Introduction
A forecast question is not fully specified by asking, “What will happen next?” A usable forecast requires a date at which the prediction is made, a precise outcome to predict, a statement of what information is available then, and a horizon measured in the data’s time units. These details are especially important in actuarial work: an observation may refer to one date but not be reported or finalized until later.
The same distinctions govern historical training data. A model may be fitted later using outcomes that became known after earlier forecast dates, but the predictors for each historical row must represent what would have been available at that row’s origin. Otherwise, the model is trained using information that could not have been used when the forecast was supposedly made.
Translate the Prompt into a Time-Indexed Question
For observations indexed by regular periods, let \(Y_t\) denote the value of the target variable in period \(t\). A forecast made at origin \(T\) for a target \(h\) periods ahead can be written as a forecast of \(Y_{T+h}\) using \(\mathcal{F}_T\). The notation does not say which forecasting method is used; it makes the timing of the question explicit.
A target need not be one observation. For example, a question may ask for the total number of claims over the next three months. If the origin is the end of month \(T\), that target is \(Y_{T+1}+Y_{T+2}+Y_{T+3}\), and its horizon is three monthly periods. A forecast of the next quarter’s total is not the same target as a forecast of each month separately, even though both refer to the same calendar quarter.
The origin is also not necessarily the date of the most recent observation in a data file. Reports can arrive late, and values can be revised. State both the forecast cutoff and the latest usable observation date when they differ. For every input, ask when it became available, not only which period it describes.
Worked Example: A Claims Forecast with a Reporting Delay
A hypothetical insurer records monthly claim counts. Each month’s count is finalized at the end of the following month. On August 31, 2025, an analyst must forecast the total count for September through November 2025. Identify the origin, target, horizon, latest finalized observation, and which historical training labels can be used if fitting occurs on August 31.
Method. Put the forecast date, target months, observation dates, and report availability on one monthly timeline. Do not treat the event month and the report-availability month as interchangeable.
Origin and information. The forecast origin is August 31, 2025. Under the stated reporting rule, July’s count is finalized at the end of August. August’s count will not be finalized until September 30. Thus, even though August has ended, the latest finalized monthly count at this origin is July’s count.
Target and horizon. The target is the sum of the three monthly counts for September, October, and November 2025. These are the first, second, and third months after the August 31 origin, so the horizon extends through \(h=3\). In symbols, the target is \(Y_{\text{Sep}}+Y_{\text{Oct}}+Y_{\text{Nov}}\).
Historical training labels. Consider a historical row whose origin was April 30, 2025, and whose target was the three-month total for May through July. The July count is finalized at the end of August, so that complete target is available on August 31. This row’s label is mature and can be used for fitting at the current cutoff. A row with origin May 31 targets June through August; its August count will not be finalized until September 30. That label is not yet available for fitting on August 31.
Check. The appropriate current target begins in September, not August, and the latest finalized observation is July, not August. The historical row ending in July is eligible; the row ending in August is not. The tempting mistake is to use every period that has ended as though its reported value were already final.
What Counts as Available Information?
Information availability is determined by the forecast process, not by whether a value can eventually be found in a historical database. A predictor is usable at origin \(T\) if its value was known, or could genuinely have been calculated, by that date. A future-dated value can qualify when it was fixed in advance: a scheduled premium change, a published calendar, or a contracted payment may be known at the origin. A future realization, such as actual claims, realized weather, or a later recorded cancellation, is not known merely because it appears in a completed data set.
There are two separate timing checks. First, each input in a forecast row must be available at that row’s origin. Second, when fitting a model at a later cutoff, the row’s target must have become observable by the fitting date. Historical targets naturally occur after their origins; that is what makes them outcomes to learn from. They must not be placed among the predictors for those earlier origins.
Write the forecast date and the exact data cutoff. Include any reporting or processing delay that affects what is actually on hand.
Specify the outcome, its measurement, and the target period. Decide whether the request is for one period, several separate values, or an aggregate.
Use the frequency of the series and count periods after the origin. Do not substitute the number of rows in a data table for elapsed forecast periods.
For predictors, check availability at each historical origin. For training labels, check that the complete target was available by the model-fitting cutoff.
Why the Information Cutoff Matters
A forecast made at \(T\) must be a function of information available at \(T\). In notation, its inputs must be measurable using \(\mathcal{F}_T\): they can depend on the past and on information already known at \(T\), but not on an unobserved future realization. This gives a direct justification for the no-future-information rule.
Suppose a proposed predictor is the realized claim count \(Y_{T+1}\), but the forecast is made at \(T\). By definition, \(Y_{T+1}\) is a future outcome and is not part of \(\mathcal{F}_T\). A prediction rule that uses it cannot have been applied at origin \(T\). In contrast, a value \(C_{T+1}\) for next month’s contracted coverage may be usable if the contract fixed it by \(T\); its future time label does not make it unknown. The relevant test is availability at the origin, not whether a variable is labeled “future.”
Historical training data use the same logic with two different dates. For a row with historical origin \(t\), its predictor vector \(X_t\) must be constructed from \(\mathcal{F}_t\). Its later outcome \(Y_{t+h}\) supplies the label only once that outcome has been reported. If model fitting occurs at cutoff \(C\), a row is eligible only if its full target is available by \(C\). For a target covering several periods, the latest target period and any reporting delay both matter.
Worked Example: A Quarterly Policy-Lapse Forecast
A hypothetical insurer wants to forecast the number of policy lapses in the fourth quarter of 2026. The forecast is made on June 30, 2026, using completed quarterly records. At that date, the first-quarter and second-quarter lapse counts are available. A premium change for each policy in the fourth quarter was already issued and is recorded. Third-quarter cancellations will only be known after the quarter ends. Determine the horizon and classify the proposed inputs.
Method. Index the completed quarters relative to the origin, then classify each variable by when it is known. The requested count is a single quarterly target, not a cumulative count from the origin.
Horizon. The origin is the end of 2026 Q2. The intervening next quarter, Q3, is one period ahead; Q4 is two periods ahead. Therefore the target is the Q4 lapse count at \(h=2\). The fact that Q3 is not the target does not turn Q4 into a one-quarter-ahead forecast.
Inputs. The Q1 and Q2 counts are available at the origin and can be considered as historical information. The issued Q4 premium change is also available because it was fixed before June 30. The actual Q3 cancellation count is not available at the origin. It cannot be used as a predictor for this forecast, even if it is later included in the insurer’s completed 2026 data.
Historical-row check. For a training row whose origin was June 30, 2025, the Q4 2025 lapse count can serve as its label only after Q4 ends and the count is reported. If the model is fitted on June 30, 2026, that Q4 2025 label is available under the stated facts. But the predictors in that historical row must still be values known on June 30, 2025; they cannot include that row’s later Q3 cancellation count.
Check. The horizon is two quarters, the pre-issued Q4 premium change is an available input, and realized Q3 cancellations are not. A common error is to use a future observed count because it is present in the final data set. The correction is to reconstruct the data as it stood at each forecast origin.
Aggregated Targets and Units of Time
A prompt may give the target as an annual total, a quarterly average, a rate over an exposure period, or a vector of future period values. State the target’s unit as well as its dates. If quarterly payments are recorded in dollars, for instance, a four-quarter total is measured in dollars; it is not a forecast for one quarterly payment. If the source data are monthly, a twelve-month horizon is twelve periods even when the result is summarized as one annual amount.
Also keep the origin’s date convention consistent. “At the end of June” usually means June’s completed observations may be available only if the reporting process makes them available by that instant. If reports arrive later, the last usable month may be May. Never silently assume that period-end values are immediately known.
Worked Example: Forecasting Annual Benefit Payments from Quarterly Data
A hypothetical benefit program has quarterly payment totals, in millions of dollars. Its valuation date is December 31, 2025, and the requested target is the total payments over calendar year 2026. Quarterly records become available ten days after each quarter ends. The observed 2024 quarterly payments were \(0.40\), \(0.42\), \(0.45\), and \(0.48\) million dollars. Determine the target, horizon, information cutoff, and whether the 2025 annual total can be used as a mature historical training label on the valuation date.
Method. Convert the annual request into its quarterly components, add the supplied historical values, and apply the ten-day publication delay to the fitting-data check.
Target and horizon. The target is the sum of the four 2026 quarterly payment totals. From December 31, 2025, the target quarters are Q1, Q2, Q3, and Q4 of 2026, so the horizon reaches four quarterly periods. The request is for one annual aggregate, even though four future quarterly outcomes contribute to it.
Information at the origin. The origin is December 31, 2025. The 2025 Q4 quarter has just ended, but its record will not be available until January 10, 2026. Therefore, as of the valuation date, the latest available quarterly record is for 2025 Q3. The eventual Q4 value cannot be used as an input on December 31.
Historical aggregate calculation. The 2024 total is
This is a complete annual label for a historical forecasting row whose origin was December 31, 2023. All four 2024 quarterly records were available by January 10, 2025, so the 2024 label is mature by December 31, 2025.
2025 label check. The 2025 annual total includes Q4 2025. That record is not available until January 10, 2026, which is after the fitting cutoff of December 31, 2025. Thus, the complete 2025 annual label cannot be used for fitting on the valuation date, even though Q1 through Q3 are available. Using those three quarters as if they were the full-year total would change the target definition and understate the label.
Check. The target is a four-quarter total at a four-quarter horizon; the 2024 total is \(1.75\) million dollars; and the 2025 annual label is not yet mature at the stated cutoff. The trap is to confuse a quarter that has ended with a quarterly record that has been reported.
Common Mistakes in Forecast-Question Translation
| Mistake | Why it fails | Correction |
|---|---|---|
| Calling the latest row the forecast origin | The row may describe an earlier period or may not yet be reported. | State the forecast date and data-availability cutoff separately. |
| Counting the target period as horizon zero | This shifts the requested prediction by one period. | Count the periods after the origin: the next month is \(h=1\), the following month is \(h=2\). |
| Using a future realization as a predictor | That value was not in the origin’s information set. | Reconstruct each historical predictor using only what was known at its own origin. |
| Training on an incomplete target window | The row’s label no longer represents the requested target. | Wait until every component of the target and any required reporting lag have elapsed. |
| Ignoring the measurement unit | A total, a rate, and a period-specific count answer different questions. | Name the variable, period, aggregation, and unit explicitly. |
Check Your Understanding
For each problem, identify the time alignment before deciding what may be used.
- A hypothetical monthly exposure series is reported five days after month-end. On October 31, an analyst forecasts November exposure. What is the forecast origin and horizon, and is the October exposure report available on that date? (Answer: Origin October 31; \(h=1\) month; the October report is not available until November 5.)
- A forecast is made at the end of 2027 Q1 for the total number of new claims in 2027 Q2 and Q3. State the target and horizon. (Answer: The target is the sum of Q2 and Q3 new claims; it covers \(h=1\) and \(h=2\), so the target window ends at a two-quarter horizon.)
- A quarterly lapse forecast is made at the end of Q2. Which is available as an input: a premium adjustment finalized in Q1 for Q4, or the realized Q3 lapse count? (Answer: The Q4 premium adjustment is available if finalized in Q1; the realized Q3 lapse count is not available at the end of Q2.)
- Monthly claim totals are finalized two months after month-end. A model is fitted on June 30 and uses historical origins at the end of January, each with a one-month target. Is the January-origin target for February available by the fitting cutoff? (Answer: Yes; February’s total is finalized at the end of April, before the June 30 fitting cutoff.)
- A forecast made on December 31, 2030 targets the total paid amount for all four quarters of 2031. Quarterly reports arrive 15 days after each quarter ends. What is the horizon, and can the full 2030 total be used as a training label on December 31, 2030? (Answer: The horizon is four quarters; no, the 2030 Q4 report is not available until January 15, 2031, so the full-year label is not yet mature.)