Evaluating Primary and Excess Loss Components

Why Evaluate Primary and Excess Losses Separately?

An experience modification uses past losses to differentiate among risks, but losses of different sizes may contribute differently to its indications. A plan may distinguish frequent, relatively small claims from the portions of large claims above a specified split point. Evaluating these components separately helps reveal whether the resulting modification reflects repeatable differences between risks or responds too strongly to volatile large-loss experience.

The split is applied to each claim, not to the total losses for a policy period. Component definitions are straightforward; the evaluation requires more care. In particular, a high observed excess-loss ratio may reflect a genuine difference in expected losses, a small and noisy excess-loss denominator, or both. Component weights should therefore be considered alongside calibration, predictive performance, and the uncertainty of the resulting indications.

Definition: For a claim with loss $X$ and a per-claim split point $L>0$, the primary loss is $P=\min(X,L)$, and the excess loss is $E=(X-L)_+=\max(X-L,0)$. Thus $X=P+E$. For a group of claims, aggregate primary and excess losses by summing the corresponding per-claim components. The split point and loss basis must be consistent with the experience plan being evaluated.

The word “excess” here means the portion of an individual loss above the split point; it does not mean the entire amount of a claim that exceeds the point. A claim of $12,000 with a $10,000 split contributes $10,000 to primary losses and $2,000 to excess losses. A claim below the split contributes no excess loss.

Component Ratios and Weighted Indications

For a risk or group, let $A_P$ and $A_E$ be observed aggregate primary and excess losses. Let $Q_P$ and $Q_E$ be the corresponding expected losses on the same exposure, coverage, trend, and loss-cost basis. When both expected amounts are positive, define the component loss ratios $r_P=A_P/Q_P$ and $r_E=A_E/Q_E$. A ratio above 1 indicates observed losses above the corresponding expectation; a ratio below 1 indicates observed losses below expectation.

A plan or evaluation can combine these ratios using weights $w_P$ and $w_E$, where $w_P\geq0$, $w_E\geq0$, and $w_P+w_E=1$. The resulting component indication is $I=w_Pr_P+w_Er_E$. This is a diagnostic representation of how the two component ratios contribute. It is not a universal experience-rating formula: an actual plan may apply credibility, ballast, caps, floors, or other provisions, and its exact formula must be evaluated as written.

Key distinction: Component weights describe how strongly the indication responds to each component ratio. They are not automatically equal to the shares of expected loss. If weights are intended to represent expected-loss shares, a common choice is $w_P=Q_P/(Q_P+Q_E)$ and $w_E=Q_E/(Q_P+Q_E)$. A plan may instead prescribe different weights, which should be identified rather than silently replaced.

Worked Example: Calculating a Weighted Component Indication

A hypothetical risk has a per-claim split point of $8,000 and four claims of $3,000, $7,000, $12,000, and $20,000. The expected aggregate primary loss is $32,000 and expected aggregate excess loss is $8,000. The evaluation uses weights of 0.75 on the primary ratio and 0.25 on the excess ratio. Calculate the components, ratios, and indication.

Method and component calculation. Apply the split separately to each claim. The primary amounts are $3,000, $7,000, $8,000, and $8,000, so $A_P=3{,}000+7{,}000+8{,}000+8{,}000=\$26{,}000$. The excess amounts are $0, $0, $4,000, and $12,000, so $A_E=0+0+4{,}000+12{,}000=\$16{,}000$. The component totals sum to $42,000, matching the original claims: $26{,}000+16{,}000=42{,}000$.

Ratios. The primary ratio is $r_P=26{,}000/32{,}000=0.8125$. The excess ratio is $r_E=16{,}000/8{,}000=2.0000$.

Weighted indication. The indication is $I=0.75(0.8125)+0.25(2.0000)=0.609375+0.500000=1.109375$. Rounded only at the final step, to four decimal places, $I=1.1094$, or 110.94% when expressed as a percentage.

As a check, the unweighted total observed-to-expected loss ratio is $42{,}000/(32{,}000+8{,}000)=1.0500$. It differs from the weighted indication because the stated component weights are not the expected-loss shares: those shares are $32/40=0.80$ primary and $8/40=0.20$ excess. Common mistake avoided: using the total loss ratio in place of the specified weighted component indication, or assigning a whole large claim to excess rather than splitting it at $8,000.

Why Component Weights Affect Volatility

A component can have a large ratio and still provide an unstable indication. Excess losses are often concentrated in fewer claims, so their ratios may vary substantially from one experience period to another. The relevant question is not simply whether excess losses are large, but how much uncertainty each component contributes after its weight is applied. The following derivation gives one way to quantify that uncertainty under a clearly stated model.

Assume the number of claims $N$ follows a Poisson distribution with mean $\lambda$, and claim severities are independent and identically distributed, independent of $N$. For each claim, let its primary and excess amounts be $P$ and $E$, with finite second moments. Aggregate losses are $A_P=\sum_{j=1}^{N}P_j$ and $A_E=\sum_{j=1}^{N}E_j$. Conditional on $N$, the component sums have variances $N\operatorname{Var}(P)$ and $N\operatorname{Var}(E)$; their conditional covariance is $N\operatorname{Cov}(P,E)$. Applying the law of total variance and covariance, including variation in $N$, gives:

\begin{aligned} E[A_P]&=\lambda E[P], & E[A_E]&=\lambda E[E],\\ \operatorname{Var}(A_P)&=\lambda E[P^2], & \operatorname{Var}(A_E)&=\lambda E[E^2],\\ \operatorname{Cov}(A_P,A_E)&=\lambda E[PE]. \end{aligned}

For example, $\operatorname{Var}(A_P)=E[\operatorname{Var}(A_P\mid N)]+\operatorname{Var}(E[A_P\mid N])$. The first term is $\lambda\operatorname{Var}(P)$, and the second is $\operatorname{Var}(N(E[P]))=\lambda(E[P])^2$. Their sum is $\lambda E[P^2]$. The same argument gives the excess variance. For the covariance, the conditional covariance contributes $\lambda\operatorname{Cov}(P,E)$, while variation in the conditional means contributes $\lambda E[P]E[E]$; together these equal $\lambda E[PE]$. This covariance need not be zero: an excess loss can occur only on a claim whose primary component is already at the split point.

Worked Example: Quantifying Component Volatility

Consider a hypothetical portfolio with $N\sim\operatorname{Poisson}(10)$ claims and independent claim severities. In thousands of dollars, a severity is $2 with probability 0.50, $8 with probability 0.30, and $20 with probability 0.20. Use a $10,000 per-claim split. Calculate the expected component totals, their variances and covariance, and the variance of an indication weighted by expected-loss shares.

Method and per-claim values. The primary amounts $P$, in thousands of dollars, are 2, 8, and 10; the corresponding excess amounts $E$ are 0, 0, and 10. Therefore, $E[P]=0.50(2)+0.30(8)+0.20(10)=1+2.4+2=5.4$, and $E[E]=0.50(0)+0.30(0)+0.20(10)=2$. Expected aggregate losses are $E[A_P]=10(5.4)=54$ and $E[A_E]=10(2)=20$, both in thousands of dollars.

Variances and covariance. First, $E[P^2]=0.50(2^2)+0.30(8^2)+0.20(10^2)=2+19.2+20=41.2$. Thus $\operatorname{Var}(A_P)=10(41.2)=412$ in squared thousands of dollars. Next, $E[E^2]=0.20(10^2)=20$, so $\operatorname{Var}(A_E)=10(20)=200$. Finally, $E[PE]=0.20(10)(10)=20$, giving $\operatorname{Cov}(A_P,A_E)=10(20)=200$. The positive covariance reflects the fact that every excess claim also has a primary component of $10,000.

Variance of the weighted indication. The expected-loss-share weights are $w_P=54/(54+20)=27/37$ and $w_E=20/(54+20)=10/37$. The component ratios are $r_P=A_P/54$ and $r_E=A_E/20$. Their variances are $412/54^2=412/2916\approx0.1413$ and $200/20^2=0.5000$; their covariance is $200/(54\cdot20)=0.1852$. Consequently,

\begin{aligned} \operatorname{Var}(I) &=w_P^2\operatorname{Var}(r_P)+w_E^2\operatorname{Var}(r_E) +2w_Pw_E\operatorname{Cov}(r_P,r_E)\\ &=\left(\frac{27}{37}\right)^2\left(\frac{412}{2916}\right) +\left(\frac{10}{37}\right)^2(0.5) +2\left(\frac{27}{37}\right)\left(\frac{10}{37}\right) \left(\frac{200}{54\cdot20}\right)\\ &=\frac{103+50+100}{1369}=\frac{253}{1369}\approx0.1848. \end{aligned}

The standard deviation is $\sqrt{0.1848}\approx0.4299$, with the exact model giving an expected indication of 1.0000. These calculations show why it is important to include both component variances and their covariance. Common mistake avoided: treating primary and excess ratios as independent just because they are recorded separately, or using $\lambda\operatorname{Var}(P)$ instead of $\lambda E[P^2]$ for the aggregate variance when the claim count is Poisson.

Does the Weighting Produce Useful Differentiation?

A plan can be stable but fail to distinguish risks with meaningfully different expected outcomes. Conversely, a plan can produce widely separated indications largely because it gives substantial weight to a component with high sampling error. A useful evaluation therefore considers both the size of indicated differences and their uncertainty. It should also check component calibration: across comparable risks, observed primary and excess losses should be assessed against their corresponding expected losses, not only against total expected losses.

A practical comparison uses component ratios, the plan’s actual weights, and uncertainty estimates obtained on an appropriate validation basis. When comparing two independently evaluated risks, the uncertainty of the difference depends on the uncertainty of each risk’s component ratios and on any within-risk covariance. Independence is a simplifying assumption, not a default fact; shared environmental or catastrophe experience can make risks’ outcomes dependent.

Worked Example: Checking Whether Component Separation Is Reliable

A hypothetical validation study estimates the following component ratios: Risk A has $r_P=0.95$ and $r_E=1.60$; Risk B has $r_P=1.05$ and $r_E=0.40$. The evaluation weights are 0.75 primary and 0.25 excess. Each risk’s estimated primary ratio has standard error 0.10 and excess ratio has standard error 0.60. Assume the two risks are independent, and within each risk the primary- and excess-ratio estimates have zero covariance. Calculate the weighted indications and assess their separation relative to its estimated standard error.

Method and indications. For Risk A, $I_A=0.75(0.95)+0.25(1.60)=0.7125+0.4000=1.1125$. For Risk B, $I_B=0.75(1.05)+0.25(0.40)=0.7875+0.1000=0.8875$. The indicated difference is $I_A-I_B=1.1125-0.8875=0.2250$.

Uncertainty of the difference. Independence between risks means the variance of the difference in primary ratios is $0.10^2+0.10^2=0.0200$; for excess ratios it is $0.60^2+0.60^2=0.7200$. With the stated zero covariance assumption, the variance of the weighted difference is $0.75^2(0.0200)+0.25^2(0.7200)=0.01125+0.04500=0.05625$. Its standard error is $\sqrt{0.05625}\approx0.2372$.

The difference divided by its standard error is $0.2250/0.2372\approx0.95$. Thus the point indications differ, but this comparison alone does not show a clear separation relative to the stated uncertainty. Common mistake avoided: interpreting a large difference in excess ratios as strong differentiation without accounting for the high excess-ratio standard errors. This is a diagnostic calculation, not a universal pass-or-fail threshold; a full evaluation should use suitable validation data and account for dependence when present.

Evaluation Checks and Common Traps

The split point defines what is measured; the weights define how the component results enter the indication. Evaluating both requires consistent data and a clear statement of the plan’s formula. For a useful comparison:

Apply the split to each claim before aggregating. Confirm that primary plus excess equals the original loss for every claim.
Match observed and expected component amounts on exposure, valuation date, trend, coverage, and loss basis. A ratio with mismatched denominators is not an interpretable component comparison.
Report the weights actually used. Do not assume they equal expected-loss shares unless that is how the plan defines them.
Assess volatility using the component variances and, when relevant, their covariance. Rare excess claims can make an excess ratio unstable, especially when its expected denominator is small.
Compare differentiation and predictive performance on appropriate validation data. A component’s historical contribution to an indication is not, by itself, evidence that it improves predictions.

A frequent error is to assign the entire amount of every claim above the split point to excess. That changes the definition: excess is only the amount above the point, while the primary component for that claim remains at the split point. Another error is to interpret a weight as a measure of credibility without checking the plan’s formula. Weights can affect responsiveness, but they do not by themselves establish how reliable the underlying experience is.

Key takeaway: Evaluate primary and excess components together, but do not conflate their roles. Split each claim correctly, compare each observed component with its matching expectation, apply the stated weights, and judge the resulting differentiation in light of component uncertainty and validation performance.

Check Your Understanding

For each problem, calculate the requested component or indication before interpreting it.

A hypothetical policy has a $6,000 split point and claims of $2,000, $9,000, and $15,000. Calculate aggregate primary and excess losses. (Answer: primary $14,000; excess $12,000.)
A risk has expected primary and excess losses of $24,000 and $16,000, and actual amounts of $30,000 and $8,000. With weights 0.60 and 0.40, calculate the weighted component indication. (Answer: 0.9500.)
For a claim, primary loss is $7,000 and excess loss is $5,000. What is the original loss, and what would be wrong with recording $12,000 as excess? (Answer: original loss $12,000; excess is only $5,000, because the $7,000 primary portion must remain in primary.)
Two independent risk indications have a difference of 0.18 and an estimated standard error for that difference of 0.24. Calculate the difference in standard-error units and state what it suggests. (Answer: 0.75 standard errors; the point difference is not large relative to the stated uncertainty.)
A plan weights primary and excess ratios by 0.80 and 0.20. Expected losses are $60,000 primary and $40,000 excess. Are the plan weights the expected-loss shares? Give the expected-loss shares and explain the distinction. (Answer: no; the expected-loss shares are 0.60 primary and 0.40 excess, so the plan places less weight on excess than its expected-loss share.)

What You'll Learn