From a Decision to a Capital Model
An economic-capital model is useful only when its design fits the decision it is meant to support. A model built to compare risk profiles over a one-year planning horizon may not answer a question about liquidity over the next month. Likewise, a precisely calculated output is not persuasive if important exposures were omitted or the result cannot be independently checked.
This workshop follows one design chain: define the decision and measure, map exposures, specify the model architecture, calibrate it, test its behavior, validate its implementation, and document governance and limitations. The numbers below describe hypothetical insurers and are not industry observations. The chosen confidence levels and calibrations are illustrative, not prescribed regulatory standards.
The definition assumes that the model represents losses on a consistent basis for the stated horizon, that the loss distribution is specified, and that expected loss and the selected quantile are calculated from that same distribution. For a discrete distribution, the quantile is the smallest loss whose cumulative probability is at least \(\alpha\); it need not be a loss with exactly \(\alpha\) probability below it. This measure does not, by itself, describe liquidity needs, losses beyond the chosen quantile, or the quality of the underlying data.
Design the Model Before Calculating Capital
A design specification should make the model’s purpose operational. It identifies the decision owner, the decision to be informed, the valuation date, the projection horizon, the loss perspective, and the output required. For example, a hypothetical insurer might use a one-year, pre-tax loss model to assess whether its risk profile has changed enough to prompt a management review. That purpose does not automatically make the resulting number suitable for pricing, liquidity planning, or every external reporting use.
Next, create an exposure map. Each material exposure should have a named source, an owner, a model component, a unit and valuation basis, and a documented treatment for uncertainty or exclusion. The map should explain where policy losses, investment losses, operating events, and risk-transfer recoveries enter the calculation. It should also identify overlaps: counting the same event in two components can inflate losses, while leaving it out of both can understate them.
The model architecture then specifies how mapped exposures become outputs. A simple architecture could use scenario-level gross losses by risk component, apply documented reinsurance recoveries, aggregate to net loss, and calculate the selected measure from the resulting distribution. The architecture must state how components interact. Adding component quantiles is not generally equivalent to calculating a quantile of aggregate loss, because the joint behavior of components matters. The model should therefore describe how its scenarios represent co-occurrence and dependence, without implying that a convenient assumption is evidence.
Name the decision, user, horizon, valuation basis, and intended output before selecting a model method.
Trace material exposures from source systems to model components, units, assumptions, and documented exclusions.
State how scenarios, component losses, recoveries, aggregation, and the capital measure connect.
Explain parameter choices, compare results with suitable evidence, and test scenarios and sensitivities.
Independently check data and implementation, record limitations, assign ownership, and define change controls.
Worked Example: Build a Scenario-Based Capital View
Worked Example: Build a Scenario-Based Capital View
A hypothetical insurer wants a one-year, pre-tax view of loss for internal risk monitoring. Its simplified model uses five mutually exclusive scenarios, intended to represent the full probability distribution for this exercise. Loss amounts are in thousands of dollars. Calculate expected loss, 95% Value at Risk, and economic capital defined as Value at Risk less expected loss. The method is a discrete scenario calculation; it avoids the mistake of adding risk-component quantiles instead of first aggregating losses within each scenario.
| Scenario | Probability | Underwriting loss ($000) | Market loss ($000) | Operational loss ($000) |
|---|---|---|---|---|
| Routine year | 0.70 | 100 | 20 | 10 |
| Claims surge | 0.20 | 300 | 30 | 20 |
| Claims and market stress | 0.07 | 250 | 180 | 40 |
| Catastrophe year | 0.02 | 700 | 40 | 100 |
| Severe combined stress | 0.01 | 900 | 250 | 150 |
Step 1: Check the distribution and aggregate each scenario. The probabilities sum to \(0.70+0.20+0.07+0.02+0.01=1.00\). Scenario loss is the sum of its three components: routine \(100+20+10=130\); claims surge \(300+30+20=350\); claims and market stress \(250+180+40=470\); catastrophe \(700+40+100=840\); and severe combined stress \(900+250+150=1{,}300\), all in thousands.
Step 2: Calculate expected loss. Multiply each aggregate loss by its probability and sum:
Expected loss is $223,700. The units remain thousands until converted to dollars.
Step 3: Find the 95% quantile. The ordered cumulative probabilities are 0.70 at loss 130, 0.90 at 350, 0.97 at 470, 0.99 at 840, and 1.00 at 1,300. The first cumulative probability at least 0.95 is 0.97, so \(\operatorname{VaR}_{0.95}(X)=470\), or $470,000.
Step 4: Calculate the defined capital measure and check it.
The result is $246,300. As a check, it is below the 95% Value at Risk because expected loss is positive, and adding expected loss back gives \(246.3+223.7=470\), as required by the definition. This model result depends on the assumed five-scenario distribution; it is not evidence that losses above $470,000 are immaterial.
Calibration, Scenarios, and Validation
Calibration translates data, expert judgment, and selected targets into model parameters or adjustments. The documentation should distinguish empirical estimates from judgmental assumptions and identify the evidence and rationale for each material choice. A calibration target can help reveal a mismatch, but matching a target does not establish that the exposure map, dependence assumptions, or implementation is sound.
Scenario testing asks how the model behaves under defined conditions. A scenario should specify the event, affected exposures, severity, duration or horizon, and any relevant management or risk-transfer response. Sensitivity analysis varies selected assumptions to show which inputs drive results. These are diagnostic tools: a plausible-looking result under one stress does not prove the model is complete or correctly implemented.
Validation should be proportionate to model use and independent enough to challenge the model’s design and calculation. It can include tracing data from source to output, recalculating selected cases outside the production engine, testing boundary conditions, comparing results with appropriate benchmarks, and reviewing whether assumptions fit the intended use. Each finding needs an owner, severity, action, and disposition. Unresolved limitations should remain visible to decision-makers rather than being hidden by a single capital number.
Worked Example: Apply a Calibration Adjustment Without Overclaiming
A hypothetical insurer’s documented one-year model currently produces a 95% Value at Risk of $4,800,000 and expected loss of $1,600,000. Following a review of the evidence, the model owner proposes a uniform 1.05 multiplier on all modeled losses as a temporary calibration adjustment. Assume the adjustment is accepted for this example and does not change scenario probabilities or ranks. Calculate the adjusted measure. The method is proportional rescaling; it avoids the mistake of adjusting the Value at Risk while leaving expected loss on the old scale.
Step 1: Apply the same multiplier to both quantities. Adjusted Value at Risk is \(1.05\times \$4{,}800{,}000=\$5{,}040{,}000\). Adjusted expected loss is \(1.05\times \$1{,}600{,}000=\$1{,}680{,}000\).
Step 2: Recalculate capital on the adjusted basis.
Before adjustment, the corresponding difference was \(\$4{,}800{,}000-\$1{,}600{,}000=\$3{,}200{,}000\). The adjusted result is $3,360,000, also equal to \(1.05\times \$3{,}200{,}000\), which checks the calculation. This arithmetic does not justify the multiplier: the model record must explain its evidence, scope, approval, monitoring, and whether a uniform adjustment is appropriate across risks. A scalar adjustment also does not repair missing exposures or incorrect scenario structure.
Worked Example: Reconcile a Reinsurance Calculation
For a hypothetical validation test, an insurer has gross claims loss of $2,600,000, operational loss of $350,000, and market loss of $250,000 in a specified one-year scenario. A treaty reimburses 20% of claims loss above a $500,000 retention, subject to a $700,000 recovery limit. Assume the treaty responds to this entire claims amount and there are no other recoveries or adjustments. Calculate net aggregate loss and compare it with an engine output of $2,830,000. The method is an independent calculation benchmark; it avoids applying the treaty percentage to all claims rather than only the layer above retention.
Step 1: Calculate the eligible layer. Claims above retention are \(\$2{,}600{,}000-\$500{,}000=\$2{,}100{,}000\).
Step 2: Apply the share and limit. The uncapped recovery is \(0.20\times \$2{,}100{,}000=\$420{,}000\). Since $420,000 is below the $700,000 limit, the recovery is $420,000.
Step 3: Calculate net aggregate loss. Net claims loss is \(\$2{,}600{,}000-\$420{,}000=\$2{,}180{,}000\). Adding the other components gives \(\$2{,}180{,}000+\$350{,}000+\$250{,}000=\$2{,}780{,}000\).
Step 4: Reconcile and assess. The engine exceeds the independent result by \(\$2{,}830{,}000-\$2{,}780{,}000=\$50{,}000\). If the hypothetical test’s investigation threshold is $25,000, the $50,000 difference exceeds it and must be investigated before the test passes. The benchmark establishes the expected result under the stated treaty interpretation; it does not, by itself, identify the cause of the engine difference.
Governance and the Model Record
A complete model record should allow another qualified reviewer to understand why the model exists, how its output is produced, and what would make the output unreliable for a particular decision. At minimum, document the purpose and users, horizon and measure, exposure map, data sources, assumptions, scenario construction, aggregation logic, calibration choices, validation evidence, known limitations, and approvals. Record material model changes with their rationale, impact assessment, testing, approval, effective date, and rollback or contingency plan where appropriate.
Governance also defines how results are used. The model owner maintains the calculation and documentation; independent reviewers challenge design and implementation; and decision-makers receive limitations and unresolved findings alongside results. Escalation triggers might include a material exposure change, a data-quality failure, an unexplained benchmark difference, or a limitation that invalidates the intended use. Controls should prevent an unresolved defect from being treated as an ordinary fluctuation in capital.
| Design choice | What to record | Why it matters |
|---|---|---|
| Decision and measure | User, purpose, horizon, confidence level, loss basis, and definition of capital | Prevents an output from being applied to a different question without justification |
| Exposure mapping | Sources, owners, components, units, exclusions, and risk-transfer treatment | Helps identify omissions, double counting, and inconsistent bases |
| Calibration | Evidence, judgment, adjustments, approvals, and monitoring plan | Separates supportable assumptions from unexplained target fitting |
| Validation and limitations | Independent tests, findings, owners, status, and use restrictions | Makes uncertainty and defects actionable rather than invisible |
| Change control | Reason, impact, testing, approval, and implementation record | Supports reproducibility and controlled evolution of the model |
Common Design Traps
- Using a measure without defining its basis. State the horizon, loss perspective, confidence level, and treatment of expected loss before interpreting capital.
- Aggregating component quantiles as if they were aggregate capital. Calculate aggregate loss under a coherent joint scenario or distribution, then apply the measure to that aggregate.
- Treating calibration as validation. A result that matches a chosen target can still omit exposures or contain calculation errors.
- Confusing a scenario test with a probability forecast. A severe scenario can reveal vulnerabilities without being assigned a reliable frequency.
- Reporting a number without its limitations. A decision-useful output includes the assumptions, unresolved findings, and conditions under which it should not be used.
Check Your Understanding
For each problem, state any assumptions needed and work from the defined model basis.
- A hypothetical insurer uses a one-year loss distribution. Explain what additional information should be specified before interpreting its reported 99% economic-capital result. (Answer: State the loss perspective and valuation basis, the definition of economic capital and treatment of expected loss, and the model’s horizon and intended decision use; also identify material assumptions and limitations affecting the result.)
- A discrete model has losses of $100,000, $300,000, $700,000, and $1,200,000 with probabilities 0.60, 0.30, 0.08, and 0.02, respectively. Using the definition in this tutorial, calculate 95% Value at Risk and economic capital. (Answer: Expected loss is $230,000; 95% Value at Risk is $700,000; economic capital is $470,000.)
- A hypothetical model reports expected loss of $2,500,000 and 95% Value at Risk of $6,000,000. A justified uniform loss multiplier of 1.08 is applied without changing scenario ranks or probabilities. Calculate the adjusted economic capital. (Answer: Adjusted expected loss is $2,700,000; adjusted Value at Risk is $6,480,000; adjusted economic capital is $3,780,000.)
- A hypothetical treaty reimburses 25% of claims above a $400,000 retention, capped at $300,000. Gross claims are $1,600,000; operational and market losses are $200,000 and $150,000. Assuming the treaty applies to all stated claims, calculate net aggregate loss. (Answer: Eligible claims are $1,200,000; the uncapped recovery is $300,000, equal to the limit; net claims are $1,300,000; net aggregate loss is $1,650,000.)
- An independent check gives a net loss of $1,650,000, while the production engine reports $1,690,000. The documented investigation threshold is $30,000. What should the validation conclusion be? (Answer: The $40,000 difference exceeds the $30,000 threshold, so the check fails pending investigation and documented disposition; the difference alone does not identify its cause.)