Evaluating Modification Caps and Floors

Why Evaluate Modification Caps and Floors?

An uncapped experience modification can vary widely, especially when a risk has limited or volatile experience. A plan may limit the modification to a specified range: a floor prevents it from falling below a minimum, while a cap prevents it from rising above a maximum. These rules can reduce sharp premium changes, but they also move the indicated modification away from the uncapped result. Evaluation therefore asks two questions: how much do the constraints change the result, and what does that change do to accuracy and stability?

The comparison should use the same risks, experience periods, and evaluation targets for the capped and uncapped results. Accuracy is assessed against a defined outcome, such as a later-period loss measure or an estimate of expected risk. Stability is assessed by comparing modifications across renewals for matched risks. These are different objectives: a modification can be more stable without being more accurate.

Definition: Let $M$ be an uncapped indicated modification, $L$ the floor, and $U$ the cap, where $L \leq U$. The capped modification is $C(M)=\min(U,\max(L,M))$. The floor binds when $M<L$, and the cap binds when $M>U$. An indication equal to a boundary is not strictly constrained by that boundary.

When the base premium is $P_0$, the indicated premium is $P_0M$ without constraints and $P_0C(M)$ with constraints. If the base premium is unchanged, the percentage change in premium from one renewal to the next is the same as the percentage change in the modification. If base premiums change, compare actual premium changes separately; otherwise, a rate or exposure change can be mistaken for a modification effect.

Measures for Comparing Capped and Uncapped Results

For each evaluated risk $j$, let $T_j$ be the chosen benchmark: for example, a later-period observed loss measure or a sufficiently credible estimate of expected relative risk. Compare $M_j$ and $C(M_j)$ with the same $T_j$. Common summaries include mean error, mean absolute error, and root mean squared error. With weights $w_j$ that represent the evaluation basis, weighted root mean squared error is

\operatorname{RMSE}_w(\widehat{T}) = \sqrt{\frac{\sum_j w_j(\widehat{T}_j-T_j)^2}{\sum_j w_j}}, \qquad \widehat{T}_j \in \{M_j,C(M_j)\}.

A positive mean error indicates that predictions average above the benchmark; a negative mean error indicates that they average below it. RMSE gives greater influence to large errors because it squares them. The weights must match the question. Equal weights describe the average risk; base-premium or exposure weights can better reflect portfolio-level financial impact. State the weighting choice rather than assuming all summaries answer the same question.

The constrained frequency is the share of evaluated indications strictly outside the allowed range before clipping. If $n$ indications are evaluated, this is the number below $L$ or above $U$, divided by $n$. Report floor and cap frequencies separately as well as their combined frequency when the distinction matters. A high constraint frequency means the bounds affect many results; it does not, by itself, show whether accuracy improves.

Why Clipping Can Help or Hurt Forecast Accuracy

A useful justification follows from squared-error forecasting. Suppose the benchmark for one risk is its true expected relative cost $\theta$, and suppose $L \leq \theta \leq U$. The capped value is never farther from $\theta$ than the uncapped value. If $M$ is already between the boundaries, the values are equal. If $M<L$, then $\theta \geq L>M$, so moving $M$ up to $L$ reduces its distance to $\theta$. If $M>U$, then $\theta \leq U<M$, so moving $M$ down to $U$ also reduces its distance. Thus

(C(M)-\theta)^2 \leq (M-\theta)^2 \quad\text{when } L\leq\theta\leq U.

This is a pointwise result about squared error when the target lies within the permitted range. If the true expected cost lies outside that range, clipping is not guaranteed to help: the boundary may move the prediction away from the target. When the benchmark is a noisy future observation $Y$ with conditional mean $\theta$, the identity $E[(\widehat{T}-Y)^2]=\operatorname{Var}(Y)+(\widehat{T}-\theta)^2$ shows why comparing forecasts against future observations can still be informative in expectation, provided the same target and evaluation design are used. Observed results also contain process variation, so one portfolio’s measured difference should not automatically be treated as a permanent effect.

Worked Example: Comparing Accuracy Against Expected Risk

A hypothetical plan evaluates five risks. The benchmark $T$ is each risk’s expected relative cost, expressed as a modification. The uncapped indications are $0.55, 0.80, 1.00, 1.25, 1.70$; the benchmarks are $0.70, 0.85, 1.00, 1.20, 1.40$. The floor is $0.75$ and the cap is $1.35$. Compare the modifications and their unweighted RMSEs.

Method. Apply $C(M)=\min(1.35,\max(0.75,M))$ to each indication. Then subtract the benchmark from each forecast, square the errors, average them, and take the square root. This comparison avoids treating the uncapped indications themselves as the accuracy target.

The capped indications are $0.75, 0.80, 1.00, 1.25, 1.35$. The uncapped errors, $M-T$, are $-0.15,-0.05,0,0.05,0.30$. Their squared errors are $0.0225,0.0025,0,0.0025,0.0900$, with sum $0.1175$. Therefore the uncapped mean squared error is $0.1175/5=0.0235$, and its RMSE is $\sqrt{0.0235}=0.1533$ to four decimal places.

The capped errors, $C(M)-T$, are $0.05,-0.05,0,0.05,-0.05$. Their squared errors are $0.0025,0.0025,0,0.0025,0.0025$, with sum $0.0100$. The capped mean squared error is $0.0100/5=0.0020$, and its RMSE is $\sqrt{0.0020}=0.0447$ to four decimal places. In this hypothetical comparison, capping and flooring reduce the observed RMSE. The result is consistent with the benchmarks lying within the permitted range; it does not establish that these bounds would improve every portfolio.

Check. The first indication is below the floor and moves from $0.55$ to $0.75$; the last is above the cap and moves from $1.70$ to $1.35$. Both benchmarks are inside the permitted range, and the resulting errors have smaller absolute values than their uncapped counterparts.

Renewal Stability and Constraint Frequency

A direct stability comparison uses matched risks observed at consecutive renewals. Calculate each risk’s absolute modification change before and after applying the same boundaries, then summarize those changes—for example, with their mean or root mean squared change. A lower capped change indicates less movement in the modification, but does not show that the capped modification is better calibrated or more adequate.

The binding frequency helps explain the stability result. If only a small share of indications is outside the range, constraints may have little portfolio-wide effect even if they matter greatly for a few risks. If many indications are constrained, the capped and uncapped distributions may differ substantially. Report the floor and cap separately: frequent floor binding and frequent cap binding have different implications for which risks are being affected.

Worked Example: Measuring Renewal Stability

For six hypothetical risks with unchanged base premiums, the prior-renewal uncapped modifications are $0.60,0.80,1.00,1.20,1.35,1.50$. The current-renewal indications are $0.65,0.75,1.10,1.35,1.65,1.80$. The floor is $0.70$, and the cap is $1.40$. Find the mean absolute change in the modification for each renewal comparison and the share of indications constrained in each period.

Method. First cap each period’s indication, then calculate matched-risk absolute changes. Count an indication as constrained only if it is strictly below $0.70$ or strictly above $1.40$; an indication equal to $1.40$ is not counted as above the cap.

The prior capped modifications are $0.70,0.80,1.00,1.20,1.35,1.40$. The current capped modifications are $0.70,0.75,1.10,1.35,1.40,1.40$. The uncapped absolute changes are $0.05,0.05,0.10,0.15,0.30,0.30$. Their sum is $0.95$, so the mean is $0.95/6=0.1583$ to four decimal places.

The capped absolute changes are $0,0.05,0.10,0.15,0.05,0$. Their sum is $0.35$, so the mean is $0.35/6=0.0583$ to four decimal places. The constraints reduce the mean absolute movement by $0.1583-0.0583=0.1000$ modification points, using the rounded means.

In the prior period, $0.60$ is below the floor and $1.50$ is above the cap: $2/6=33.3\%$ are constrained. In the current period, $0.65$ is below the floor, while $1.65$ and $1.80$ are above the cap: $3/6=50.0\%$ are constrained. The current value $1.40$ equals the cap and is not counted. With base premiums unchanged, the percentage movement in premium for each risk matches its percentage movement in modification; the mean absolute modification change calculated here is not itself a mean percentage premium change.

Portfolio Weights and Premium Impact

Unweighted forecast accuracy can give a small-base-premium risk the same influence as a large one. A weighted measure can better reflect financial impact when the weights are appropriate to the decision being evaluated. The following example uses base premium as the weight and separately calculates total indicated premium. The weighted RMSE is still an error measure in modification points; it is not a dollar error.

Worked Example: Weighted Accuracy and Total Premium

Three hypothetical risks have base premiums of $ \$10{,}000$, $ \$20{,}000$, and $ \$30{,}000$. Their uncapped modifications are $0.50,1.20,1.80$, and their expected relative-cost benchmarks are $0.60,1.00,1.50$. Apply a $0.70$ floor and a $1.40$ cap. Calculate the base-premium-weighted RMSE and total indicated premium before and after constraints.

Method. The capped modifications are $0.70,1.20,1.40$. Use the base premiums as weights for the weighted RMSE, equivalently using proportional weights $1,2,3$. Calculate each indicated premium as base premium times modification.

The uncapped errors are $0.50-0.60=-0.10$, $1.20-1.00=0.20$, and $1.80-1.50=0.30$. The weighted squared-error sum using weights $1,2,3$ is $1(0.10^2)+2(0.20^2)+3(0.30^2)=0.01+0.08+0.27=0.36$. The total weight is $1+2+3=6$, so the weighted mean squared error is $0.36/6=0.06$ and the weighted RMSE is $\sqrt{0.06}=0.2449$ to four decimal places.

The capped errors are $0.70-0.60=0.10$, $1.20-1.00=0.20$, and $1.40-1.50=-0.10$. The weighted squared-error sum is $1(0.10^2)+2(0.20^2)+3(0.10^2)=0.01+0.08+0.03=0.12$. The weighted mean squared error is $0.12/6=0.02$, and the weighted RMSE is $\sqrt{0.02}=0.1414$ to four decimal places.

The uncapped indicated premiums are $ \$10{,}000(0.50)=\$5{,}000$, $ \$20{,}000(1.20)=\$24{,}000$, and $ \$30{,}000(1.80)=\$54{,}000$, totaling $ \$83{,}000$. The capped indications are $ \$7{,}000$, $ \$24{,}000$, and $ \$42{,}000$, totaling $ \$73{,}000$. The benchmark premiums total $ \$10{,}000(0.60)+\$20{,}000(1.00)+\$30{,}000(1.50)=\$6{,}000+\$20{,}000+\$45{,}000=\$71{,}000$. Thus the uncapped total is $ \$12{,}000$ above the benchmark total, and the capped total is $ \$2{,}000$ above it. The lower weighted RMSE and smaller aggregate difference are results for these supplied data, not a general guarantee about caps and floors.

Common Mistakes and Practical Interpretation

Comparing different targets. Accuracy comparisons are meaningful only when capped and uncapped forecasts are evaluated against the same benchmark, time period, and population.
Counting boundary values as constrained. With strict binding definitions, an indication exactly equal to $L$ or $U$ is at the boundary, not outside it.
Assuming more stability means more accuracy. A floor or cap can reduce renewal movement while increasing error for risks whose true expected cost lies beyond that boundary.
Ignoring the weighting basis. An unweighted average describes an average risk, while a premium-weighted result emphasizes larger base premiums. State which is being reported.
Applying a cap to a premium but not to its modification—or vice versa. When the base premium is fixed and the constraint is on the modification, calculate the premium using $P_0C(M)$, not $P_0M$.

Key takeaway: Evaluate a modification range by reporting accuracy against a stated benchmark, renewal stability for matched risks, and separate floor and cap binding frequencies. No single measure establishes that a range is preferable: the interpretation depends on the target, weighting, and whether the risks’ expected costs lie within the permitted range.

Check Your Understanding

For each problem, apply the stated boundaries before evaluating the requested measure.

An indication is $0.62$, with floor $0.75$ and cap $1.30$. What is the capped modification? (Answer: $0.75$)
Eight uncapped indications are $0.60,0.75,0.80,1.00,1.25,1.40,1.50,1.60$. The floor is $0.70$ and the cap is $1.50$. What share is strictly constrained? (Answer: $2/8=25.0\%$)
Four risks have prior indications $0.60,0.90,1.20,1.50$ and current indications $0.80,1.10,1.40,1.70$. With a $0.70$ floor and $1.40$ cap, calculate the mean absolute change in capped modifications. (Answer: $0.125$)
Two risks have benchmark modifications $0.80,1.30$, uncapped indications $0.60,1.60$, and weights $1,3$. Apply a $0.70$ floor and $1.40$ cap. What is the weighted RMSE of the capped indications? (Answer: $0.1000$)
A risk’s benchmark is $1.60$, its uncapped indication is $1.50$, and the cap is $1.40$. Compare squared errors before and after capping. (Answer: Uncapped squared error $0.01$; capped squared error $0.04$, so capping worsens squared error.)

What You'll Learn