8. Premium Rating, Reserving and Insurance Analytics

Introduction and Context

Insurance converts uncertain future losses into a present financial obligation. The premium is the price charged for accepting risk, while reserves represent amounts set aside to meet obligations that have already arisen or are expected to arise. Premium rating and reserving are therefore central to insurer solvency, fairness, competitiveness and long-term profitability. If premiums are inadequate, an insurer may grow rapidly while accumulating losses. If premiums are excessive, customers may leave and competition may weaken the portfolio. Similarly, inadequate reserves can produce misleading profits and threaten the insurer’s ability to pay claims.

Modern insurance management combines actuarial principles, statistical analysis, underwriting judgment, financial management and technology. The objective is not simply to calculate an average loss. It is to estimate the cost of risk, reflect uncertainty, allocate expenses and capital, and make decisions that remain robust when actual experience differs from expectations.

This chapter develops the logic of premium rating, explains frequency and severity models, distinguishes technical premiums from market prices, examines credibility and experience rating, introduces the principal forms of reserves, and shows how loss ratios, combined ratios and predictive analytics support managerial decisions.

Learning Objectives

After studying this chapter, the student should be able to:

1. Explain frequency, severity, expected loss, pure premium and technical premium.

2. Distinguish risk, expense, profit, catastrophe and reinsurance loadings.

3. Apply basic aggregate-loss and loss-ratio methods.

4. Explain credibility theory and experience rating.

5. Distinguish case, incurred-but-not-reported and premium reserves.

6. Interpret loss ratios, expense ratios and combined ratios.

7. Evaluate the uses and limitations of predictive analytics in insurance.

8. Discuss governance, fairness, documentation and professional judgment in pricing and reserving.

The Architecture of an Insurance Price

A useful starting point is the aggregate loss of a portfolio. Let \(N\) denote the number of claims during a period and \(X_i\) the cost of claim \(i\). Aggregate claims are:

\[

S=X_1+X_2+...+X_N.

\]

If claim frequency is \(E[N]\) and average claim severity is \(E[X]\), then, under the usual independence assumption,

\[

E[S]=E[N]E[X].

\]

For a portfolio of exposure units, the expected loss per exposure is often called the pure premium or expected loss cost. Frequency measures how often claims occur; severity measures the size of a claim when it occurs. Their product is a powerful decomposition because different underwriting, prevention and economic factors affect each component.

For example, suppose a motor portfolio has 10,000 vehicle-years, an expected frequency of 0.08 claims per vehicle-year and an average severity of ₹75,000. Expected claims are:

\[

10,000\times0.08\times₹75,000=₹60,000,000.

\]

The expected loss cost is ₹6,000 per vehicle-year. This is not necessarily the premium charged. It excludes many costs and uncertainty adjustments.

A simplified technical premium can be expressed as:

\[

P_T = L + E_v + E_f + R + C + M,

\]

where \(L\) is expected loss, \(E_v\) variable expenses, \(E_f\) fixed expenses allocated to the policy, \(R\) risk or volatility loading, \(C\) catastrophe or reinsurance cost, and \(M\) a margin for capital and profit. Taxes, levies and commissions may be included separately depending on the pricing system.

An alternative percentage-loading formula is:

\[

P=\frac{L+F+R+M}{1-v},

\]

where \(F\) represents fixed monetary expenses and \(v\) is the proportion of premium consumed by variable expenses. This formulation is important because a variable expense expressed as a percentage of premium cannot simply be added to loss cost; it must be grossed up.

The technical premium is an actuarial indication, not an automatic market price. The final price may be influenced by competition, customer value, distribution costs, regulatory constraints, strategic objectives, capacity, reinsurance arrangements and the quality of the individual risk.

Frequency, Severity and the Distribution of Losses

Frequency data are usually counts and may be modelled using Poisson, binomial or negative binomial distributions. The Poisson model assumes that claims occur independently at a stable average rate. Real portfolios often display over-dispersion because policyholders differ in risk or claims cluster over time; the negative binomial model can then be more appropriate.

Severity is positive and often highly skewed. A small number of large claims may account for a substantial proportion of total cost. Lognormal, gamma, Weibull and Pareto distributions are commonly used in actuarial work. The choice of distribution should be guided by diagnostic testing, business knowledge and the intended use of the model rather than by statistical fit alone.

The aggregate loss distribution matters because two portfolios may have the same expected loss but very different volatility. A portfolio with many small independent claims is generally more predictable than one exposed to infrequent, extremely large losses. The latter requires greater risk capital, reinsurance or catastrophe loading.

Important adjustments include deductibles, limits, coinsurance and policy conditions. A deductible reduces the amount paid on smaller claims and changes both claim frequency and severity. A policy limit truncates the upper tail. Inflation, medical-cost escalation, wage growth, litigation trends and changes in repair technology require historical claims to be developed to a common cost basis. This is known as trending or on-level adjustment.

From Historical Experience to a Prospective Rate

Historical loss data cannot be used mechanically. A prospective rate should reflect expected future conditions. The principal adjustments include:

• changes in exposure mix, such as vehicle type, geography or occupation;
• changes in policy terms, deductibles and limits;
• claims development for losses not yet fully settled;
• economic and social inflation;
• legal and regulatory changes;
• catastrophe exposure and reinsurance costs; and
• changes in underwriting, claims handling and fraud controls.

The loss-ratio method is often useful when reliable exposure or claim-level data are unavailable. If the current premium is \(P\), the expected future loss ratio is \(LR_e\), and the permissible loss ratio is \(LR_p\), the indicated premium may be approximated by:

\[

P_{indicated}=P\times\frac{LR_e}{LR_p}.

\]

If current premium is ₹100 million, expected loss ratio is 72 per cent and permissible loss ratio is 60 per cent, the indicated premium is ₹120 million. This means a 20 per cent increase is required, assuming the underlying assumptions are sound.

The permissible loss ratio is derived from the target combined ratio and expense structure. If the target combined ratio is 95 per cent and the expense ratio is 35 per cent, the permissible loss ratio is approximately 60 per cent. The method is simple but can conceal differences among risks and may be distorted by unusual catastrophe years or premium changes.

Credibility and Experience Rating

A central pricing problem is deciding how much weight to give an individual policyholder’s experience. A large commercial account may have credible internal claims data, while a small account’s experience may be dominated by random fluctuation. Credibility theory combines individual experience with broader collective information:

\[

P_{cred}=ZP_{experience}+(1-Z)P_{manual},

\]

where \(Z\) is the credibility factor between zero and one. A higher \(Z\) is justified when the account has a large volume of homogeneous, reliable and relatively stable data.

Limited-fluctuation credibility sets the credibility factor according to the probability that observed experience will fall within an acceptable error around the true expected value. A simplified full-credibility standard may require a specified number of claims or exposure units. Partial credibility uses only a proportion of the account’s own experience and supplements it with manual rates.

Experience rating is a pricing mechanism in which the premium responds to the insured’s past losses. In workers’ compensation, fleet motor insurance and large property programmes, experience modification factors may increase or decrease the manual premium. Effective experience rating must distinguish avoidable risk from random variation. A single large fire should not automatically redefine the long-term risk of an otherwise well-managed account, although it may reveal weaknesses requiring corrective action.

Retrospective rating determines part of the premium after the policy period, based on actual losses subject to minimum and maximum limits. It provides incentives for loss prevention but transfers greater uncertainty to the insured. Such arrangements require careful treatment of claims development, collateral, data quality and moral hazard.

Technical Premiums, Market Premiums and Risk Selection

The technical premium represents the estimated economic cost of providing insurance, including an appropriate allowance for uncertainty and capital. The market premium is the amount actually charged. The difference may arise from competitive strategy, cross-subsidisation, customer retention, regulatory pricing restrictions or deliberate growth objectives.

Underpricing is not always immediately visible. A new portfolio may show favourable early results because claims have not yet emerged or developed. Conversely, a portfolio may appear unprofitable after a single catastrophe even though its long-term technical price is adequate. Management should therefore analyse results by accident year, underwriting year, product, territory, channel and risk segment.

Risk classification improves fairness when variables are relevant, measurable and legally permissible. Excessively refined segmentation may create complexity, privacy concerns or exclusion. The objective is not to charge every person a perfectly individualised price at any cost; it is to achieve a defensible balance between actuarial accuracy, solidarity, accessibility and regulation.

Reserving: Measuring Obligations After the Event

Premium rating looks forward from exposure to expected claims. Reserving looks backward from reported and unreported events to the insurer’s remaining obligation. A claim reserve is the expected future payment for a claim that has occurred but is not fully settled. The total reserve should also reflect claims incurred but not reported (IBNR), claims reported but not adequately valued, reopened claims, adjustment expenses and, where appropriate, claims handling costs.

A case reserve is established for an individual reported claim using information available to the claims handler. IBNR includes claims that have occurred but have not yet been reported and development on reported claims that are currently understated. Some actuarial systems use the broader term incurred but not enough reported (IBNER) for adverse development on known claims.

Unearned premium reserve relates to the portion of written premium applicable to future coverage. It is not a claims reserve. If an annual policy is written on 1 October and the reporting date is 31 December, three months of coverage have elapsed and nine months of premium remain unearned, subject to the applicable accounting basis. Premium deficiency or liability adequacy testing may be required when unearned premium is insufficient to cover expected future claims and expenses.

The Chain-Ladder Approach

The chain-ladder method projects cumulative claims development from historical development patterns. A triangle may show paid or incurred claims by accident period and development age. Development factors are calculated from older periods and applied to more recent, incomplete periods. The difference between projected ultimate claims and reported or paid claims is the reserve.

The method is attractive because it is transparent and relatively easy to implement. However, it assumes that historical development patterns remain relevant. It can be unreliable after changes in claims handling, settlement speed, policy terms, inflation, legal rules or portfolio composition.

Bornhuetter–Ferguson and Expected-Loss Methods

The Bornhuetter–Ferguson method combines an a priori expected loss ratio with the proportion of losses expected to have emerged. It is less sensitive than chain-ladder to random early experience, making it useful for immature accident years or volatile portfolios. Its weakness is dependence on the quality of the initial expected loss assumption.

An expected-loss method may be used when little credible claims experience exists. Ultimate loss is estimated as exposure multiplied by an expected loss cost or loss ratio. As experience matures, the method can be blended with actual development. In practice, actuaries often compare several methods and select a best estimate supported by diagnostics and judgment rather than treating one technique as universally correct.

Reserves should be presented with uncertainty ranges where material. A point estimate alone may conceal the possibility of adverse development. Management needs to understand not only the central estimate but also reserve risk, tail risk, discounting assumptions and the sensitivity of results to inflation, settlement patterns and large claims.

Performance Measures for Insurance Decisions

The loss ratio is:

\[

Loss\ Ratio=\frac{Incurred\ Losses}{Earned\ Premium}.

\]

It measures the proportion of earned premium consumed by claims. The expense ratio is generally underwriting expenses divided by written or earned premium, depending on the reporting convention. The combined ratio is:

\[

Combined\ Ratio=Loss\ Ratio+Expense\ Ratio.

\]

A combined ratio below 100 per cent indicates underwriting profit before investment income and certain other items; a ratio above 100 per cent indicates underwriting loss on that basis. The ratio must be interpreted carefully. It may exclude catastrophe volatility, reserve releases, reinsurance effects or changes in earned premium. Investment income can make an insurer profitable despite an underwriting loss, but reliance on investment returns may be inappropriate when interest rates, duration or liquidity conditions change.

Suppose earned premium is ₹500 million, incurred losses are ₹325 million and underwriting expenses are ₹150 million. The loss ratio is 65 per cent, the expense ratio is 30 per cent and the combined ratio is 95 per cent. If ₹20 million of reserve releases improved incurred losses, the underlying loss ratio would be higher. Analysts should therefore separate current accident-year performance from reserve movements relating to prior years.

Other useful measures include frequency, average severity, retention ratio, renewal rate, claim settlement time, expense per policy, return on capital, rate adequacy and loss ratio by segment. No single metric is sufficient. A low loss ratio may result from restrictive coverage, declining exposure quality or delayed reporting rather than superior underwriting.

Predictive Analytics in Insurance

Predictive analytics uses historical data, statistical models and machine-learning methods to estimate future outcomes such as claim frequency, severity, lapse, fraud probability or settlement duration. Generalised linear models remain valuable because their coefficients and distributional assumptions are relatively interpretable. Decision trees, random forests, gradient boosting and neural networks may capture nonlinear relationships, but their complexity creates additional governance requirements.

A typical pricing analytics workflow includes:

1. defining the business question and target variable;

2. assembling policy, exposure, claims, external and contextual data;

3. cleaning records and resolving missing, duplicate and inconsistent values;

4. separating training, validation and test data;

5. selecting variables and avoiding leakage;

6. fitting and validating the model;

7. assessing calibration, discrimination, stability and fairness;

8. translating predictions into rates or underwriting actions; and

9. monitoring outcomes and recalibrating the model.

A frequency-severity framework may estimate expected claim cost as the product of predicted claim frequency and predicted conditional severity. Alternative models estimate aggregate pure premium directly. The model must be validated out of sample because a model that fits historical data extremely well may perform poorly on future business.

Data leakage is a common error. A variable recorded after a claim occurs cannot legitimately be used for initial underwriting. Correlation is not causation, and a predictive variable may act as a proxy for a protected or socially sensitive characteristic. Governance should include documentation, approval authority, change control, explainability appropriate to the decision, cybersecurity and human review.

Illustrative Case: Repricing a Commercial Motor Portfolio

Background

A fictional insurer, Meridian General, insures 40,000 commercial vehicles. Its portfolio has expanded rapidly through broker channels. Management observes that the combined ratio has risen from 96 per cent to 108 per cent over two years.

Situation and Analysis

The pricing team finds an overall reported loss ratio of 74 per cent. However, claims from the latest accident year are immature. Development analysis indicates that the ultimate loss ratio is likely to be 82 per cent. Expenses are 29 per cent of premium, and catastrophe and large-loss reinsurance costs add an effective 3 per cent. The technical combined ratio is therefore approximately 114 per cent.

Segment analysis reveals that urban delivery fleets have high frequency but moderate severity, while long-distance heavy vehicles have lower frequency but substantially higher severity. The broker channel has attracted a disproportionate share of newer, less experienced operators. In addition, claims inflation for vehicle parts has exceeded the trend used in the previous rate filing.

Approach

Meridian first recalculates exposure-adjusted frequency and severity by vehicle class, territory, use and driver experience. It updates claim costs for inflation and revises development factors after reviewing claims settlement changes. A frequency-severity model is developed, but the insurer retains a transparent generalised linear model for the initial rating implementation.

The indicated technical rate is increased by segment rather than uniformly. Large fleets with credible and improving safety experience receive experience-rated terms. Smaller fleets are priced using credibility-weighted rates. Deductible options and telematics-based loss-prevention programmes are offered, but participation is voluntary and subject to data governance controls. Claims managers investigate whether faster repair networks can reduce severity without reducing service quality.

Outcome

The revised programme produces more differentiated prices, improves the alignment between premium and expected risk, and identifies loss-prevention opportunities. Management does not judge success solely by the first quarter’s combined ratio. It establishes monitoring by accident year, rate change, retention, mix, frequency, severity and customer complaints.

Learning Points

The case demonstrates that reported results can understate ultimate claims, averages can conceal segment differences, and pricing must incorporate expenses, reinsurance and claims inflation. It also shows that analytics should support—not replace—underwriting judgment, governance and communication with customers.

Risks, Limitations and Professional Governance

Pricing and reserving errors commonly arise from small or biased samples, changing exposure, inappropriate development assumptions, double-counting inflation, ignoring expenses, confusing written and earned premium, and treating a point estimate as certainty. Catastrophe events can invalidate ordinary patterns. Operational changes, such as claims automation or a new repair arrangement, may alter historical development and make past data non-comparable.

The principal ethical issue is fairness. Rating variables should have a rational relationship to risk and should comply with insurance law, anti-discrimination rules and regulatory guidance. Data obtained from connected devices, social platforms or third parties require lawful collection, purpose limitation, security and transparency. Customers should not be denied meaningful explanation merely because a model is complex.

Actuaries, underwriters, finance professionals, data scientists and senior management have distinct but connected responsibilities. Models require independent validation, version control, access controls, documentation and periodic review. Reserve committees should challenge assumptions and investigate material deviations. Stress testing should consider adverse claims inflation, catastrophe accumulation, reserve deterioration, lapse changes and reinsurance failure.

Current Developments and Best Practice

Emerging practice includes real-time or near-real-time underwriting, usage-based insurance, automated claims estimation, geospatial catastrophe analytics, climate-risk scenario analysis and advanced fraud detection. These developments can improve responsiveness and loss prevention, but they also increase dependence on data quality, model controls and technology resilience.

Sound practice can be summarised as follows:

• begin with a clearly defined exposure, coverage and decision objective;
• reconcile pricing, claims, finance and actuarial data;
• separate experience by meaningful risk and accident dimensions;
• adjust historical data for development, inflation and structural change;
• combine models with informed professional judgment;
• quantify uncertainty and test adverse scenarios;
• validate models independently and monitor drift;
• ensure legal, ethical and privacy compliance; and
• communicate assumptions, limitations and decisions clearly.

Chapter Summary

Premium rating estimates the cost of future insurance obligations and converts expected losses into a sustainable price. Frequency and severity provide the basic decomposition of expected loss, while expenses, uncertainty, catastrophe exposure, reinsurance, capital and profit produce the technical premium. Credibility and experience rating balance individual claims history against broader portfolio information.

Reserving estimates obligations arising from past exposure. Case reserves, IBNR, unearned premium reserves and premium deficiency provisions address different liabilities. Chain-ladder, Bornhuetter–Ferguson and expected-loss methods are useful but depend on data quality, stable development patterns and professional judgment. Loss ratios and combined ratios support performance analysis, but must be interpreted in relation to reserve movements, maturity, catastrophe effects and investment income.

Predictive analytics can improve segmentation, pricing, claims management and fraud detection. Its value depends on sound data, appropriate validation, fairness, explainability, security and ongoing monitoring. The central managerial lesson is that actuarial estimates are decision tools under uncertainty, not mechanically precise facts.

Key Terms

Frequency: Number of claims relative to a defined exposure base.

Severity: Monetary size of an individual claim.

Expected loss: Probability-weighted average cost of future claims.

Pure premium: Expected loss cost per exposure unit before expenses and profit loadings.

Technical premium: Premium indicated to cover expected losses, expenses, risk and capital requirements.

Loading: Amount added to expected loss for expenses, uncertainty, catastrophe, reinsurance or profit.

Credibility: Degree of reliance placed on an individual risk’s own experience.

Experience rating: Pricing based partly on the insured’s historical claims experience.

Manual rate: Base rate developed from collective portfolio experience.

Aggregate loss: Total claims cost from a portfolio during a specified period.

Case reserve: Estimated future payment for a reported individual claim.

IBNR: Claims incurred but not reported, including relevant development on known claims.

Unearned premium reserve: Premium allocated to coverage that has not yet expired.

Loss ratio: Incurred losses divided by earned premium.

Expense ratio: Underwriting expenses divided by the relevant premium base.

Combined ratio: Loss ratio plus expense ratio.

Chain-ladder method: Reserving method based on historical claims development factors.

Bornhuetter–Ferguson method: Reserving method combining expected losses with observed emergence.

Predictive analytics: Use of data and statistical or machine-learning models to estimate future outcomes.

Model drift: Deterioration in model performance as underlying conditions change.

Review Questions

Short Answer Questions

1. Distinguish claim frequency from claim severity.

2. What is meant by a technical premium?

3. Differentiate a case reserve from an IBNR reserve.

4. State the formula for the combined ratio.

5. Why is credibility important in experience rating?

Long Answer Questions

1. Explain how an insurer moves from expected loss to a final premium, including major loadings.

2. Discuss the assumptions, uses and limitations of the chain-ladder and Bornhuetter–Ferguson methods.

3. Explain how deductibles, limits, inflation and claims development affect premium rating.

4. Discuss the managerial interpretation of loss ratio, expense ratio and combined ratio.

5. Explain the governance and ethical requirements for predictive analytics in insurance.

Analytical / Application Questions

1. An insurer has 20,000 exposures, expected frequency of 0.06 and average severity of ₹50,000. Calculate expected aggregate claims and expected loss cost per exposure. Identify additional information required to determine the technical premium.

2. A portfolio has earned premium of ₹800 million, incurred losses of ₹520 million and underwriting expenses of ₹240 million. Calculate the loss ratio, expense ratio and combined ratio, and interpret the result.

3. A small commercial account has poor recent claims experience, but two claims are unusually large and the broader class remains profitable. Explain how credibility-weighted experience rating could be used and what further information should be examined before repricing the account.

22. Multiple Choice Questions with Answers

Question 1

An insurer observes an expected claim frequency of 0.05 and an average severity of ₹40,000 per exposure. What is the expected loss cost per exposure?

A. ₹2,000

B. ₹8,000

C. ₹40,050

D. ₹800,000

Correct Answer: A. ₹2,000

Question 2

Which item is normally included in a technical premium but not in pure premium?

A. Only claim frequency

B. Expected losses only

C. Expense and risk loadings

D. Historical exposure count

Correct Answer: C. Expense and risk loadings

Question 3

A credibility factor of 0.70 means that pricing should generally give what weight to the account’s own experience?

A. 0.07

B. 30 per cent

C. 70 per cent

D. 100 per cent

Correct Answer: C. 70 per cent

Question 4

Which reserve relates to coverage that remains unexpired at the reporting date?

A. Case reserve

B. Unearned premium reserve

C. IBNR reserve

D. Catastrophe reserve only

Correct Answer: B. Unearned premium reserve

Question 5

An insurer has a 68 per cent loss ratio and a 27 per cent expense ratio. What is its combined ratio?

A. 41 per cent

B. 68 per cent

C. 95 per cent

D. 105 per cent

Correct Answer: C. 95 per cent

Question 6

Why may the Bornhuetter–Ferguson method be preferred for a very immature accident year?

A. It ignores all expected losses

B. It is less sensitive to random early claims experience

C. It requires no assumptions

D. It uses only fully settled claims

Correct Answer: B. It is less sensitive to random early claims experience

Question 7

Which situation is most likely to make historical chain-ladder development factors unreliable?

A. Stable claims handling over many years

B. A major change in settlement practices

C. A large homogeneous data set

D. Consistent policy conditions

Correct Answer: B. A major change in settlement practices

Question 8

A predictive pricing model uses information recorded after a claim occurs to price the policy at inception. This is primarily an example of:

A. Credibility weighting

B. Data leakage

C. Full credibility

D. Premium earning

Correct Answer: B. Data leakage

Question 9

If current premium is ₹50 million, expected loss ratio is 72 per cent and permissible loss ratio is 60 per cent, what indicated premium follows the loss-ratio method?

A. ₹36 million

B. ₹50 million

C. ₹60 million

D. ₹72 million

Correct Answer: C. ₹60 million

Question 10

Which statement best describes the combined ratio?

A. It measures only investment income

B. It equals loss ratio plus expense ratio

C. It equals premium divided by claims

D. It is always calculated using written premium only

Correct Answer: B. It equals loss ratio plus expense ratio