Why Data Sources and Observation Windows Matter
A predictive result is only as interpretable as the data behind it. A report may state that a model predicts future claims from policy information, but readers also need to know which records supplied those data, when the records were extracted, what period the predictors describe, and how long claims were observed. Without those details, a reported result can appear relevant while being based on information unavailable at prediction time or outcomes observed over a different period than the intended use.
This tutorial focuses on a concise account of data provenance and observation windows. It does not require a technical inventory of every file or a long description of every record. The goal is to give readers enough dates, sources, and timing conventions to assess whether the evidence fits the predictive question.
Describe the Data Lineage That Affects Interpretation
A useful provenance statement names the source systems or data collections, the relevant records they supplied, and the date the analytic extract was produced. If records from multiple systems were joined or transformed, briefly state the consequential steps. For example, a report might say that policy characteristics came from an administration system, claim outcomes from a claims system, and the two were joined by policy identifier before analysis. This is more informative than saying only that the data came from “internal records.”
Keep the account selective. Readers generally do not need a catalogue of file names, database tables, or routine processing steps. They do need to know about choices that could change the interpretation: a source covering only certain dates, a change in source system, an extract made before all outcomes had time to be reported, or a transformation that changes the measurement period.
In particular, distinguish three dates that are often confused:
- Event date: when the event of interest occurred, such as a service date or policy effective date.
- Availability date: when the event or record became available to the analyst or decision process.
- Extraction date: when the data were retrieved for the analysis.
These dates need not match. A claim can have a service date in one month, be submitted in a later month, and appear in an extract later still. For a model intended to make predictions at a particular time, the availability date can matter as much as the event date: information recorded later cannot be treated as known at the earlier prediction time.
Worked Example: A Service Date Is Not an Availability Date
A hypothetical service model is intended to score members at 00:00 on April 1, 2025. Its predictor window is the 30 calendar days immediately before that time. The source extract contains six candidate service records with service dates from March 2 through March 31. Five were available in the system by March 31 at 23:59; the sixth has a March 31 service date but was not entered until April 2. Determine how many records can be used as predictors under the intended timing rule, and describe the window precisely.
Method. Define the predictor period as the half-open interval from March 2 at 00:00 through April 1 at 00:00: the start is included and the end is excluded. This is 30 days because March 2 to April 1 is 30 elapsed calendar days.
The six service dates fall within the event-date window, but the scoring time is April 1 at 00:00. The five records available by March 31 at 23:59 were known by scoring time and can be considered. The sixth was entered on April 2, after scoring time, so it cannot be treated as information available for that prediction. Thus, five records are eligible under this rule.
Report wording: “Service predictors covered March 2 through March 31, 2025, and were restricted to records available by the April 1 scoring time.” The exact interval convention prevents ambiguity about whether an April 1 record is included.
Check. The result is five, not six: the excluded record has an in-window service date but a post-score availability date. The common mistake is to use event date alone and thereby include information that was not yet available.
Specify Predictor and Outcome Periods Around the Prediction Time
The observation windows should be stated in relation to the index date. A predictor window describes the period of information used to make the prediction. An outcome window describes the future period during which the target is measured. State both the duration and the endpoint convention, especially when a boundary date could otherwise be counted twice or assigned inconsistently.
A simple timing rule helps explain why these definitions matter. Let \(t_0\) be the index time, \(W_X\) the predictor lookback duration, and \(W_Y\) the outcome follow-up duration. If the intended prediction uses only information known at \(t_0\), predictors must be available no later than \(t_0\), and the future outcome period begins at or after \(t_0\), according to the stated convention.
The brackets mean that the left endpoint is included and the right endpoint is excluded. This convention gives adjoining periods a clear boundary: an event exactly at \(t_0\) belongs to the outcome period, not the predictor period. Other conventions can be appropriate, but a report must state them consistently.
The timing rule follows from the meaning of a prediction. At \(t_0\), the model is intended to use what is known then to predict what happens later. If a predictor uses a value first available after \(t_0\), it contains future information relative to the decision. If the target period ends before \(t_0\), it measures past rather than future outcomes. Thus the timing conditions are not just formatting choices; they determine whether the reported analysis represents the stated predictive task.
A report should also clarify the unit to which a window applies when that affects interpretation. For example, state whether each policy has its own effective-date index or whether all policies share a common calendar scoring date. The exact population definition is a separate reporting decision, but the window description should make its time anchor unambiguous.
Worked Example: Set a Renewal Prediction Window
A hypothetical insurer wants to predict whether a policy will have a reported claim during the six months after its renewal score is produced at 00:00 on July 1, 2025. Candidate predictors summarize the preceding 12 months. Use half-open intervals and identify the predictor and outcome dates.
Step 1: Predictor period. The 12-month lookback starts at 00:00 on July 1, 2024. It ends just before 00:00 on July 1, 2025. In half-open form, the period is \([{\rm July\ 1,\ 2024}, {\rm July\ 1,\ 2025})\). It includes July 1, 2024 through June 30, 2025.
Step 2: Outcome period. Six months after the score date ends at 00:00 on January 1, 2026. The outcome interval is \([{\rm July\ 1,\ 2025}, {\rm January\ 1,\ 2026})\), including July 1 through December 31, 2025.
Step 3: State what is measured. If the target is a reported claim, the report must say so; an incurred claim and a reported claim are not interchangeable descriptions. A concise statement is: “Predictors summarized the 12 months before each July 1, 2025 score; the target was whether a claim was reported during the following six calendar months, July 1 through December 31, 2025.”
Check. The predictor period ends exactly when the outcome period begins, so no date is counted in both windows. The common mistake is to call the outcome period “the six months of 2025,” which would describe a different interval.
Follow-Up, Reporting Lag, and Run-Out
An outcome window can have ended even though its records are not yet complete. For example, an event may occur during follow-up but be submitted or processed later. The data cutoff is the latest date through which the analytic extract includes information. Run-out is additional time allowed after the outcome period for relevant events to be reported or recorded. State the run-out convention when it materially affects which outcomes are included.
A statement such as “six months of outcomes were observed” is insufficient if the extract was created before the full reporting lag elapsed. A clearer account distinguishes the event period from the time allowed for later reporting. It should not imply that all future events are known; rather, it should identify the operational rule used to define the target as observed in the analysis.
Worked Example: Check Whether an Outcome Window Has Run-Out
A hypothetical analysis measures claims with event dates during calendar year 2025 and allows 90 calendar days after the outcome period closes for claims to be reported. The extract was produced on March 31, 2026. Assume the outcome period closes at 00:00 on January 1, 2026, and the 90-day run-out is measured forward from that close. Determine whether the cutoff provides the stated run-out.
Step 1: Locate the outcome close. Calendar-year 2025 is represented by \([{\rm January\ 1,\ 2025}, {\rm January\ 1,\ 2026})\). Its close is 00:00 on January 1, 2026.
Step 2: Add 90 calendar days. January 2026 contributes 30 days from January 1 to January 31, leaving \(90-30=60\) days. February 2026 has 28 days, leaving \(60-28=32\) days. Adding 32 days into March reaches April 1, 2026. The 90-day point is therefore April 1, 2026.
Step 3: Compare with the cutoff. The March 31, 2026 extract precedes April 1, 2026 by one day. It does not provide the full 90 days of run-out under the stated convention. The report should not say that all 2025 outcomes had a complete 90-day reporting period.
Check. The comparison is date-for-date: March 31 is one day earlier than the April 1 run-out point. The common mistake is to count “January, February, and March” as three full months and assume that a March 31 extract is automatically 90 days after January 1.
Write a Compact, Auditable Account
A concise report can combine the relevant information in a few sentences. It should let a reader reconstruct the timing without requiring them to infer whether “past year” means a calendar year, a trailing year, or a policy year. A practical sequence is:
Identify which source supplied predictors and which supplied outcomes, using plain names that distinguish them.
Give the date through which records were retrieved, and mention a material joining or transformation if it changes what is measured.
State the index date or scoring time, predictor lookback, outcome follow-up, and endpoint convention where needed.
State the reporting lag or run-out rule and whether the extract reaches the required cutoff.
For example: “Policy characteristics came from the administration system and claim outcomes from the claims system; records were linked by policy identifier. For scores produced at renewal on July 1, 2025, predictors covered the preceding 12 months and the target was a reported claim during July through December 2025. The extract was created March 31, 2026, and included claims reported through that date.” If the target definition requires a specified run-out, the report should compare that requirement with the cutoff rather than leave completeness implied.
The wording should describe what was actually done, not what would have been ideal. If a source does not cover the full intended period, identify the limitation plainly. If different sources have different cutoffs, report each cutoff when the difference affects the evidence. Do not hide a mismatch behind a single broad phrase such as “data through year-end.”
Common Timing Traps
- Using event date as a proxy for availability. A record with an earlier event date can still arrive after the score time. Apply the availability rule relevant to the intended prediction.
- Leaving “past year” undefined. It could mean a calendar year or 12 months before each index date. Name the exact period and anchor.
- Confusing outcome follow-up with run-out. Follow-up is the period in which outcomes occur; run-out is additional time for those outcomes to enter the data.
- Claiming complete outcomes without checking the cutoff. Compare the extract date with the required reporting period, using the stated calendar convention.
- Mixing outcome definitions. “Claim occurred,” “claim was reported,” and “claim was paid” refer to different events and may have different dates.
A reliable account is brief but specific: source, cutoff, index time, predictor period, outcome period, and any material lag or run-out. These details help readers decide whether the evidence corresponds to the prediction being discussed, without overwhelming them with technical data-processing detail.
Check Your Understanding
For each problem, state the timing or reporting conclusion using the conventions given.
- A score is produced at 00:00 on October 1, 2026. Predictors use the preceding 14 days, represented as a half-open interval. What are the interval’s start and end dates? (Answer: \([{\rm September\ 17,\ 2026}, {\rm October\ 1,\ 2026})\), including September 17 through September 30.)
- A model predicts events during the 30 days beginning at 00:00 on May 10, 2026, using half-open intervals. What is the outcome interval and final included date? (Answer: \([{\rm May\ 10,\ 2026}, {\rm June\ 9,\ 2026})\), including May 10 through June 8.)
- A transaction has an event date of February 28, 2026, but was first available to the scoring process on March 3. A score was produced on March 1. Can the transaction be used as information available for that score? Explain the relevant date. (Answer: No; it became available on March 3, after the March 1 score time.)
- A target period closes at 00:00 on October 1, 2025, and the required run-out is 30 calendar days. What is the run-out date, and would a September 30, 2025 extract satisfy it? (Answer: October 31, 2025; no, the extract predates the target-period close.)
- A hypothetical report uses policy-system predictors for scores on January 1, 2026, and claim-system outcomes for the following calendar quarter. The extract was created June 30, 2026. Name two timing details the report should state beyond the source names. (Answer: Define the predictor lookback and the quarter’s exact outcome dates, and state the outcome reporting or run-out rule and how the June 30 cutoff relates to it.)