Exposure-Weighted and Equal-Weight Segment Summaries

Why the Weighting Rule Matters

A portfolio may contain segments of very different sizes: for example, a large group of standard risks and a small group of specialized risks. Each segment can have its own prediction error summary. Combining those summaries requires a choice about how much influence each segment receives. That choice is not merely arithmetic: it determines whether the result describes error across the portfolio's exposure or gives each segment the same influence.

This tutorial focuses on combining segment-level error summaries. It assumes that the predictions and outcomes being compared use a consistent target, scale, and evaluation period across the segments. The formulas below require no distributional assumption about the errors. The earlier material on mean squared error and root mean squared error supplies the underlying metrics; the key issue here is how segment results are aggregated.

Definition: Let a portfolio be divided into $J$ non-overlapping segments. For segment $j$, let $n_j$ be its exposure amount, $MSE_j$ its mean squared error, and $RMSE_j=\sqrt{MSE_j}$. An exposure-weighted pooled RMSE gives each segment influence in proportion to its exposure. An equal-weight segment summary gives every segment the same influence, regardless of its exposure.

Here, exposure could be a count of evaluated records or a specified amount such as policy-years. The interpretation depends on the unit used to define it. If every evaluated record has equal exposure, $n_j$ is simply the number of records in segment $j$. If exposures differ across records, the same logic applies using the stated record-level exposure weights.

Exposure-Weighted Pooled Error

Suppose $MSE_j$ is calculated using the observations in segment $j$, and $n_j$ is the number of equally weighted observations in that segment. The pooled MSE is the sum of squared errors across all segments divided by the total number of observations. Because the segment sum of squared errors is $n_jMSE_j$, the pooled RMSE is:

MSE_{\text{pooled}}=\frac{\sum_{j=1}^{J} n_j MSE_j}{\sum_{j=1}^{J} n_j}, \qquad RMSE_{\text{pooled}}=\sqrt{\frac{\sum_{j=1}^{J} n_j MSE_j}{\sum_{j=1}^{J} n_j}}.

The exposure weights apply to the segment MSEs inside the square root. In general, this is not the same as taking an exposure-weighted average of the segment RMSEs. Squaring and taking a square root are nonlinear operations, so the order matters.

Justification: Let $e_{ij}$ be the prediction error for observation $i$ in segment $j$. The segment MSE is $MSE_j=(1/n_j)\sum_{i=1}^{n_j}e_{ij}^2$, so its sum of squared errors is $\sum_{i=1}^{n_j}e_{ij}^2=n_jMSE_j$. The pooled MSE divides the sum of squared errors over all segments by the total number of observations: $$ MSE_{\text{pooled}} =\frac{\sum_{j=1}^{J}\sum_{i=1}^{n_j}e_{ij}^2}{\sum_{j=1}^{J}n_j} =\frac{\sum_{j=1}^{J}n_jMSE_j}{\sum_{j=1}^{J}n_j}. $$ Taking the nonnegative square root gives the pooled RMSE formula.

Worked Example: Pooling Errors From Two Segments

A hypothetical pricing team evaluates predictions for two segments. Segment A has four records with errors $-1, 1, 2,$ and $-2$. Segment B has two records with errors $0$ and $4$. Calculate each segment's MSE and RMSE, the pooled RMSE, and the equal-weight average of the segment RMSEs. Explain which result reflects the records collectively.

Method. Square each error to obtain the sum of squared errors for its segment, divide by the segment record count to get its MSE, and take the square root to get its RMSE.

For Segment A, the squared errors are $1, 1, 4,$ and $4$. Their sum is $1+1+4+4=10$. Therefore, $MSE_A=10/4=2.5$, and $RMSE_A=\sqrt{2.5}=1.5811$ to four decimal places.

For Segment B, the squared errors are $0$ and $16$. Their sum is $0+16=16$. Therefore, $MSE_B=16/2=8$, and $RMSE_B=\sqrt{8}=2.8284$ to four decimal places.

There are $4+2=6$ records in total. The pooled MSE is $(4(2.5)+2(8))/6=(10+16)/6=26/6=4.3333$. Thus the pooled RMSE is $\sqrt{26/6}=2.0817$ to four decimal places.

Giving each segment equal weight instead, the average segment RMSE is $(1.5811+2.8284)/2=2.20475$, or $2.2048$ to four decimal places. This is a segment-level summary, not the pooled RMSE. The pooled RMSE reflects error across the six records, so Segment A has twice the influence of Segment B. Common mistake avoided: averaging the two RMSEs equally and reporting that result as the portfolio RMSE would silently give the two segments equal influence despite their different record counts.

Check. The pooled squared-error total is $10+16=26$, matching the numerator used above. Since the pooled RMSE is the square root of $26/6$, its square is approximately $4.3333$, the pooled MSE.

Equal Weight for Each Segment

Sometimes the evaluation question is not “What is the typical error across all portfolio exposure?” but “How does performance look when each segment counts equally?” For instance, an analyst may want to prevent a very large segment from dominating a summary of segment performance. There is more than one reasonable equal-segment summary, so the chosen form should be named explicitly.

Definition: The equal-weight average segment RMSE is $(1/J)\sum_{j=1}^{J}RMSE_j$. The equal-weight root-mean segment MSE is $\sqrt{(1/J)\sum_{j=1}^{J}MSE_j}$. Both give each segment equal weight, but they are generally different because the square root is applied at different stages.

The average segment RMSE answers a direct question about the arithmetic average of segment-level RMSEs. The root-mean segment MSE first averages squared error across segments, then returns the result to the original error scale. Neither should be described simply as “the” equal-weight RMSE without saying which calculation was used.

Worked Example: Comparing Three Ways to Summarize Segment RMSEs

A hypothetical claims model has three segments with exposures 80, 20, and 10 records, and segment RMSEs of 2, 4, and 10 dollars, respectively. Calculate the exposure-weighted pooled RMSE, the equal-weight average segment RMSE, and the equal-weight root-mean segment MSE. Round to four decimal places.

Method. First square the segment RMSEs to recover the segment MSEs: $2^2=4$, $4^2=16$, and $10^2=100$ dollars squared.

The total exposure is $80+20+10=110$. The pooled MSE is $(80(4)+20(16)+10(100))/110=(320+320+1000)/110=1640/110=14.9091$. Therefore, the exposure-weighted pooled RMSE is $\sqrt{14.9091}=3.8612$ dollars.

For the equal-weight average segment RMSE, average the three RMSEs directly: $(2+4+10)/3=16/3=5.3333$ dollars.

For the equal-weight root-mean segment MSE, average the segment MSEs first: $(4+16+100)/3=120/3=40$ dollars squared. Taking the square root gives $\sqrt{40}=6.3249$ dollars.

The three results answer different questions: $3.8612$ dollars weights records by exposure; $5.3333$ dollars is the arithmetic average of segment RMSEs; and $6.3249$ dollars is the square root of the equally weighted average of segment MSEs.

Common mistake and correction. An exposure-weighted average of the RMSEs would be $(80(2)+20(4)+10(10))/110=(160+80+100)/110=340/110=3.0909$, to four decimal places. That arithmetic is correct, but $3.0909$ is not the pooled RMSE. The pooled calculation weights the segment MSEs and then takes a square root, giving $3.8612$ dollars. The difference arises because averaging RMSEs does not preserve the pooled sum of squared errors.

Check. The pooled result squared is approximately $3.8612^2=14.909$, which agrees, allowing for rounding, with the pooled MSE of $1640/110=14.9091$.

Using Exposure Amounts Rather Than Record Counts

Exposure need not be an integer count. Suppose the metric is defined using a positive exposure weight for each observation, such as a fraction of a policy-year. For record $i$ in segment $j$, let $w_{ij}$ be that weight. Define the segment weighted MSE as the weighted sum of squared errors divided by the segment's total weight. Then the pooled weighted MSE is a weighted average of the segment weighted MSEs, with each segment's total exposure as its weight.

MSE_j^{(w)} =\frac{\sum_i w_{ij}e_{ij}^2}{\sum_i w_{ij}}, \qquad MSE_{\text{pooled}}^{(w)} =\frac{\sum_{j=1}^{J}W_jMSE_j^{(w)}}{\sum_{j=1}^{J}W_j}, \qquad W_j=\sum_i w_{ij}.

This formula assumes nonnegative exposure weights and a common definition of error across segments. A zero total weight contributes no defined segment weighted MSE and should not be treated as an ordinary segment in this calculation. If the segment metric was calculated with some other weighting rule, its reported MSE cannot automatically be combined using this formula; the aggregation weights and metric definitions must match.

Worked Example: Fractional Exposures and Equal Segment Influence

A hypothetical portfolio has three segments with total exposures of $0.5$, $1.5$, and $3.0$ policy-years. Their exposure-weighted segment MSEs are $9$, $4$, and $1$ dollars squared. Calculate the pooled exposure-weighted RMSE, the equal-weight average segment RMSE, and the equal-weight root-mean segment MSE.

Method. Weight each segment MSE by its total exposure for the pooled result. For the equal-weight summaries, first give all three segments weight $1/3$.

The total exposure is $0.5+1.5+3.0=5.0$ policy-years. The weighted sum of segment MSEs is $0.5(9)+1.5(4)+3.0(1)=4.5+6+3=13.5$. Thus the pooled MSE is $13.5/5.0=2.7$ dollars squared, and the pooled RMSE is $\sqrt{2.7}=1.6432$ dollars.

The segment RMSEs are $\sqrt{9}=3$, $\sqrt{4}=2$, and $\sqrt{1}=1$ dollars. Their equal-weight average is $(3+2+1)/3=2$ dollars.

The equal-weight root-mean segment MSE is $\sqrt{(9+4+1)/3}=\sqrt{14/3}=\sqrt{4.6667}=2.1602$ dollars.

Thus, exposure weighting gives $1.6432$ dollars, while the two equal-segment summaries give $2$ dollars and $2.1602$ dollars. The large-exposure segment has the smallest error, so it pulls the pooled result downward. Equal weighting deliberately gives the smaller segments more influence than their exposure would warrant in a portfolio-wide calculation.

Check. The weighted MSE is $2.7$, which lies between the smallest segment MSE, $1$, and the largest, $9$. Its square root, $1.6432$, is on the same dollar scale as the segment RMSEs.

Choosing and Reporting the Summary

The weight rule should follow the question being answered. Use exposure weighting when the target is performance across the evaluated portfolio or exposure base. Use an equal-weight segment summary when the target is a comparison of segments in which a small segment should count as much as a large one. If both questions matter, report both, label them clearly, and explain why the numbers differ.

For a squared-error-based metric, state whether the calculation averages segment MSEs and then takes a square root, or averages segment RMSEs directly. These quantities have the same units after the square root, but they are not interchangeable. Also state the exposure definition, evaluation population, and whether segment values are counts or weighted exposure totals.

Confirm the segment metrics.
Check that every segment uses the same outcome, prediction scale, error definition, and evaluation period.

Choose the target question.
For portfolio-wide error, use exposure weights. To give each segment equal influence, choose and name an equal-segment summary.

Combine at the correct stage.
For pooled RMSE, combine segment MSEs using exposure weights, then take the square root. Do not substitute a weighted average of segment RMSEs.

Report the interpretation.
Name the weighting rule and clarify whether the result describes records or exposure collectively, or treats segments equally.

A useful exam check is to ask what an individual “vote” in the summary represents. In an exposure-weighted result, each record or exposure unit contributes according to its weight. In an equal-segment result, each segment contributes one equal share, even when the segments have very different sizes.

Check Your Understanding

For each problem, identify the requested weighting rule and keep the MSE and RMSE calculations distinct. Answers are rounded to four decimal places where needed.

A hypothetical portfolio has two segments with 30 and 10 records and MSEs of 4 and 9 dollars squared. Calculate the pooled RMSE. (Answer: $\sqrt{(30(4)+10(9))/40}=\sqrt{5.25}=2.2913$ dollars.)
Two segments have exposures of 9 and 1 records and RMSEs of 3 and 5 dollars. Calculate both the pooled exposure-weighted RMSE and the equal-weight average segment RMSE. (Answer: pooled RMSE $=\sqrt{(9(3^2)+1(5^2))/10}=\sqrt{10.6}=3.2558$ dollars; equal-weight average segment RMSE $=(3+5)/2=4$ dollars.)
Three equally weighted segments have MSEs of 1, 9, and 16 dollars squared. Calculate the equal-weight average segment RMSE and the equal-weight root-mean segment MSE. (Answer: average segment RMSE $=(1+3+4)/3=2.6667$ dollars; root-mean segment MSE $=\sqrt{(1+9+16)/3}=\sqrt{26/3}=2.9439$ dollars.)
Three segments have exposures of 2, 1, and 1 policy-years and exposure-weighted MSEs of 1, 4, and 25 dollars squared. Calculate pooled RMSE and the equal-weight average of the segment RMSEs. (Answer: pooled RMSE $=\sqrt{(2(1)+1(4)+1(25))/4}=\sqrt{7.75}=2.7839$ dollars; equal-weight average segment RMSE $=(1+2+5)/3=2.6667$ dollars.)
A two-segment portfolio has exposures of $0.25$ and $0.75$ policy-years and weighted segment MSEs of 16 and 4 dollars squared. Calculate the pooled RMSE and the equal-weight root-mean segment MSE. State why they differ. (Answer: pooled RMSE $=\sqrt{(0.25(16)+0.75(4))/1}=\sqrt{7}=2.6458$ dollars; equal-weight root-mean segment MSE $=\sqrt{(16+4)/2}=\sqrt{10}=3.1623$ dollars; they differ because the first uses exposure weights and the second gives each segment equal weight.)

What You'll Learn