From Rate Specification to a Monitoring Decision
A rate change is not complete when its indicated factors are approved or configured. The deployed system must calculate the intended premiums, production must be reconciled to a suitable benchmark, and early loss results must be interpreted with attention to exposure, mix, and claim maturity. A sound monitoring decision links those stages rather than treating a single unfavorable statistic as proof that the rate is wrong.
This workshop follows a hypothetical launch of a commercial equipment policy. The filed change applies a 1.08 overall rate-level factor. The examples use dollars, no investment income or discounting, and no mortality assumption: neither interest nor mortality is needed for this property-insurance monitoring case. Where a statistical model is used, its distribution and parameters are stated explicitly.
The case uses three linked questions: Did the system implement the specified rate? Do actual written and earned premiums reconcile to the expected premium for the risks written? Do mature-adjusted losses provide enough evidence to change the rate, or is the appropriate response to investigate and continue monitoring?
Worked Example: Translate the Filed Change Into System Tests
Worked Example: Translate the Filed Change Into System Tests
The hypothetical filing increases the rate level by 8%. A rating test uses two risk profiles. Profile A has a $1,000 base premium, a territory factor of 1.20, and a protection factor of 0.90. Profile B has a $1,500 base premium, a territory factor of 0.80, and a protection factor of 1.10. The rate-level factor applies to both profiles; all other factors remain unchanged. Compute each premium before and after the change, then check the weighted test total for 300 Profile A policies and 200 Profile B policies. State the method and a common mistake avoided.
Method. Calculate each premium as base premium multiplied by the applicable rating factors. Apply the new rate-level factor once to the otherwise calculated premium. This checks the rating specification at policy level as well as in aggregate.
For Profile A, the pre-change premium is:
The post-change premium is:
For Profile B, the pre-change premium is:
The post-change premium is:
For 300 Profile A policies, the pre-change total is \(300 \times \$1{,}080.00=\$324{,}000.00\), and the post-change total is \(300 \times \$1{,}166.40=\$349{,}920.00\). For 200 Profile B policies, the corresponding totals are \(200 \times \$1{,}320.00=\$264{,}000.00\) and \(200 \times \$1{,}425.60=\$285{,}120.00\).
The aggregate pre-change total is \(\$324{,}000+\$264{,}000=\$588{,}000\). The post-change total is \(\$349{,}920+\$285{,}120=\$635{,}040\). The ratio is \(\$635{,}040/\$588{,}000=1.08\), so the test totals agree with the specified 8% increase.
Check. Profile A's increase is \(\$86.40/\$1{,}080=8\%\); Profile B's is \(\$105.60/\$1{,}320=8\%\). Both policy calculations and the aggregate calculation pass. The common mistake is to add 0.08 to a rating factor, rather than multiply the premium by 1.08. For example, changing Profile A's territory factor from 1.20 to 1.28 would not apply the filed 8% change to the total premium.
A production test should also verify boundary conditions and rating-rule compliance, not just these arithmetic examples. In this case, the two profiles are test cases with known expected results; a production audit would compare the rating inputs, applied factors, effective date, and calculated premium for sampled policies with the filed rules.
Reconcile Production Premium Before Reading Loss Results
Written premium records premium on policies written or renewed during a period. Earned premium recognizes the portion attributable to coverage already provided. The deployment comparison should not assume that written premium is fully earned: policy effective dates and the observation date determine how much coverage has elapsed. When a cohort has a common effective date and a common elapsed fraction, the reconciliation is direct. With varied dates, the calculation is performed policy by policy or by an appropriate exposure-weighted method.
The identity follows because the cohort has the same fraction of each policy's written premium attributable to elapsed coverage. If policy \(j\) has written premium \(P_j\), its earned premium is \(eP_j\). Summing across policies gives \(\sum_j eP_j=e\sum_jP_j\). The shortcut does not apply when policies have different earned fractions unless \(e\) is replaced by each policy's own fraction.
Worked Example: Reconcile Written and Earned Premium
For the first six months of a hypothetical launch, the expected full-term premium for the risks written, under the new rate level, is $1,080,000. Those policies have a common earned fraction of 50% at the measurement date. Actual written premium is $1,048,000. Calculate expected and actual earned premium, their difference, and the percentage shortfall of actual earned premium relative to expected earned premium. State the method and a common mistake avoided.
Method. Apply the same 50% earned fraction to the expected full-term benchmark and to actual written premium. This puts the two premium measures on the same elapsed-coverage basis.
Expected earned premium is:
Actual earned premium is:
The actual-minus-expected difference is \(\$524{,}000-\$540{,}000=-\$16{,}000\). The shortfall as a fraction of expected earned premium is:
Rounded to the nearest dollar, actual earned premium is $524,000, which is $16,000 or 2.96% below the expected earned benchmark. The common mistake is to compare actual earned premium with expected written premium, which would mix a six-month measure with a full-term measure. The reconciliation flags a premium difference to investigate—such as policy mix, rating inputs, or transaction-level effects—but does not by itself identify its cause.
Compare Losses on a Consistent Basis
Loss results require their own benchmark. In the case below, the expected loss ratio at the new rate level is 70%, and the expected earned premium for the observed risks is $540,000. The rate-level expected loss amount is therefore $378,000. This is a benchmark amount for comparison, not a prediction that observed losses must equal it in a short period. Observed losses also require suitable treatment of claim maturity and risk mix; immature reported amounts can understate ultimate losses.
The actual earned premium of $524,000 differs from its $540,000 benchmark. For that reason, report both the actual loss ratio and the loss amount relative to the rate-level expected loss benchmark. The actual loss ratio uses actual earned premium. The relative loss measure compares mature-adjusted observed losses with expected losses at the rate level for the observed risks.
Worked Example: Interpret Early Loss Experience
At the same measurement date, reported incurred losses are $410,000. A stated claim-development adjustment estimates ultimate losses of $385,000 for this cohort. Use the expected loss ratio and expected premium benchmark above to calculate expected losses, the actual ultimate loss ratio, and ultimate losses as a percentage of expected losses. Then assess the result using a one-sided statistical alert. For this alert only, assume ultimate aggregate loss is approximately normally distributed with mean $378,000 and standard deviation $60,000, so its variance is \((\$60{,}000)^2\). Use a 5% upper-tail threshold, with standard normal critical value 1.645. State the method and a common mistake avoided.
Method. First calculate the expected loss amount on the benchmark earned-premium basis. Then calculate actual ultimate loss ratio using actual earned premium, and compare ultimate loss dollars with expected loss dollars. Finally, evaluate whether the deviation exceeds the stated statistical alert threshold.
Expected losses are:
The actual ultimate loss ratio is based on the actual earned premium of $524,000:
The amount of ultimate losses above expected is \(\$385{,}000-\$378{,}000=\$7{,}000\). Relative to expected losses, the result is:
Thus, observed ultimate losses are 1.85% above the rate-level expected amount. For the specified normal alert model, the upper threshold is:
The observed $385,000 is below $476,700. Equivalently, the standardized deviation is \((\$385{,}000-\$378{,}000)/\$60{,}000=0.1167\), far below 1.645. The one-sided alert is not triggered under the stated model. The common mistake is to call 73.47% the loss ratio against the expected benchmark: it is the actual loss ratio on actual earned premium. The benchmark loss ratio is 70%, while the separate loss-amount comparison is $385,000 versus $378,000. These measures answer different questions and should be labelled accordingly.
This alert conclusion is conditional on the assumed normal model, mean, standard deviation, and data adjustments. It is not proof that the rate is adequate, nor does it rule out a premium implementation issue. A premium shortfall and a loss result can have different causes and should be investigated separately.
Derive the Decision, Not Just the Statistic
A monitoring statistic becomes useful when it is connected to a decision rule and an operational follow-up. The case results support a restrained conclusion: the policy-level rate tests pass; actual earned premium is 2.96% below the expected benchmark; and mature-adjusted losses are only 1.85% above expected and do not trigger the specified statistical alert. Those findings do not justify an immediate portfolio-wide rate change. They do justify reconciling the premium shortfall and continuing observation.
The alert threshold has a simple justification. If \(L\) is ultimate aggregate loss and the stated model is \(L\sim N(\$378{,}000,(\$60{,}000)^2)\), standardization gives \(Z=(L-\$378{,}000)/\$60{,}000\), with \(Z\sim N(0,1)\). A one-sided 5% upper-tail rule flags results above the 95th percentile of this standard normal distribution, \(Z=1.645\). Solving \(Z>1.645\) for \(L\) gives \(L>\$378{,}000+1.645(\$60{,}000)=\$476{,}700\). This derivation explains why the threshold depends on both the expected amount and the assumed variability; a fixed dollar threshold would not have the same interpretation across cohorts of different sizes.
In practice, the standard deviation must be supportable for the specific cohort and period, and model assumptions should be reviewed. If the amount of data is small, losses are strongly skewed, or the cohort is heterogeneous, a normal approximation may be unsuitable. An alert is a screening rule, not a substitute for checking exposure, mix, development, and data quality.
Compare policy-level system calculations with the filed factors and effective-date rules. A failed test calls for an operational investigation, not an unsupported change to the indicated rate.
Compare written and earned premium with the expected benchmark on matching exposure and elapsed-coverage bases. Investigate the $16,000 difference before attributing loss results to rate adequacy.
Use the stated ultimate-loss estimate and rate-level expected losses. Review the alert assumptions and whether the cohort is sufficiently comparable for the result to be informative.
For this hypothetical case, retain the deployed rate, investigate premium reconciliation, and continue monitoring under the established observation window. Escalate only if subsequent, sufficiently mature and comparable results meet the pre-defined evidence threshold.
Common Deployment and Monitoring Traps
- Using the wrong premium basis. Comparing earned premium with a full-term written benchmark creates a false gap. Align the earning period or calculate the appropriate benchmark earned premium.
- Treating a loss ratio as a loss-cost comparison. State whether a ratio uses actual or expected earned premium, and compare loss dollars with the expected loss amount when that is the question.
- Reacting to an unadjusted early loss total. Claim maturity can change the estimate. Use a stated development treatment and revisit it as data mature.
- Confusing operational and actuarial corrections. A system defect or unexplained premium variance calls for diagnosis and controlled correction; it does not automatically support a rate-level change.
- Over-interpreting a threshold. A result below one alert threshold is not proof of no problem. A result above it is a prompt for review, not automatic authorization for a production change.
Check Your Understanding
For each hypothetical case, work through the stated bases before deciding what the result means.
- A $800 base premium has a territory factor of 1.10 and a protection factor of 0.95. A new factor of 1.05 applies to the resulting premium. What is the new premium to the cent? (Answer: $877.80)
- A cohort has $720,000 written premium, and 60% is earned by the observation date. The expected written-premium benchmark for the same cohort is $750,000, with the same earned fraction. Calculate actual and expected earned premium and the actual-minus-expected difference. (Answer: $432,000 actual, $450,000 expected, and -$18,000)
- Expected losses are $315,000 on an expected earned-premium benchmark of $450,000. Actual earned premium is $480,000 and mature-adjusted actual losses are $336,000. Calculate the expected loss ratio, actual loss ratio, and actual losses as a percentage of expected losses. (Answer: 70.00%, 70.00%, and 106.67%, respectively)
- Assume aggregate losses are normally distributed with mean $250,000 and standard deviation $20,000. For a one-sided 5% alert, use critical value 1.645. Is an observed loss of $287,000 above the alert threshold? Show the threshold or standardized deviation. (Answer: Yes; the threshold is $282,900 and the standardized deviation is 1.85)
- Two segments have expected losses of $100,000 and $300,000, respectively; their mature-adjusted observed losses are $110,000 and $270,000. Calculate the combined observed-to-expected ratio and state one reason not to apply a uniform correction based only on that combined ratio. (Answer: 95.00%; the segments move in opposite directions, so their results should be investigated separately before considering a uniform correction)