Credibility, Risk Measures, Capital and Reinsurance Decisions

Introduction and context

Insurance management requires decisions under uncertainty. An insurer must estimate future claims, determine how much uncertainty it can retain, purchase reinsurance, hold sufficient capital, and evaluate whether the return earned adequately compensates for risk. These decisions are connected: credibility determines how much weight is given to an insurer’s own experience; risk measures translate loss distributions into decision indicators; capital protects policyholders and other stakeholders; and reinsurance changes both expected profit and the distribution of adverse outcomes.

This chapter introduces the principal actuarial tools used for these decisions. The emphasis is on practical interpretation as well as mathematical structure. The methods are deliberately introductory, but their limitations and governance requirements are important at postgraduate level. No single measure or model can replace professional judgement, data validation, stress testing and regulatory oversight.

Learning objectives

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

1. Explain limited-fluctuation and Bühlmann credibility.
2. Calculate and interpret Value at Risk (VaR) and Tail Value at Risk (TVaR).
3. Distinguish frequency, severity, aggregate loss and ruin concepts.
4. Assess retention and reinsurance alternatives.
5. Explain diversification, solvency capital and risk-adjusted performance.
6. Identify model, data, basis and governance risks in actuarial decisions.

1. Credibility: combining experience with a broader expectation

1.1 The credibility problem

Suppose an insurer has only a small number of observations for a new product. Its observed loss ratio may be unusually high or low because of random fluctuation. Using the observation without adjustment may produce unstable premiums and capital estimates. Conversely, ignoring the insurer’s own experience may discard valuable information.

Credibility theory combines:

• Individual experience, which is relevant but may be statistically volatile; and
• Collective or manual experience, representing a larger and usually more stable reference group.

A general credibility estimate is:

where \bar{X} is the observed experience, P₀ is the collective estimate, and Z is the credibility factor, with 0 ≤ Z ≤ 1. The estimate has the same units as \bar{X} and P₀, such as claims per policy or loss ratio.

When Z = 1, the individual experience is used completely. When Z = 0, only the collective estimate is used. Credibility is therefore a structured form of shrinkage toward a benchmark.

1.2 Limited-fluctuation credibility

Limited-fluctuation credibility determines the exposure needed for observed experience to lie within a specified tolerance of its expected value with a chosen confidence level. Let:

• N = number of independent exposure units;
• X_i = loss or claim outcome for unit i;
• μ = expected outcome per unit;
• σ² = variance per unit;
• k = reliability constant determined by the desired confidence and tolerance.

Under independence and finite variance, the variance of the sample mean is:

A normal approximation gives the approximate requirement:

where r is the allowable relative error and k is commonly associated with a two-sided confidence probability. This formula is an approximation, not a universal law. It is less reliable for heavy-tailed claims, dependent observations or very small samples.

For claim counts, a common special case assumes a Poisson distribution with mean λ per exposure. Then the variance also equals λ, and the exposure requirement can be derived accordingly. The analyst must verify whether the Poisson assumptions—independence, homogeneous risk and equality of mean and variance—are reasonable.

1.3 Bühlmann credibility

Bühlmann credibility models variation between risks and random variation within each risk. Let X denote an observation for a risk with hypothetical mean Θ. Define:

• Process variance: v(Θ) = Var(X ∣ Θ);
• Expected process variance: EPV = E[v(Θ)];
• Variance of hypothetical means: VHM = Var(E[X ∣ Θ]).

For n observations, the Bühlmann credibility factor is:

The Bühlmann estimate is:

where m = E[Θ] is the collective mean. The ratio K is the equivalent number of observations required to balance process variation against heterogeneity between risks.

Assumptions and limitations. The classical model assumes exchangeability, a stable underlying risk structure and appropriate moments. It does not automatically address inflation, changes in underwriting, claims settlement delays, catastrophes or dependence. Bühlmann–Straub credibility extends the framework to unequal exposure volumes by weighting observations according to exposure.

Worked example: credibility-weighted loss ratio

An insurer observes a loss ratio of 78% over a sufficiently large block. The collective estimate is 65%, and the credibility factor is 0.60. The credibility estimate is:

Thus, the indicated loss ratio is 72.8%. The result is neither the raw experience nor the manual estimate; it reflects both sources of information. If the observation period includes an exceptional catastrophe, the actuary should consider catastrophe normalization rather than applying credibility mechanically.

2. Aggregate loss and the need for risk measures

Let aggregate annual loss be:

where N is the random claim count and X_i are individual claim severities. This is a collective risk model. If N is independent of the severities and has mean E[N], while each severity has mean E[X], then:

If, additionally, the severities are identically distributed with variance Var(X) and N is independent of X_i, then:

The first term represents severity uncertainty conditional on claim count; the second represents frequency uncertainty. The formula is not valid without the stated independence and identical-distribution assumptions. Dependence, inflation, catastrophe clustering and policy limits may materially change the result.

Expected loss alone is inadequate for capital decisions. Two portfolios can have the same expected loss but very different probabilities of a large loss. Risk measures summarize the adverse portion of the loss distribution.

3. Value at Risk and Tail Value at Risk

3.1 Value at Risk

For a loss random variable L and confidence level α, Value at Risk is the α-quantile:

For a continuous distribution, this is the loss level exceeded with probability approximately 1 − α. For example, annual VaR at 99.5% is a threshold exceeded in approximately 0.5% of years under the model.

VaR is easy to communicate and is widely used in regulatory and internal reporting. However, it says little about the magnitude of losses beyond the threshold. It can also fail the subadditivity property: diversification may appear not to reduce VaR under some distributions.

3.2 Tail Value at Risk

For a continuous loss distribution, Tail Value at Risk is the conditional mean beyond VaR:

For distributions with discontinuities, a quantile-integral definition is safer:

TVaR is also called Expected Shortfall in much of the risk-management literature. It measures the average severity of the worst (1 − α) × 100% of outcomes. Under suitable integrability conditions, TVaR is a coherent risk measure: it satisfies monotonicity, translation invariance, positive homogeneity and subadditivity. Subadditivity supports diversification analysis, although diversification benefits remain model-dependent.

Worked example: discrete loss distribution

Consider a portfolio with the following annual losses:

At the 95% confidence level, the smallest loss with cumulative probability at least 95% is 100,000, so VaR_95% = 100,000. The worst 5% consists of probability mass 3% at 500,000 and 1% at 2,000,000, plus 1% of the 100,000 outcome needed to make 5%. Therefore:

The large difference between VaR and TVaR shows why a threshold alone may understate catastrophe exposure. For a discrete distribution, careless use of “average losses exceeding VaR” can produce a different result if the threshold has a probability mass; the quantile-integral convention should be specified.

Choice of risk measure

The selected measure should match the decision:

Risk measures must state the horizon, confidence level, valuation basis, currency, treatment of recoveries and whether they describe gross or net losses.

4. Ruin, solvency and capital

4.1 Classical risk model

In the classical compound Poisson surplus model, insurer surplus at time t is:

where u is initial surplus, c is premium income rate, N(t) is a Poisson claim count with rate λ, and X_i are claim severities. The insurer experiences ruin if surplus becomes negative:

The finite-time ruin probability is P(τ ≤ T), and ultimate ruin probability is P(τ < ∞). In the classical model, if the safety loading is positive, written as:

then the surplus has positive expected drift. This condition does not eliminate ruin; it only makes indefinite survival possible and gives a basis for a decreasing ultimate ruin probability as initial capital increases.

The adjustment coefficient R, when it exists, solves the Lundberg equation:

where M_X(R) = E[e^RX] is the moment-generating function of severity. The equation requires the moment-generating function to be finite at a positive R. Heavy-tailed distributions may not satisfy this condition, so exponential ruin bounds can be inappropriate.

4.2 Solvency capital

Solvency capital is the financial resource available to absorb adverse deviations while maintaining obligations and continuing operations. A simplified economic balance-sheet view is:

A capital requirement is commonly based on a risk measure of one-year change in own funds, subject to a specified confidence level and regulatory methodology. A simplified solvency ratio is:

The ratio is meaningful only when numerator and denominator use compatible valuation bases, time horizons and aggregation rules. Regulatory capital is a floor, not necessarily the insurer’s optimal economic capital. Internal capital may incorporate risk appetite, ratings objectives, liquidity, concentrations and strategic resilience.

4.3 Diversification and aggregation

For two loss risks L₁ and L₂, covariance is:

where ρ is correlation and σ₁, σ₂ are standard deviations. Thus:

If ρ < 1, aggregation may reduce risk relative to the sum of standalone risks. However, correlation can increase during crises, catastrophe risks may be dependent, and linear correlation does not describe tail dependence. Diversification benefits should therefore be stress-tested and not treated as guaranteed capital.

5. Retention and reinsurance decisions

5.1 Retention as a risk–return decision

A retention is the part of a loss the insurer keeps. For an occurrence loss X and per-occurrence excess-of-loss treaty with retention d, insurer net loss is:

where y^+ = max(y,0). The reinsurer’s recovery is:

For a layer with attachment a and limit l, recovery is:

These formulas apply to an occurrence layer and ignore reinstatement premiums, aggregate limits, exclusions, expenses and timing. Those contractual features must be incorporated in implementation models.

A higher retention generally reduces reinsurance premium and may improve expected underwriting profit, but it increases net volatility, tail exposure and required capital. A lower retention produces the opposite trade-off. The correct retention is not necessarily the one with the lowest expected cost; it should be selected subject to risk appetite, liquidity, capital and counterparty constraints.

Worked example: excess-of-loss recovery

A claim of 1,200,000 is subject to a retention of 500,000 and a layer limit of 600,000. Recovery is:

The insurer retains 600,000. If the claim were 800,000, recovery would be 300,000 and the insurer would retain 500,000. If the claim were 2,000,000, recovery would still be 600,000, leaving the insurer with 1,400,000 because the layer is exhausted.

5.2 Forms of reinsurance

• Quota share: the insurer and reinsurer share premiums and losses in a fixed proportion. It provides broad proportional risk transfer and may support growth or capital relief.
• Surplus share: the reinsurer participates above a retention measured in lines, subject to treaty capacity.
• Per-risk excess of loss: protects against large individual risks.
• Catastrophe excess of loss: protects against accumulation from one event.
• Aggregate excess of loss: responds when total losses exceed an attachment during a period.
• Stop-loss: protects an aggregate loss ratio or amount, often subject to important definitions and limits.

The economic value of reinsurance includes protection against tail losses, capital relief, earnings stability, expertise and access to capacity. Costs include premium, frictional expenses, collateral or security costs, basis risk, counterparty default and possible moral hazard.

5.3 Reinsurance and risk-adjusted performance

A decision can be evaluated using a risk-adjusted return measure. A simple form is:

The exact numerator depends on the organization’s accounting and management framework. Expected loss should not be subtracted twice, and capital must be measured consistently with the risk being evaluated. An alternative is return on risk-adjusted capital, using allocated capital for a line of business, treaty or portfolio.

A reinsurance purchase may lower accounting profit but increase RAROC if it reduces economic capital more than it reduces risk-adjusted earnings. The analysis should include the reinsurer’s credit risk and the possibility that recoveries are delayed or disputed.

6. Integrated decision framework

An insurer can structure a retention and capital decision as follows:

1. Define the objective: profitability, solvency, earnings stability, growth or rating protection.
2. Specify the exposure basis: policy year, accident year, underwriting year or calendar year.
3. Validate data: exposure, claims development, inflation, limits, deductibles and catastrophe coding.
4. Model gross losses: frequency, severity, dependence and parameter uncertainty.
5. Apply reinsurance terms: attachment, limit, reinstatements, exclusions, premiums and recoverability.
6. Calculate risk measures: expected loss, VaR, TVaR, stress loss and liquidity needs.
7. Determine capital: aggregate risks using an appropriate dependence model and diversification treatment.
8. Evaluate return: expected profit, RAROC, solvency ratio and sensitivity to assumptions.
9. Test alternatives: retention levels, treaty structures, capital targets and adverse scenarios.
10. Govern the decision: document assumptions, approvals, monitoring indicators and model limitations.

Illustrative case: selecting a catastrophe retention

Background

A fictional property insurer has annual catastrophe loss scenarios after ordinary claims but before catastrophe reinsurance. Management is considering a per-event excess-of-loss treaty with a retention of 20 million and a limit of 80 million. The insurer’s risk appetite requires that a modeled 99% annual catastrophe loss should not exceed 35 million net of reinsurance.

Problem

The insurer must determine whether the proposed retention is consistent with risk appetite and whether the premium is justified by capital relief. The scenario model estimates the following annual gross catastrophe losses:

Analysis

With a 20 million retention and an 80 million limit, net losses are 0, 15 million, 20 million and 70 million respectively. The 99% VaR is 70 million because the cumulative probability reaches 99% at the extreme-event net loss. This breaches the 35 million risk appetite.

Management considers a second layer: retention 10 million and limit 30 million, producing net losses of 0, 10 million, 30 million and 120 million if this layer is the only cover. The extreme event remains highly exposed because the layer is exhausted. This illustrates why attachment and limit must be considered together; reducing retention alone does not eliminate an upper tail.

Suppose instead the insurer purchases a 10 million retention with a 100 million limit. Net losses become 0, 10 million, 40 million and 50 million. The 99% VaR is 50 million—an improvement, but still above appetite. A further aggregate or catastrophe layer is required, or the appetite must be reconsidered. The final choice should also consider treaty premium, reinsurer security, reinstatement terms, modeled uncertainty and liquidity.

Outcome and learning points

The case demonstrates that:

• A treaty must be evaluated on the full distribution, not only expected recovery.
• A limit can leave substantial residual tail risk.
• VaR may conceal the difference between 50 million and much larger losses beyond the threshold; TVaR and stress scenarios are also required.
• Capital relief depends on net tail risk and counterparty credit quality.
• Risk appetite is a decision constraint, not merely a reporting statistic.

7. Limitations, governance and contemporary practice

Model and data limitations

Credibility estimates can be distorted by changing exposure mix, claims inflation, incomplete development and inadequate collective benchmarks. VaR and TVaR depend on distributional assumptions, simulation quality and treatment of rare events. Catastrophe models may contain parameter, event-set and vulnerability uncertainty. Reinsurance models may omit wording ambiguity, reinstatement costs or collection delays.

Model risk management

Good practice includes independent validation, version control, back-testing where possible, sensitivity analysis, scenario testing, documentation of expert judgement and clear ownership. Parameter uncertainty can be represented through simulation of model parameters rather than treating estimates as known constants. Reverse stress testing asks what combination of events would threaten solvency, even if the scenario is outside the central model.

Regulatory and ethical considerations

Capital models affect pricing, underwriting access, dividends and policyholder security. Governance should prevent deliberate understatement of risk to achieve commercial targets. Regulators commonly require adequate technical provisions, eligible own funds, risk management, reinsurance assessment and supervisory reporting. Professional actuarial standards require appropriate data, transparent assumptions, communication of uncertainty and avoidance of misleading precision.

Technology and AI

Cloud computing, automated claims data and machine-learning models can improve segmentation and claims prediction. However, predictive accuracy does not guarantee sound capital decisions. AI systems require controls for data drift, explainability, fairness, cybersecurity, human review and monitoring of tail performance. Rare catastrophic outcomes are especially difficult for data-hungry algorithms because historical samples contain few observations. Expert judgement and scenario analysis remain essential.

Best-practice checklist

• State the horizon, confidence level, currency and gross/net basis for every risk measure.
• Use credibility only after assessing homogeneity, exposure changes and data quality.
• Report both central estimates and uncertainty ranges.
• Use TVaR, stress testing and liquidity analysis alongside VaR.
• Model reinsurance wording, limits, reinstatements, exclusions and counterparty risk.
• Separate regulatory capital from economic capital and explain reconciliation.
• Test diversification assumptions under crisis correlations and tail dependence.
• Evaluate treaties using expected profit, capital relief, RAROC and solvency resilience.
• Document model limitations and obtain appropriate independent review.

Chapter Summary

Credibility provides a disciplined method for combining an insurer’s own experience with collective information. Limited-fluctuation methods focus on the exposure needed for statistical reliability, while Bühlmann credibility separates process variation from variation between risks. Aggregate loss models combine frequency and severity uncertainty.

VaR identifies a loss quantile, whereas TVaR measures the average loss in the tail and is generally more informative for severe-loss capital decisions. Ruin theory links premium income, initial surplus and claims to the probability of insolvency, but classical models rely on restrictive assumptions. Solvency capital must be interpreted consistently with its valuation basis, horizon and aggregation method.

Retention and reinsurance decisions involve a trade-off between expected cost, volatility, tail protection, capital, liquidity and counterparty risk. Diversification can reduce capital, but dependence and crisis correlation may substantially reduce its value. Strategic actuarial decisions therefore require integrated modeling, stress testing, governance and professional judgement rather than reliance on one metric.

Key Terms

Credibility: A method of weighting individual experience and collective expectation.

Credibility factor: The weight assigned to individual experience, usually between 0 and 1.

Limited-fluctuation credibility: An approach based on achieving a specified statistical reliability with sufficient exposure.

Bühlmann credibility: A credibility model separating process variance from variance between hypothetical risk means.

Process variance: Random variation within a particular risk.

Collective mean: The expected value used as the benchmark for the portfolio or reference class.

Aggregate loss: The total loss from all claims in a specified period.

Frequency: The number of claims in a period.

Severity: The monetary size of an individual claim.

Value at Risk: A specified quantile of a loss distribution.

Tail Value at Risk: The average loss in the tail beyond a specified confidence level.

Ruin: The event that insurer surplus falls below zero.

Solvency capital: Financial resources available to absorb adverse deviations and support obligations.

Retention: The portion of loss retained by the insurer.

Excess-of-loss reinsurance: Cover responding to losses above an attachment or retention, subject to a limit.

Counterparty risk: The possibility that a reinsurer cannot meet its contractual obligation.

Diversification: Reduction in aggregate risk arising from imperfect dependence between risks.

RAROC: Risk-adjusted return on capital.

Model risk: The possibility that model assumptions, design or implementation produce misleading results.

Review Questions

Short Answer Questions

1. What is the purpose of a credibility factor?
2. Distinguish VaR from TVaR.
3. State two assumptions of the classical compound Poisson risk model.
4. What is meant by an insurer’s retention?
5. Why can diversification reduce required capital?

Long Answer Questions

1. Explain Bühlmann credibility, including EPV, VHM and the credibility factor.
2. Compare VaR and TVaR as measures for solvency capital.
3. Discuss the relationship between premium income, surplus and ruin probability.
4. Explain the principal economic advantages and disadvantages of catastrophe excess-of-loss reinsurance.
5. Describe an integrated framework for selecting retention and reinsurance arrangements.

Analytical / Application Questions

1. An insurer has an observed loss ratio of 82%, a collective estimate of 68% and credibility factor 0.45. Calculate the credibility-weighted estimate and interpret it.
2. A claim of 4 million is subject to a retention of 1 million and a per-occurrence limit of 2 million. Calculate the reinsurer recovery and insurer net loss. Explain the result if the claim rises to 6 million.
3. Two portfolios have equal expected loss, but Portfolio A has a much larger TVaR. Which portfolio requires greater attention in capital and reinsurance planning, and why?

Multiple Choice Questions with Answers

Question 1

An insurer observes a loss ratio of 80%, has a collective estimate of 60%, and uses credibility factor Z = 0.25. What is the credibility estimate?

A. 62%

B. 65%

C. 70%

D. 75%

Correct Answer: B. 65%

Question 2

In Bühlmann credibility, K is defined as:

A. VHM ÷ EPV

B. EPV + VHM

C. EPV ÷ VHM

D. EPV × VHM

Correct Answer: C. EPV ÷ VHM

Question 3

For a continuous loss distribution, VaR at the 99% level is best interpreted as:

A. The average annual loss

B. The loss threshold exceeded with probability about 1%

C. The maximum possible loss

D. The average of all losses

Correct Answer: B. The loss threshold exceeded with probability about 1%

Question 4

Which measure focuses on the average loss in the worst tail of the distribution?

A. Expected frequency

B. Loss ratio

C. VaR

D. TVaR

Correct Answer: D. TVaR

Question 5

If claim frequency and severity are independent, expected aggregate loss equals:

A. E[N] + E[X]

B. E[N] − E[X]

C. E[N] × E[X]

D. E[N] ÷ E[X]

Correct Answer: C. E[N] × E[X]

Question 6

A claim of 900,000 is subject to a retention of 400,000 and an excess-of-loss limit of 300,000. What is the reinsurer recovery?

A. 300,000

B. 400,000

C. 500,000

D. 900,000

Correct Answer: A. 300,000

Question 7

In the classical risk model, positive safety loading requires premium income rate c to be:

A. Less than λE[X]

B. Equal to zero

C. Greater than λE[X]

D. Equal to λ ÷ E[X]

Correct Answer: C. Greater than λE[X]

Question 8

Which statement about a reinsurance limit is correct?

A. It guarantees that all losses are reimbursed

B. It caps the reinsurer’s payment under the relevant layer

C. It eliminates counterparty risk

D. It always increases the insurer’s expected profit

Correct Answer: B. It caps the reinsurer’s payment under the relevant layer

Question 9

If two risks have correlation below 1, their aggregate variance may be reduced because:

A. Their expected losses become zero

B. The covariance contribution is less than under perfect positive dependence

C. Their standard deviations disappear

D. Frequency is necessarily Poisson

Correct Answer: B. The covariance contribution is less than under perfect positive dependence

Question 10

RAROC generally relates risk-adjusted return to:

A. Economic capital

B. Gross written premium only

C. Claim count only

D. The number of policies issued

Correct Answer: A. Economic capital