Excessive accumulation of correlated exposures.

Model drift: Change over time in data, relationships or prediction performance.

Override: Human alteration of an automated recommendation.

Model governance: Processes controlling model development, validation, use, monitoring and retirement.

Review Questions

Short Answer Questions

1. Distinguish technical price from charged price.
2. Why is exposure included as an offset in a frequency GLM?
3. What problem does credibility address?
4. Define target leakage and give one insurance example.
5. Why can an individually profitable policy be undesirable for a portfolio?

Long Answer Questions

1. Explain how frequency, severity, expenses, capital and profit combine in an indicated premium.
2. Compare GLMs with gradient-boosting models for insurance pricing and underwriting.
3. Discuss how competition and price elasticity should influence rate decisions.
4. Explain the principal fairness and governance risks of predictive underwriting.
5. Describe a complete validation and monitoring framework for a production pricing model.

Analytical / Application Questions

1. An insurer estimates frequency of 0.12 claims per policy-year and severity of $4,500. Fixed expense is $60, capital cost is $25, profit provision is $20 and variable expense is 20% of premium. Calculate the indicated premium, stating your assumptions.
2. A proposed renewal model improves overall lift but produces poor calibration for older buildings and much higher referral rates in one neighbourhood group. Recommend an investigation and remediation plan.
3. A regional portfolio has attractive standalone margins but is heavily concentrated in one flood zone. Explain how covariance, catastrophe scenarios, reinsurance and underwriting limits should affect the acceptance strategy.

Multiple Choice Questions with Answers

Question 1

Which expression represents a frequency–severity loss-cost decomposition?

A. L = λ × μ

B. L = λ + μ

C. L = λ/μ

D. L = μ − λ

Correct Answer: A. L = λ × μ

Question 2

In a frequency GLM, the exposure term is commonly included as:

A. A response variable

B. A logarithmic offset

C. A fairness constraint

D. A severity adjustment

Correct Answer: B. A logarithmic offset

Question 3

If fixed components total $620 and variable expenses are 18% of premium, what premium follows from the stated pricing formula?

A. $620.00

B. $731.60

C. $756.10

D. $1,000.00

Correct Answer: C. $756.10

Question 4

What is the principal purpose of credibility weighting?

A. To remove all random variation from claims

B. To combine internal experience with a complement

C. To guarantee a competitive market price

D. To replace exposure measurement

Correct Answer: B. To combine internal experience with a complement

Question 5

Which validation design is generally most realistic for a model used to price future renewals?

A. Randomly mixing all policy years

B. Using only training-set performance

C. Temporal validation using later periods as test data

D. Validating only the largest claims

Correct Answer: C. Temporal validation using later periods as test data

Question 6

A price increase of 4% causes demand to fall by 6%. What is the approximate price elasticity?

A. −1.5

B. −0.67

C. +1.5

D. +10

Correct Answer: A. −1.5

Question 7

Which factor explains why portfolio loss variance can exceed the sum of individual variances?

A. Premium discounts

B. Covariance among risks

C. Policy documentation

D. Model interpretability

Correct Answer: B. Covariance among risks

Question 8

What is target leakage?

A. A deliberate premium discount

B. A variable unavailable at decision time entering the model

C. A decline in renewal retention

D. A capital allocation method

Correct Answer: B. A variable unavailable at decision time entering the model

Question 9

Which practice is most appropriate for a high-impact automated underwriting decline?

A. No explanation or appeal mechanism

B. Automatic acceptance of all similar risks

C. Human review, reason codes and an audit trail

D. Removing all variables from the model

Correct Answer: C. Human review, reason codes and an audit trail

Question 10

Why may removing a protected attribute fail to eliminate discriminatory outcomes?

A. Models cannot use numerical data

B. Correlated proxy variables may reproduce its effect

C. Calibration always creates discrimination

D. Exposure is never relevant to pricing

Correct Answer: B. Correlated proxy variables may reproduce its effect