Integrated Operational Loss Modeling Workshop

Introduction

An operational loss model is useful only when its parts form a coherent chain: event records must represent the intended exposure, fitted distributions must match the data treatment, and the annual aggregate must include every modeled component exactly once. A capital estimate then depends both on the assumed loss process and on the chosen risk measure. This workshop connects those steps in one hypothetical modeling exercise.

The examples use amounts in thousands of dollars unless stated otherwise. This keeps calculations readable, but units must remain consistent: a variance stated in squared thousands of dollars is not a dollar amount. The modeling assumptions are deliberately explicit. In practice, data limitations and model uncertainty should be reported alongside the resulting estimates.

Definition: In a frequency-severity model, $N$ is the number of loss events during one year, $X_j$ is the severity of event $j$, and the annual aggregate loss is $S=\sum_{j=1}^{N}X_j$, with $S=0$ when $N=0$. A compound Poisson model assumes $N\sim\operatorname{Poisson}(\lambda)$, event severities are independent and identically distributed, and the severities are independent of $N$. Here $\lambda$ is the expected annual event count.

Prepare the Data Before Fitting

Start by defining the modeled event, observation period, and loss amount. For example, a model might include external payment-processing errors recorded when they are recognized, with each record expressed as a gross loss in the same price level. The definition should also clarify how repeated entries, recoveries, reopened claims, and losses spanning multiple dates are handled. These are data rules, not distributional assumptions.

For a basic frequency-severity fit, the annual event counts and individual event severities must refer to the same population of losses. If losses below a reporting threshold are absent, the observed severity sample is truncated; fitting an untruncated distribution to it without adjustment can misstate the severity distribution. Likewise, missing years or inconsistent event definitions can distort frequency. The examples below assume complete, consistently defined records and no reporting threshold.

Specify the event and loss basis.
State what qualifies as one event, which amounts are included, and the currency and valuation date.

Build consistent annual counts.
Use one count for each complete year in the exposure period, including zero-loss years.

Build the severity sample.
Retain one modeled severity per event under the chosen event definition; reconcile totals to the underlying records.

Fit, aggregate, and document.
Estimate parameters under stated assumptions, calculate the annual loss distribution or its summaries, and record limitations and uncertainty.

Worked Example: From Event Records to a Fitted Model

A hypothetical operations unit has four complete years of records. Annual event counts are $1,0,2,1$. The four gross event severities, in thousands of dollars, are $10,20,30,40$. Assume all events are captured, the annual counts are independent Poisson observations with a constant rate, and severities are independent exponential observations with a common mean. Fit the model and calculate its expected annual loss.

Frequency fit. The total count is $1+0+2+1=4$, and the exposure is four years. The Poisson maximum-likelihood estimate is $\hat{\lambda}=4/4=1$ event per year. The zero-event year belongs in the exposure denominator; omitting it would overstate the fitted rate.

Severity fit. The total observed severity is $10+20+30+40=100$ thousand dollars, across four events. For an exponential distribution with mean parameter $\theta$, the density is $f_X(x)=\theta^{-1}e^{-x/\theta}$ for $x\geq0$. Its likelihood for $n$ observations is proportional to $\theta^{-n}e^{-\sum x_j/\theta}$; maximizing its log-likelihood gives $\hat{\theta}=\sum x_j/n=100/4=25$ thousand dollars.

Expected annual loss. A compound Poisson model has $E[S]=\lambda E[X]$, so the fitted expected annual loss is $1(25)=25$ thousand dollars, or $25,000. This is a mean, not a high-quantile capital estimate.

Check and common mistake. The fitted mean severity is $25,000 and the fitted annual event rate is one, so their product has units of dollars per year and equals $25,000 per year. Excluding the zero-count year would instead give $4/3$ events per year, incorrectly treating the observed exposure as only three years.

Aggregate Frequency and Severity

The compound Poisson model connects event-level assumptions to annual loss. If severity has finite mean and variance, independence gives $E[S]=\lambda E[X]$. For variance, condition first on the event count: $E[S\mid N]=NE[X]$ and $\operatorname{Var}(S\mid N)=N\operatorname{Var}(X)$. The law of total variance then yields:

\begin{aligned} \operatorname{Var}(S) &=E[\operatorname{Var}(S\mid N)]+\operatorname{Var}(E[S\mid N])\\ &=\lambda\operatorname{Var}(X)+\lambda(E[X])^2\\ &=\lambda E[X^2]. \end{aligned}

The final equality uses $E[X^2]=\operatorname{Var}(X)+(E[X])^2$. The result depends on Poisson frequency and independence of frequency and severity. It does not generally apply unchanged if the count is overdispersed, severities depend on the count, or event severities are dependent. Also, $\operatorname{Var}(S)$ is not $\lambda\operatorname{Var}(X)$ alone: that expression omits variation in the number of events.

For the exponential fit in the preceding example, $E[X^2]=2\theta^2$. With $\hat{\lambda}=1$ and $\hat{\theta}=25$ thousand dollars, the fitted aggregate variance is $1(2)(25^2)=1{,}250$ squared thousands of dollars. The corresponding standard deviation is $\sqrt{1{,}250}=35.3553$ thousand dollars. This standard deviation describes annual process variability under the fitted model; it is not a confidence interval for the parameters.

Scenario Incorporation and Capital

A scenario can be represented as an additional annual loss variable when its interpretation and dependence assumptions are specified. The next example assumes a one-time annual scenario amount that is independent of the baseline loss. It does not treat the scenario as an event already included in the baseline data. For discrete annual loss, define the $p$-level Value-at-Risk by $\operatorname{VaR}_p(S)=\inf\{s:F_S(s)\geq p\}$, where $F_S(s)=P(S\leq s)$. Because the loss distribution has jumps, the quantile must be found from the cumulative probability, not by interpolation.

Worked Example: Add an Independent Scenario and Estimate Capital

A hypothetical baseline has $N\sim\operatorname{Poisson}(1)$ events per year, each with a fixed severity of $10,000. Separately, a distinct scenario causes a $50,000 loss in a year with probability 0.02 and causes no loss otherwise. Assume the scenario indicator is independent of baseline frequency. Find expected annual loss, variance, 99% VaR, and unexpected-loss capital defined as 99% VaR minus expected annual loss.

Expected loss. Baseline expected loss is $1(10)=10$ thousand dollars. Scenario expected loss is $0.02(50)=1$ thousand dollars. By additivity of expectation, total expected annual loss is $10+1=11$ thousand dollars, or $11,000.

Variance. For the fixed-severity compound Poisson baseline, the variance is $\lambda E[X^2]=1(10^2)=100$ squared thousands of dollars. The scenario is a Bernoulli amount: its variance is $p(1-p)(50^2)=0.02(0.98)(2{,}500)=49$ squared thousands of dollars. Independence permits adding variances, giving $100+49=149$ squared thousands of dollars. The standard deviation is $\sqrt{149}=12.2066$ thousand dollars.

99% VaR. Write $S=10N+50I$, in thousands of dollars, where $I$ is one when the scenario occurs and zero otherwise. At a loss of 50 or less, either the scenario is absent and $N\leq5$, or it is present and $N=0$. Thus $P(S\leq50)=0.98P(N\leq5)+0.02P(N=0)$. For a Poisson(1) count, $P(N\leq5)=0.999405815$ and $P(N=0)=e^{-1}=0.367879441$. Therefore $P(S\leq50)=0.98(0.999405815)+0.02(0.367879441)=0.9867753$, below 0.99.

At 60 or less, the scenario-absent case allows $N\leq6$, while the scenario-present case allows $N\leq1$. Hence $P(S\leq60)=0.98P(N\leq6)+0.02P(N\leq1)$. Using $P(N\leq6)=0.999916758$ and $P(N\leq1)=0.735758882$ gives $P(S\leq60)=0.98(0.999916758)+0.02(0.735758882)=0.9946336$. Since the CDF is below 0.99 at 50 and at least 0.99 at 60, the 99% VaR is $60,000.

Capital and check. Unexpected-loss capital is $60,000 − $11,000 = $49,000. The calculation avoids a common trap: at $50,000 the cumulative probability is only 0.9867753, so that amount is not the 99% VaR. Adding the scenario’s $1,000 expected loss to the baseline expectation is correct, but it does not by itself determine the tail quantile.

Uncertainty in the Estimate

Annual loss variability and uncertainty in estimated model parameters are different. Process variability concerns what the next year’s losses might be if the model parameters were known. Parameter uncertainty concerns how much the fitted parameters themselves might change with another sample. A capital result based on a single fitted model does not automatically include the second source.

A simple approximate illustration uses the frequency and severity estimates from the first example. Under independent Poisson counts observed for $T$ years, $\operatorname{Var}(\hat{\lambda})=\lambda/T$. For $n$ independent exponential severities with mean $\theta$, $\operatorname{Var}(\hat{\theta})=\theta^2/n$. If the estimates are independent, a first-order delta-method approximation for the variance of $\hat{\lambda}\hat{\theta}$ is:

\operatorname{Var}(\hat{\lambda}\hat{\theta}) \approx \theta^2\operatorname{Var}(\hat{\lambda}) +\lambda^2\operatorname{Var}(\hat{\theta}).

For the hypothetical data, use $T=4$, $n=4$, $\hat{\lambda}=1$, and $\hat{\theta}=25$, with monetary amounts in thousands. The estimated variances are $1/4=0.25$ for $\hat{\lambda}$ and $25^2/4=156.25$ for $\hat{\theta}$. The approximate variance of the estimated expected annual loss is $25^2(0.25)+1^2(156.25)=312.5$ squared thousands, giving an approximate standard error of $\sqrt{312.5}=17.6777$ thousand dollars.

A rough normal interval using 1.96 standard errors is $25\pm1.96(17.6777)$, or approximately $(-9.65,59.65)$ thousand dollars. A negative lower endpoint is not a plausible expected loss; it signals that a normal approximation is crude for this small sample, not that expected loss can be negative. The interval is not a capital confidence interval and should not be presented as a precise uncertainty assessment. More extensive methods, such as bootstrap or model-uncertainty analysis, can be used when warranted by the decision.

Interpreting the Complete Model

A useful model report separates what is estimated from what is assumed. The frequency and severity parameters come from the chosen data and fitting method. The annual aggregate results additionally depend on the compound Poisson structure, the independence assumptions, and any scenario treatment. The risk measure then determines how those results are summarized. Expected loss, standard deviation, and VaR answer different questions and should not be substituted for one another.

State the exposure basis. Give the event definition, observation period, data completeness, and any threshold or adjustment.
Keep units visible. If losses are in thousands, variances are in squared thousands; convert only after completing the calculation.
Respect discrete quantiles. Confirm that the CDF immediately below a proposed VaR is less than the target probability and that the CDF at the VaR reaches it.
Separate uncertainty sources. Process variability is not the same as parameter, scenario, or model uncertainty.
Explain the scenario layer. State whether it is distinct from modeled events and specify its dependence on baseline losses.

This end-to-end discipline is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus: a technically correct formula is not enough unless its assumptions, data basis, units, and interpretation also fit the modeled risk.

Check Your Understanding

Use thousands of dollars for monetary calculations unless a question says otherwise. Assume independence wherever it is explicitly stated.

A hypothetical operation records annual counts $0,2,1,3$ over four complete years. Under a constant-rate Poisson frequency model, calculate the maximum-likelihood annual frequency estimate. (Answer: 1.5 events per year)
Three complete exponential severity observations are 12, 18, and 30. If the annual Poisson frequency estimate is 1.5, find the fitted mean severity and expected annual loss. (Answer: mean severity = $20,000; expected annual loss = $30,000)
A compound Poisson model has frequency parameter $\lambda=0.5$ and fixed severity $20,000. Find expected annual loss, aggregate variance in squared thousands, and aggregate standard deviation. (Answer: expected loss = $10,000; variance = 200; standard deviation = $14,142.14)
A baseline has Poisson(1) annual frequency and fixed severity $10,000. An independent distinct scenario pays $40,000 with probability 0.05 each year. Find total expected annual loss and variance in squared thousands. (Answer: expected loss = $12,000; variance = 176)
A baseline has Poisson(1) frequency and fixed severity $5,000. An independent distinct scenario pays $25,000 with annual probability 0.01. Calculate the 99% VaR and unexpected-loss capital, defined as VaR minus expected annual loss. (Answer: 99% VaR = $25,000; expected annual loss = $5,250; unexpected-loss capital = $19,750)

What You'll Learn