Why Prediction Timing Must Be Clear
A task such as “predict which policies will lapse” sounds straightforward, but it leaves a central question unanswered: when must the prediction be available? A score produced before an outreach decision serves a different purpose from one produced after outreach or near the policy’s expiry. The outcome being predicted must also be placed on a timeline: a policy might lapse by its expiry date, or it might lapse within a stated period after that date.
Prediction timing is part of the task definition, not a detail to add after a model has been built. A task becomes clearer when it specifies both the point at which the prediction is made and the future period or event to which the prediction refers. That framing helps distinguish a genuinely prospective prediction from a summary of something that has already happened.
Write the Task Along a Timeline
A useful first step is to place three items in order: the prediction time, any action or decision that uses the prediction, and the outcome period. The outcome period must make sense in relation to the prediction time. If a task asks about an outcome that is already settled before the prediction is made, it is not asking for a future prediction of that outcome.
For an outcome that is whether an event occurs, let \(t\) denote the prediction time, \(h\) the length of the horizon, and \(T\) the time the event occurs. Define the target as an indicator that equals 1 if the event occurs after \(t\) and by \(t+h\), and 0 otherwise. Under a prediction framed as a probability, the quantity of interest is the probability of that event conditional on the information available at the prediction time.
The event window in the definition explains why the timing matters. The condition \(t<T\) makes the event future to the prediction; the condition \(T\leq t+h\) sets the deadline. If either boundary is omitted, the target can refer to different outcomes in different cases.
This is a definition of one possible time-specific target, not a claim that every prediction task must use an event indicator. For example, a task may predict a future dollar amount rather than whether an event occurs. In either case, the target’s timing needs to be stated. The event-indicator form does not apply unchanged when the target is a continuous amount; that target needs its own definition and horizon.
A Short Justification: Why the Horizon Is Part of the Target
Suppose the same event is of interest at two horizons, \(h_1\) and \(h_2\), where \(h_1<h_2\). The target with the longer horizon includes all events counted by the shorter horizon, along with events occurring after \(t+h_1\) but by \(t+h_2\). Thus the two targets do not ask the same question, even if both are casually described as predicting “whether the event will happen.”
This follows directly from the event bounds: if \(t<T\leq t+h_1\), then \(t<T\) still holds and \(T\leq t+h_1<t+h_2\), so \(T\leq t+h_2\). Therefore, extending the horizon can change which cases count as positive outcomes. A model or score described only as predicting “the event” is underspecified until the relevant time period is known.
Worked Examples
Worked Example: Renewal Outreach Before Expiry
A hypothetical insurer wants to identify policies for a renewal outreach campaign. Policy expiry is June 30. Staff need the prediction on May 31 to plan outreach that begins June 1. A manager proposes the target “whether the policy lapses.” Rewrite the task with a clear prediction time and outcome. Then explain why “lapses within 30 days after expiry” is a different target.
Method: place the decision and outcome on a calendar. The required prediction time is May 31, because the score is needed to plan the June 1 outreach. The relevant expiry date is June 30.
From June 1 through June 30 there are 30 calendar days: June has 30 days, and the full period after May 31 through expiry is June 1–30. A clear task is therefore: “On May 31, predict whether each policy will lapse by its June 30 expiry date.”
The proposed alternative uses a different deadline. “Lapses within 30 days after expiry” refers to July 1 through July 30 if the 30-day period begins the day after June 30. It counts events after expiry that the first target does not count. The two targets can classify a policy differently, so they cannot be treated as interchangeable.
Check: The score is available before outreach, and the stated lapse outcome is future to May 31 and has a definite deadline, June 30. The common mistake avoided is leaving “lapses” undefined as to whether it means by expiry or after expiry.
Worked Example: Claim Severity at a Stated Horizon
A hypothetical claims team wants a severity estimate for each newly reported property claim. The first report arrives on November 4, 2025. The estimate is needed seven days later, on November 11. The team proposes predicting “claim severity” but does not say what amount or time period it means. Define a specific target if the intended amount is cumulative paid loss through 180 days after the first report. Use the convention that the event window ends at the close of the 180th day after November 4.
Method: identify the prediction date, calculate the outcome deadline, and name the amount. The prediction date is November 11, seven days after the first report. The deadline is 180 elapsed days after November 4.
Count the elapsed days to the deadline: November 4 to November 30 is 26 days; December adds 31, for 57; January adds 31, for 88; February 2026 adds 28, for 116; March adds 31, for 147; April adds 30, for 177; and three more days reach May 3, 2026, at 180 days.
A precise task is: “On November 11, predict the cumulative paid amount for the claim through the close of May 3, 2026, which is 180 elapsed days after the first report.” This target is a future amount from the November 11 prediction point. “Ultimate claim severity” would not be an equivalent definition: it does not state a common cutoff and may refer to an amount settled at a different time.
Check: The date arithmetic gives \(26+31+31+28+31+30+3=180\) days. The target specifies both the amount (cumulative paid) and the cutoff. The common mistake avoided is using “severity” as if it necessarily meant the same amount at every prediction time.
Worked Example: Fraud Review With a Future Deadline
A hypothetical insurer assigns a suspected-fraud referral on April 12. It wants a score at assignment to help prioritize cases. A proposed task says, “Predict whether the referral is fraud.” Instead, define a future outcome as whether an investigator substantiates fraud within 45 days after assignment. State the outcome deadline and explain what makes this a prospective task.
Method: anchor the outcome window to the stated prediction time. The score is produced on April 12, so the 45-day period starts from that date. There are 18 elapsed days from April 12 to April 30; the next 27 days run through May 27. Since \(18+27=45\), the deadline is May 27.
A clear task is: “On April 12, predict whether fraud will be substantiated from April 13 through May 27, inclusive.” The outcome is a future investigation result, rather than a general label of whether the referral was fraudulent in some unspecified sense.
The task is prospective because it asks for a prediction at assignment about an outcome that can occur after assignment and by a stated deadline. A different task could ask whether the case is already substantiated on April 12, but that would be a status question at the prediction time, not the future outcome defined here.
Check: The specified 45-day period ends May 27, and the stated outcome window begins after the April 12 prediction. The common mistake avoided is confusing an existing suspicion or status with a future substantiation outcome.
A Practical Framing Sequence
When a prediction request is vague, use the following sequence before discussing a model or score. Each step forces a separate timing choice into the open.
State the date or decision point when a score must be available, such as before a renewal call or when a claim is assigned.
Use an observable, specific result rather than a broad label such as “risk,” “severity,” or “fraud.”
Specify how long after the prediction the outcome is assessed, or the date by which it must occur.
Check that the prediction precedes the outcome being predicted. If the requested outcome is already settled, clarify whether the task is a status description or a future prediction.
This sequence does not determine which prediction is most valuable or which action should follow a score. It makes the prediction question itself explicit, so that different stakeholders are not discussing different deadlines or outcomes under one short label.
Common Timing Traps
- Leaving out the decision point. “Predict lapse” does not say whether the prediction is needed before renewal outreach, at expiry, or later.
- Leaving the horizon open. “Will the claim be severe?” could refer to a stated early period, a fixed payment cutoff, or a later final amount. Those are distinct targets.
- Moving the deadline without changing the task description. A 30-day event target and a 90-day event target do not count the same cases; the longer period includes events after day 30.
- Calling an already-known result a future prediction. If the event occurred before the stated prediction time, the event is not future to that prediction. Clarify whether the goal is to describe current status or predict a later outcome.
- Using one phrase for different timing conventions. Terms such as “by expiry,” “after expiry,” and “within 30 days” need an explicit starting point and endpoint.
Check Your Understanding
For each scenario, identify the missing or specified timing elements and decide how the task should be stated.
- A hypothetical insurer needs a nonpayment prediction on September 1 for policies expiring September 30. State a target if the intended outcome is nonpayment by expiry. (Answer: On September 1, predict whether each policy will have a nonpayment lapse by September 30.)
- A score is required on January 10, and the intended target is whether a reported event occurs within 14 days after that date. State the event window and deadline. (Answer: The event window is January 11–24, inclusive; the prediction is made January 10.)
- A team says, “Predict whether a claim is large,” but gives no prediction date and no assessment date. Name both missing elements needed to turn this into a time-specific task. (Answer: Specify when the prediction is made and define the amount/outcome and its deadline or assessment horizon.)
- A prediction is made on March 5. An event occurs on March 4, but the proposed target is event occurrence within 10 days after prediction. Does the March 4 event meet that target? (Answer: No. It occurred before March 5, so it is not in the future event window after the prediction.)
- A hypothetical case is assigned on June 2, with a target of investigation substantiation within 20 days after assignment. What is the deadline, and what is the prospective outcome window? (Answer: June 22 is the deadline; the prospective window is June 3–22, inclusive.)