Empirical Stop-Loss Transforms

Why the Empirical Stop-Loss Transform Matters

A stop-loss contract with attachment point $d$ pays the amount by which a loss exceeds $d$, if any. When aggregate losses are available as observations rather than as a fully specified probability model, the empirical stop-loss transform summarizes the average payment at each attachment point. It is useful for comparing attachment choices and checking how a change in the threshold affects the observed average excess.

The calculation is direct, but details matter. An observation equal to the attachment produces no payment; the denominator is the number of observations, not just the number that exceed the attachment. The empirical transform is an average over the full observed sample, so it is not the same quantity as the average excess conditional on exceeding the attachment.

Definition: Let $x_1,\ldots,x_n$ be observed aggregate losses, with $n\geq 1$. Their empirical distribution assigns probability $1/n$ to each observation, counting repeated observations with their full multiplicity. The empirical stop-loss transform at attachment point $d$ is $$ \widehat{\Pi}(d)=\frac{1}{n}\sum_{j=1}^{n}(x_j-d)^{+}, \qquad (z)^{+}=\max(z,0). $$ It is the average payment per observation under the empirical distribution. It is not, by itself, a loaded premium or a prediction that accounts for sampling uncertainty.

The observations may be listed in any order. For each one, subtract $d$; replace a negative result by zero; then add the resulting payments and divide by $n$. If the data contain repeated values, each occurrence is included. This is equivalent to assigning the empirical probability mass for a value in proportion to its frequency in the data.

Calculating the Transform and Handling Ties

An observation above $d$ contributes $x_j-d$; an observation at or below $d$ contributes zero. Thus, the exceedance count uses the strict condition $x_j>d$. This distinction is especially important at an attachment that equals one or more observed losses.

Worked Example: Two Attachments and an Exact Tie

A hypothetical portfolio has 12 observed annual aggregate losses, in thousands of dollars: $10,20,30,40,50,100,100,110,120,130,140,150$. Calculate the empirical stop-loss transform at $d=50$ and $d=100$, and check the change between the two values.

Method. Apply $(x_j-d)^+$ to every observation and divide the total payment by all 12 observations. This method avoids dropping observations that make no payment from the denominator.

At $d=50$. The first five observations are at or below 50 and contribute zero. The seven observations above 50 give payments $100-50=50,\ 100-50=50,\ 110-50=60,\ 120-50=70,\ 130-50=80,\ 140-50=90,\ 150-50=100$. Their total is $50+50+60+70+80+90+100=500$. Therefore,

\widehat{\Pi}(50)=\frac{500}{12}=41.6667\text{ thousand dollars}.

At $d=100$. The two observations equal to 100 contribute zero, as do all observations below 100. The five observations strictly above 100 give payments $10,20,30,40,50$, with total $150$. Therefore,

\widehat{\Pi}(100)=\frac{150}{12}=12.5\text{ thousand dollars}.

Change check. For attachment points from 50 up to but not including 100, seven observations remain above the attachment point. Raising the attachment by 50 reduces each of those seven payments by 50, and observations at or below the attachment contribute zero. The transform therefore falls by $50(7/12)=350/12=29.1667$ thousand dollars. This agrees with $41.6667-12.5=29.1667$ thousand dollars. At $d=100$, only five observations strictly exceed the attachment; the two observations equal to 100 no longer count as exceedances.

The calculation also gives a useful decomposition. Let $N_d$ denote the number of observations strictly greater than $d$. If $N_d>0$, the average excess among those observations is $\frac{1}{N_d}\sum_{x_j>d}(x_j-d)$. Multiplying that conditional average by the empirical proportion exceeding $d$, $N_d/n$, gives the transform:

\widehat{\Pi}(d) =\frac{N_d}{n} \left(\frac{1}{N_d}\sum_{x_j>d}(x_j-d)\right).

This identity separates two influences on the average payment: how often the observed loss exceeds the attachment, and how large the excess is when it does. If $N_d=0$, the transform is zero; a conditional average over exceedances is then undefined and should not be reported.

Worked Example: Exceedance Proportion Times Conditional Excess

A hypothetical insurer records six simulated annual aggregate losses, in thousands of dollars: $0,40,80,120,160,200$. For an attachment of $d=100$, calculate the transform both directly and using the exceedance decomposition.

Direct calculation. The payments are $0,0,0,20,60,100$. Their sum is $180$, and the sample size is six, so

\widehat{\Pi}(100)=\frac{180}{6}=30\text{ thousand dollars}.

Decomposition check. Three of the six losses strictly exceed 100, so the empirical exceedance proportion is $3/6=0.5$. Their excesses are $20,60,100$, whose mean is $(20+60+100)/3=60$ thousand dollars. The product is $0.5(60)=30$ thousand dollars, matching the direct calculation.

Common mistake avoided. Reporting 60 thousand dollars as the stop-loss transform would report the average excess conditional on exceedance, not the average payment across all six observations. The factor $3/6$ is required for the empirical stop-loss transform.

Why the Transform Is Piecewise Linear

Between observed loss values, the set of observations exceeding $d$ stays fixed. Increasing $d$ by a small amount reduces each positive payment by that amount, while the other payments remain zero. This explains both the linear segments of the transform and why its slope changes at observed loss values.

Result: For $d_1<d_2$, the empirical transform satisfies $$ \widehat{\Pi}(d_1)-\widehat{\Pi}(d_2) =\frac{1}{n}\sum_{j=1}^{n} \left[(x_j-d_1)^+-(x_j-d_2)^+\right]. $$ If no observation lies between $d_1$ and $d_2$, inclusive of $d_2$ but exclusive of $d_1$, then the observations strictly above $d_2$ each reduce their payment by $d_2-d_1$. In that case, $$ \widehat{\Pi}(d_1)-\widehat{\Pi}(d_2) =(d_2-d_1)\frac{\#\{j:x_j>d_2\}}{n}. $$

In particular, within an interval containing no observed loss values, the slope is minus the proportion of observations above the attachment. At an observed loss value the transform remains continuous, but the slope changes: observations equal to the attachment stop contributing as soon as the attachment reaches that value. The transform is nonincreasing and convex, since its slope is nonpositive and becomes less negative as the attachment increases.

Tail-Integral Representation

The empirical distribution function is $\widehat{F}(t)=\frac{1}{n}\#\{j:x_j\leq t\}$. Its empirical survival function is $\widehat{S}(t)=1-\widehat{F}(t)=\frac{1}{n}\#\{j:x_j>t\}$. The strict inequality in this survival function is consistent with an observation exactly at $t$ making no excess above $t$.

For any fixed observed loss $x_j$, the positive part $(x_j-d)^+$ is the length of the interval of thresholds $t\geq d$ for which $x_j>t$. Equivalently, it is the integral of an indicator that is one precisely while the threshold remains below the observed loss. Averaging these identities over the observations gives the empirical tail-integral formula:

\begin{aligned} \widehat{\Pi}(d) &=\frac{1}{n}\sum_{j=1}^{n}(x_j-d)^+\\ &=\frac{1}{n}\sum_{j=1}^{n}\int_d^\infty \mathbf{1}_{\{x_j>t\}}\,dt\\ &=\int_d^\infty \widehat{S}(t)\,dt. \end{aligned}

The last equality follows by interchanging a finite sum and an integral: at each $t$, the fraction of indicators equal to one is exactly $\widehat{S}(t)$. For finite data, the integrand is zero above the largest observed loss, so the integral is finite. This representation makes the connection to the empirical loss distribution precise: the stop-loss transform is the area under its empirical survival curve to the right of $d$.

Worked Example: Repeated Outcomes and a Change in Slope

A hypothetical set of eight aggregate-loss observations, in thousands of dollars, has the following frequencies: loss 0 occurs twice, loss 50 occurs once, loss 100 occurs three times, and loss 200 occurs twice. Calculate the transform at $d=75$ and $d=100$, then verify the change using the empirical exceedance proportion.

At $d=75$. The two zero losses and the loss of 50 contribute zero. Each of the three losses of 100 contributes $100-75=25$, for a total of $3(25)=75$. Each of the two losses of 200 contributes $200-75=125$, for a total of $2(125)=250$. Thus the total payment is $75+250=325$, and

\widehat{\Pi}(75)=\frac{325}{8}=40.625\text{ thousand dollars}.

At $d=100$. The three observations equal to 100 contribute zero. Each of the two observations of 200 contributes $200-100=100$, for a total of 200. Therefore,

\widehat{\Pi}(100)=\frac{200}{8}=25\text{ thousand dollars}.

Change check. For $75\leq d<100$, five observations strictly exceed $d$: the three losses of 100 and the two losses of 200. Raising the attachment from 75 to 100 reduces each of those five payments by 25. The resulting decrease is $25(5/8)=15.625$ thousand dollars, which agrees with $40.625-25=15.625$. At exactly $d=100$, only the two losses of 200 strictly exceed the attachment.

Common mistake avoided. Counting only distinct loss values would incorrectly treat the three observations of 100 as one observation. Empirical probabilities and averages use the full frequency of every repeated outcome.

Practical Interpretation and Exam Traps

The empirical transform is a descriptive calculation under the empirical distribution. If the observations are simulated from a fitted aggregate-loss model, the result is the average simulated excess for those particular simulations; it does not eliminate simulation error. If the observations are historical losses, the empirical calculation likewise does not automatically adjust for changes in exposure, inflation, or claim conditions. Those adjustments must be addressed before treating observations as comparable.

When working with empirical stop-loss values, check the following points:

Use the correct denominator. Divide the total of all excess payments by $n$, including zero-payment observations.
Use a strict exceedance count. An observation equal to $d$ contributes zero and is not counted in $\#\{j:x_j>d\}$.
Keep repeated values. Each observation contributes once, even when several observations have the same loss.
Distinguish unconditional and conditional averages. The empirical transform includes zero payments; the mean excess among exceedances does not.
Track units. If losses are recorded in thousands of dollars, the transform is also in thousands of dollars until converted.

As a final check, the transform should never increase when the attachment increases: for every observation, $(x_j-d)^+$ is nonincreasing in $d$. Also, the transform must be zero whenever $d$ is at least the largest observed loss. These checks catch many arithmetic and threshold errors.

Key takeaway: The empirical stop-loss transform is the sample average of $(x_j-d)^+$. Its value combines the empirical probability of exceeding the attachment with the average excess conditional on exceedance, and its tail-integral form is the area under the empirical survival function beyond that attachment. This treatment is aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.

Check Your Understanding

For each problem, use the empirical distribution that assigns equal weight to every listed observation, including repeated values.

A hypothetical set of aggregate losses is $15,35,65,95$, in thousands of dollars. Calculate the empirical stop-loss transform at $d=50$. (Answer: 15 thousand dollars)
Six simulated losses are $0,40,80,120,160,200$, in thousands of dollars. At $d=100$, give the empirical exceedance proportion, the mean excess conditional on exceedance, and the empirical stop-loss transform. (Answer: $0.5$, 60 thousand dollars, and 30 thousand dollars)
Eight hypothetical outcomes have frequencies $0:2,\ 50:1,\ 100:3,\ 200:2$, where losses are in thousands of dollars. Calculate the transform at $d=75$. (Answer: 40.625 thousand dollars)
For the observations $20,60,60,100,180$, in thousands of dollars, calculate the transform at $d=60$. State how many observations strictly exceed the attachment. (Answer: 32 thousand dollars; 2 observations)
Using the same eight outcomes in Question 3, calculate the transform at $d=100$, then verify the decrease from the value at $d=75$ using the number of observations strictly above 100. (Answer: 25 thousand dollars; the decrease is $15.625$ thousand dollars, equal to $25(5/8)$)

What You'll Learn