Why a Model May Need to Stop
Monitoring a model is useful only if there is a clear response when the evidence shows that its predictions cannot safely support the intended decision. A warning is not the same as a decision to continue using the model: unreliable inputs, persistent performance failures, or a decision context that no longer exists can require an immediate limit on use. This tutorial is aligned with the SOA Exam PA (Predictive Analytics) syllabus.
The action should match the reason for concern. A temporary data-feed failure may call for a suspension while the feed is repaired. A problem limited to one product or operating region may justify restricting the model to the remaining supported cases. If the model’s purpose has permanently disappeared, continued monitoring or retraining may not make sense; retirement may be appropriate.
These actions are governance decisions, not merely statistical labels. A useful rule states what evidence triggers action, who has authority to act, which uses are affected, and what process replaces the model. The trigger should be set before a problem occurs where feasible. Otherwise, there is a risk that a concerning result will be dismissed because the organization has not agreed on what it means.
Conditions That Can Require Restriction, Suspension, or Retirement
A model should not be used as though its predictions were reliable when important inputs are missing, invalid, or measured in a form the model was not designed to handle. Examples include a failed data feed, a change from dollars to thousands of dollars, unexpected categories, or values outside a documented operating range. A small number of flagged records may be handled individually if the model is restricted to valid records and the excluded records receive a safe alternative. A widespread or undetected input failure may require suspension.
Performance evidence can also justify a limit. An isolated alert may be a reason to investigate, but a prespecified rule may require action when an important metric exceeds its limit repeatedly, when a minimum evidence volume is met, or when results are too uncertain to support the intended decision. The threshold and the metric must be interpreted in context: an average error that is acceptable for one decision may be unacceptable for another.
Finally, a model can become unsuitable even if its measured performance has not deteriorated. The population, product, decision, or operational process may have changed so that the model no longer addresses the question for which it was built. If the target outcome or action is no longer relevant, a good historical score does not establish that the model remains useful.
| Evidence | Possible response | Key question |
|---|---|---|
| A limited set of records has invalid inputs | Restrict use to records with validated inputs; route excluded records to an alternative | Can affected records be reliably identified before a prediction is used? |
| A critical input feed is broadly unreliable | Suspend model use until input reliability is established | Can the model’s required inputs be verified for current decisions? |
| A prespecified performance rule is breached | Restrict or suspend, depending on the rule and scope of the breach | Is the concern sustained and supported by adequate evidence? |
| The original decision context has ended | Retire the model for that use | Does the prediction still support a real, authorized decision? |
A Conservative Check When Inputs Are Unreliable
A simple bound explains why invalid inputs cannot be treated as a minor detail just because the model performs well on the records that pass validation. Suppose a metric is an average loss, and the records are divided into a valid fraction \(1-q\) and an invalid fraction \(q\). Let \(L_V\) be the average loss on valid records and \(L_I\) the average loss on invalid records. The overall average loss is the weighted average of the two groups:
This follows by adding the total loss in each group and dividing by the total number of records. If the loss is known to be between \(0\) and a maximum \(M\), then \(L_I\leq M\), so:
This is an upper bound, not a prediction of the actual loss. It is useful only when the stated maximum is genuine and the groups are counted consistently. If loss on invalid records is unbounded or unknown, this calculation cannot certify acceptable overall performance. In that case, the safer response may be to exclude those records from model use rather than assume their errors resemble the valid records.
Worked Example: Invalid Inputs and a Performance Limit
A hypothetical property screening model is evaluated on 120 records. A review finds 18 records with an invalid replacement-cost input; those records are separate from the valid records. On the 102 valid records, the average capped absolute prediction error is $42. The operational metric caps each record’s absolute error at $500, and the maximum acceptable average error is $80. A governance rule also requires suspension if more than 5% of records have invalid inputs. Decide whether continued unrestricted use is supported.
Method. Calculate the invalid-input proportion, then use the bounded-loss result to assess whether the overall error can be certified below the limit. The cap of $500 is an explicit assumption for this metric; without it, the bound would not apply.
The invalid-input proportion is \(18/120=0.15\), or 15%. The valid proportion is \(102/120=0.85\). The input rule is breached because \(15\%>5\%\).
For the conservative performance bound, substitute \(q=0.15\), \(L_V=\$42\), and \(M=\$500\):
The bound of $110.70 exceeds the $80 limit, so these data do not establish acceptable overall performance. The calculation does not prove that actual overall error is $110.70; it shows that the available evidence cannot rule out an error above the limit under the stated cap.
Decision. Unrestricted use is not supported: the invalid-input rate breaches the prespecified rule, and the conservative bound does not certify performance below the limit. Suspend predictions that depend on the affected input until it is validated. If individual invalid records can be reliably detected, the model might be restricted to records with validated inputs, with the others sent to a documented alternative process.
Distinguishing a Warning From a Suspension Rule
An alert is evidence to review; a suspension condition is a rule that requires action. A well-specified performance condition identifies the metric, the limit, the time window, any minimum sample size, the number of breaches needed, and the scope of the response. For example, a rule could require suspension after three consecutive monthly windows with mean absolute error above a stated limit, provided each window contains at least a stated number of verified outcomes. The exact rule depends on the decision and the monitoring plan; it should not be selected after seeing a disappointing result simply to avoid action.
Averaging across time can conceal a sustained breach. If the condition is defined window by window, each window must be checked against the condition. A better overall average does not erase a required trigger. Conversely, a single noisy window should not be called a sustained failure unless the rule says that one breach is sufficient.
Worked Example: Applying a Sustained-Error Rule
A hypothetical claims triage model has a prespecified rule: suspend if mean absolute error exceeds $100 in three consecutive monthly windows, with at least four verified claims in each window. The absolute errors in the three most recent windows are, respectively, $90, $110, $100, $120; $105, $115, $100, $130; and $95, $110, $115, $120. Determine whether the rule is met.
Method. Compute each monthly mean absolute error separately, check the sample-size condition, and then count consecutive breaches. Equality to $100 does not exceed a limit of $100.
Each window has four verified claims, so all three satisfy the minimum volume requirement.
For month 1, the total absolute error is \(90+110+100+120=420\), so the mean is \(420/4=\$105\). This exceeds $100.
For month 2, the total is \(105+115+100+130=450\), so the mean is \(450/4=\$112.50\). This exceeds $100.
For month 3, the total is \(95+110+115+120=440\), so the mean is \(440/4=\$110\). This exceeds $100. There are three consecutive qualifying breaches, so the suspension rule is met.
As a check, the pooled mean across all 12 claims is \((420+450+440)/12=1310/12=\$109.17\), rounded to the nearest cent. It also exceeds $100, but the pooled calculation is not a substitute for checking the three required windows.
Decision. Suspend the model under the stated rule and use the approved alternative process while the cause and scope of the deterioration are assessed. Do not treat the fact that the second window contains one error of exactly $100 as a reason to disregard that month’s mean: the rule applies to the monthly mean, not to each individual error.
When the Decision Context Has Changed
A model’s purpose should be stated in operational terms: what it predicts, for whom or what, and what decision the prediction supports. If that decision is removed or materially changed, a historical measure of predictive performance may no longer answer the relevant question. In particular, a model should not be kept in production just because it still generates a score or because its original evaluation was strong.
The distinction between suspension and retirement depends partly on whether the use might reasonably resume. Suspension is appropriate when the concern is temporary or still being investigated. Retirement is appropriate when the model’s intended decision has permanently ended, or when no authorized use remains. Retirement should include removing the model from the decision process, recording its final scope and rationale, and preserving required documentation. Retiring a model does not mean discarding records that must be retained for audit or governance purposes.
Worked Example: An Obsolete Decision Context
A hypothetical insurer used a model to rank policies for a manual review process associated with a particular legacy product. The product has been permanently closed, and a new operating process routes all current policies through the same review pathway without using a ranking. During a final system check, 640 records are found in the current scoring feed. Determine whether restriction or retirement is the more appropriate action.
Method. Compare the model’s documented purpose with the current decision, then count how many current records belong to the original use. A model is relevant only if its prediction still supports an actual decision within its intended scope.
The documented use was ranking policies for the legacy-product review process. That product is permanently closed, so the number of current records in the original decision context is zero. The current process does not use a ranking; therefore, none of the 640 records has a decision for which this model’s score is intended.
The current in-scope proportion is \(0/640=0\%\). Restricting the model to the current feed would not solve the problem because there are no in-scope records to retain. Nor would a strong historical performance measure make the old ranking relevant to the new process.
Decision. Retire the model for this use, remove its score from the operational workflow, and document the reason. Continue with the current approved review pathway. If a future process introduces a genuinely new ranking decision, that use would need its own appropriate evaluation and authorization; it is not automatically covered by the retired model’s former purpose.
Making the Response Safe and Auditable
A model stop decision should leave no ambiguity about what happens next. The alternative must be explicit, available to the people handling affected cases, and appropriate to the scope of the problem. A manual review is not automatically safe merely because it is human; reviewers need usable information, a consistent procedure, and a way to record the decision. If a model is restricted, the system should prevent excluded records from receiving its output as if it were valid.
A practical response can be organized as follows:
Record the evidence, metric or input condition, threshold, time period, and affected records or decisions.
Decide whether the evidence supports record-level restriction, broader suspension, or retirement because the intended use has ended.
Specify how affected decisions will proceed, who is responsible, and how the process will handle missing or uncertain information.
Record the decision-maker, rationale, start date, scope, required follow-up, and the conditions for any return to use or permanent retirement.
For a temporary suspension, the conditions for resuming use should be stated rather than assumed. They might require verified inputs, adequate evidence over a defined period, and approval by the responsible governance authority. If those conditions cannot be met, the organization should consider whether the model remains suitable for the decision at all.
Common Exam Traps
- Treating every alert as proof of failure. An alert prompts investigation; apply the actual prespecified action rule and its evidence requirements.
- Using a strong average to excuse invalid records. Performance on valid inputs does not establish performance on records whose inputs cannot be trusted.
- Ignoring the rule’s time structure. If action depends on consecutive windows, calculate and inspect each window rather than relying only on a pooled average.
- Confusing restriction with retirement. Restriction preserves an identified supported use. Retirement is appropriate when the intended use has ended or no longer makes sense.
- Stopping without a safe alternative. State how decisions will proceed while the model is restricted or suspended; do not leave an operational gap.
- Assuming a model is still relevant because it can still score records. A score is useful only if it supports a current, intended decision.
Check Your Understanding
For each situation, identify the appropriate action and apply the stated condition carefully.
- A model accepts 200 records, of which 12 have an input format that cannot be verified. The rule requires restriction if more than 5% of records are invalid. Calculate the invalid percentage and state the required action. (Answer: 6%; restrict use to validated records and route the 12 invalid records to the approved alternative.)
- A prespecified rule suspends a model when its monthly error exceeds $75 in two consecutive windows, with at least three observations per window. The two windows have errors of $60, $90, $90 and $70, $80, $90. Does the rule trigger? (Answer: Yes; the window means are $80 and $80, both exceed $75, and each window has three observations.)
- A bounded loss is between 0 and $300. The invalid-input fraction is 0.20 and the valid-record mean loss is $50. Calculate the conservative upper bound on overall mean loss. (Answer: \(0.80(50)+0.20(300)=\$100\).)
- A model is restricted to one region after a data-feed failure in another region. The restriction is lifted automatically after one successful test record, although the documented resumption condition requires 100 verified records. Identify the error and state what should happen. (Answer: One record does not meet the 100-record condition; keep the restriction in place until the documented condition is satisfied and approval is given.)
- A model’s documented decision was permanently removed, and the current workflow contains 900 records but no decisions of the original type. Should the model be restricted or retired for that use? (Answer: Retire it for that use; none of the 900 current records belongs to an active in-scope decision.)