Why Define the Renewal Opportunity First?
A retention estimate is only meaningful when its unit, eligible population, and outcome are clear. “The renewal rate was 80%” is incomplete unless it tells us what could renew, when the prediction was made, and which records count as renewals. A customer may hold several policies, one policy may receive a new identifier at renewal, and a policy can end before its scheduled expiration. Each detail can change the count.
For ratemaking and demand analysis, renewal is usually a decision about continuing a particular policy’s coverage for a subsequent term. It is not automatically a decision about whether the customer stays with the insurer in any form. The definitions below make that distinction explicit and are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
The observation unit is not necessarily a customer, household, quote, or transaction. For a policy-level renewal analysis, it is one policy-term opportunity, commonly identified by the policy in force and its scheduled expiration date. If a policy has several transaction rows, those rows do not create several renewal opportunities. If the same insured has multiple policies, each policy can have its own opportunity.
Set the Prediction Date and Eligibility Rules
The prediction date must correspond to the decision the analysis is meant to support. For example, an insurer might want a retention estimate 45 days before expiration, when a renewal offer can still be changed. In that case, the model’s information set must be limited to information available by that date. A prediction made at a different point in the renewal process has a different information set and may answer a different business question.
Eligibility rules should establish, at minimum, the target line or product, the scheduled expiration period, the required policy status at the prediction date, and any exclusions. A useful record for each opportunity includes the policy identifier, scheduled expiration, prediction date, and eligibility status. These fields help distinguish an expiring policy from an arbitrary policy transaction or an account-level record.
A policy that ended before the prediction date is not an active expiring-policy opportunity under a rule requiring coverage to be in force at the prediction date. A policy still in force at the prediction date may, however, cancel before its scheduled expiration. That cancellation is not a term-end nonrenewal. Keep it distinct from the renewal outcome; if the target is renewal conditional on reaching the scheduled expiration, do not silently count that midterm cancellation as a nonrenewal.
This distinction affects the meaning of a calculated rate. A rate among policies that reach scheduled expiration describes renewal conditional on reaching that point. A rate among all policies in force at the prediction date would need a defined treatment for policies that cancel before expiration. State which target is intended rather than combining the two denominators.
Define the Outcome and the Policy-Level Unit
The continuity rule matters because a renewal may be issued under a new policy number. If the replacement identifier represents continuation of the same insured coverage at the term boundary, the renewal should be linked to the expiring opportunity rather than recorded as a loss merely because the identifier changed. Conversely, a customer’s other policy continuing does not, by itself, mean that the policy under study renewed.
The outcome is usually binary once it is known: renewed or did not renew at the term boundary. But the available records may not yet establish either outcome. An unresolved or incomplete record should remain unknown until its status can be determined under the stated rule. Coding every missing outcome as a nonrenewal creates a false negative label and can understate retention.
For a cohort of \(n\) eligible opportunities that reach the scheduled expiration and have resolved outcomes, let \(R_j=1\) if opportunity \(j\) renewed and \(R_j=0\) if it did not. The observed policy-level renewal rate is the number renewed divided by the number of resolved opportunities in that target population:
The denominator is not a minor bookkeeping choice. It expresses the population to which the rate applies. A policy-weighted rate gives each policy-term opportunity one count. A customer-weighted rate instead gives each customer one count under a separately defined customer outcome. The two rates can differ even when computed from the same book.
Why the Renewal-Rate Formula Works
For a fixed set of \(n\) resolved opportunities, each indicator \(R_j\) contributes exactly one to the numerator when the opportunity renewed and zero otherwise. Therefore, \(\sum_{j=1}^{n}R_j\) is precisely the number of renewals. Dividing that count by \(n\) gives the fraction of eligible, resolved opportunities that renewed.
The same expression estimates a probability when the opportunities are treated as observations from a defined population with renewal probability \(p\). If they have a common probability \(p\), then \(E[R_j]=p\), so:
Thus, the sample proportion is unbiased for the common renewal probability under that sampling model. The common-probability assumption is not necessary to calculate the observed cohort rate, but it is needed for this simple interpretation of its expectation. If renewal probabilities differ across opportunities, the average indicator still describes the observed policy-weighted proportion; it does not automatically represent a particular customer-weighted or future-cohort probability.
Worked Examples
Worked Example: Midterm Cancellations in an Expiring Cohort
A hypothetical insurer plans to estimate renewal 45 days before expiration. On February 14, it identifies 100 policies scheduled to expire on March 31. The eligibility rule requires the policy to be in force on February 14. Seven of the 100 had already cancelled before that date. After February 14, three more cancel effective before March 31. Of the policies that reach March 31, 72 continue into a new term and 18 do not. Calculate the policy-level renewal rate for policies reaching scheduled expiration, and identify the incorrect calculation that treats midterm cancellation as nonrenewal.
Method and cohort count. Start with the 100 scheduled expirations. Remove the seven policies already cancelled by the prediction date, because they do not meet the stated in-force eligibility rule: \(100-7=93\) eligible opportunities on February 14.
Resolve the term-boundary population. Three of those 93 cancel before March 31. They are midterm cancellations, not term-end nonrenewals. The number reaching scheduled expiration is \(93-3=90\).
Calculate and check. There are 72 renewals and 18 term-end nonrenewals, and \(72+18=90\), matching the denominator. Therefore:
The common mistake is to treat the three midterm cancellations as nonrenewals and calculate \(72/93=0.7742\), or 77.42%. That answers a different, incompletely specified question. Under the stated target, the denominator contains only policies that reach the scheduled expiration, and the checked renewal and nonrenewal counts sum to that denominator.
Worked Example: Policy-Level and Household-Level Retention
In a hypothetical renewal cohort, household H has two eligible expiring property policies and one eligible expiring auto policy. Two of H’s three policies renew. Household G has one eligible policy, which renews. Household J has four eligible policies; three renew and one does not. Assume every policy reaches expiration and has a resolved outcome. Calculate the policy-level renewal rate, then compare it with the proportion of households that retain at least one policy.
Policy-level calculation. There are \(3+1+4=8\) policy-term opportunities. Renewals total \(2+1+3=6\), and nonrenewals total \(1+0+1=2\). The check is \(6+2=8\). Thus the policy-level rate is:
Household-level calculation. H retains at least one policy, G retains its policy, and J retains at least one policy. Therefore, all three households retain at least one policy: \(3/3=1.0000\), or 100.00%.
These answers are both correct for their respective units. The household figure does not mean that 100% of policies renewed; it means that every household kept at least one policy. The common mistake is to report a customer or household result as if it were a policy-level renewal rate.
Worked Example: Successor Policy Numbers and the Prediction Snapshot
A hypothetical carrier sets May 1 as the prediction date for policies scheduled to expire on June 30. Five eligible policies are in force on May 1. At expiration, three continue into a new term, one does not, and one cancels effective June 10. One of the three continuing policies is issued under a new policy number, but records confirm it is continuous coverage for the same insured risk. Calculate renewal conditional on reaching expiration and explain which records can be used as May 1 predictors.
Link the renewal correctly. The successor number represents continuation of the expiring coverage, so count that opportunity as renewed. A number change alone does not create a nonrenewal. There are three term-boundary renewals in total.
Set the denominator. The policy cancelling on June 10 terminates before the scheduled expiration. It is a midterm cancellation, not a term-end nonrenewal. Four opportunities reach June 30: three renew and one does not. The counts reconcile as \(3+1=4\), so the conditional renewal rate is:
Respect the prediction date. A predictor must have been available on or before May 1. A fact first recorded after May 1, such as a later customer response to an offer, cannot be treated as information available at the prediction snapshot. The common mistakes are to count the June cancellation as a term-end nonrenewal, to miss the successor-number renewal, or to use post-May 1 information as if it were known at prediction time.
Practical Checks Before Reporting a Rate
Before interpreting a retention result, verify that every row represents one policy-term opportunity and that duplicates from transactions, quotes, or administrative updates have not inflated the count. Confirm that the prediction date and eligibility rules are applied consistently. Then reconcile the number of resolved opportunities against the distinct outcome categories, keeping unresolved and midterm-terminated records visible rather than forcing them into a binary label.
A concise definition should state the policy population, the prediction date, the scheduled expiration period, what counts as continuity, and how midterm cancellations and unresolved records are handled. These choices determine what the resulting rate describes; they are not interchangeable technical details.
- Do not substitute customer retention for policy renewal. A customer may renew one policy while allowing another to end.
- Do not equate identifier continuity with coverage continuity. A new policy number may represent a renewal if the defined linkage rule supports it.
- Do not label an early termination a term-end nonrenewal. The effective termination date must be compared with the scheduled expiration date.
- Do not turn an unknown outcome into zero by default. An unresolved status is not evidence of nonrenewal.
Check Your Understanding
For each problem, identify the relevant unit and apply the stated cohort and outcome rules.
- A hypothetical book has 80 policies scheduled to expire in September. Six are no longer in force on the prediction date. Of the remaining policies, two cancel before expiration; 54 of those reaching expiration renew and 18 do not. What is the renewal rate conditional on reaching expiration? (Answer: \(54/72=0.7500\), or 75.00%.)
- Three households have respectively 2, 3, and 1 eligible policies. Their numbers of renewed policies are 1, 2, and 1. What is the policy-level renewal rate? (Answer: \(4/6=0.6667\), or 66.67%.)
- A policy expires on October 31. A replacement policy number begins November 1 and records establish continuous coverage for the same insured risk. Under a continuity rule based on coverage rather than identifier, how is the expiring opportunity labelled? (Answer: Renewed.)
- A policy is in force at a prediction date 30 days before expiration but cancels effective 10 days before its scheduled expiration. In a target defined as term-end renewal conditional on reaching expiration, should it be counted as a nonrenewal in the binary denominator? (Answer: No; record it as a midterm cancellation, distinct from term-end nonrenewal.)
- There are 50 eligible opportunities reaching expiration: 37 renew, 11 do not, and 2 have unresolved outcomes. If the reported rate is among resolved opportunities, what is it, and how many outcomes remain unresolved? (Answer: \(37/(37+11)=0.7708\), or 77.08%; 2 remain unresolved.)