Is a twelve-step horizon always one year?

No. It is one year for monthly data, but twelve weeks for weekly data.

Can one model produce several horizons?

Yes. The resulting forecasts can be evaluated individually and as a path.

Must uncertainty always grow with horizon?

No universal rule applies. Some models approach a limiting uncertainty, and seasonal or known-event structure can create uneven patterns.

AI Glossary / Time series and signal processing / Intermediate

Forecast horizon

Definition

A forecast horizon is the distance from the forecast's information cutoff or origin to the future period being predicted. It may be expressed as a number of observation steps, an elapsed duration, or a sequence of future periods. A horizon must be interpreted together with the data frequency and the time at which the forecast is issued.

What it means

A next-day forecast and a next-quarter forecast ask different questions, even if they concern the same business. The horizon determines how far ahead the model must look and which uncertainties may accumulate before the outcome occurs.
A forecast issued today for four weeks ahead has a four-week horizon. If a new forecast for the same target week is issued next week, its horizon is three weeks. The target date stays fixed, but the available information and remaining lead time change.

How it works

Define the forecast origin T and the target T+h. The notation h denotes the number of steps ahead, so h = 4 means four days for daily data or four months for monthly data. If a report includes horizons one through twelve, it contains a forecast path rather than just one future point.
Choose horizons from the decisions they support, then evaluate errors separately at those horizons. Multi-step forecasts may be generated recursively, directly, or jointly. Uncertainty often changes with horizon, but the pattern depends on the process and model; it need not increase without limit or in a perfectly smooth way.

Three practical examples

A shop needs a forecast covering supplier lead time to decide replenishment quantities. Tomorrow's accuracy alone does not establish whether the model supports that decision.
A staffing team issues schedules several weeks before shifts occur. It evaluates forecasts made with the information available at that scheduling deadline.
A household estimates next month's energy use for budgeting while using a shorter horizon to adjust tomorrow's consumption. The same underlying data can support distinct forecast tasks.

Think of it this way

Think of planning a journey. The road immediately ahead is visible in detail, while conditions farther away depend on more things that have not yet happened. The distance matters, but so does the type of road: a predictable highway and an unfamiliar junction can create different uncertainties even at similar distances.

Worked example

Consider an original hypothetical weekly process with current level 120. Assume each future week's change equals a known drift of 3 plus an independent normally distributed innovation with mean zero and variance 16. Treat the current level and all model parameters as known for this example.
At horizon h, the forecast is 120 + 3h. The forecast-error variance is 16h because h independent innovations accumulate. Thus the one-week forecast is 123 with standard deviation 4; the four-week forecast is 132 with standard deviation 8; and the nine-week forecast is 147 with standard deviation 12.
An approximate 95% prediction interval at four weeks is 132 ± 1.96 × 8, or [116.32, 147.68]. The four-week standard deviation is twice, not four times, the one-week value. The variance has increased fourfold, while standard deviation takes its square root.
These calculations concern the value at the target week. They do not describe the total over the next four weeks, whose uncertainty would require a different calculation involving dependence among those future levels. Parameter estimation, changing drift, or dependent innovations would also alter the stated uncertainty.

Illustrative case study

Illustrative fictional case: Juniper Supplies selects a forecasting model because it predicts next week's orders well. Purchasing staff then discover that suppliers require six weeks of notice. A comparison at the six-week horizon changes the ranking of the candidate models.
The team revises its evaluation to match purchasing decisions, records the exact issue date of each forecast, and keeps short-horizon forecasts for a separate staffing use. This fictional case illustrates alignment between evaluation and action; it does not imply that any one forecasting method always wins at longer horizons.

Common misconceptions and mistakes

Forecast horizon is not the amount of historical training data. A model can use five years of observations to predict one day ahead, or only a short history to attempt a long forecast, though the latter may be poorly supported.
Do not compare models at different horizons and attribute the entire performance difference to model quality. Also distinguish a horizon from a prediction interval: one concerns time ahead, the other uncertainty about an outcome. Terms such as short-term and long-term are context-dependent; fixed month cutoffs from a glossary are not universal across business applications.

Business use

A clear horizon specification helps teams choose models, data refresh schedules, and performance measures. It also reveals whether a forecast arrives early enough to affect the relevant decision. A highly accurate forecast issued after an ordering deadline may have little operational value.
Document the target frequency, issue timestamp, information cutoff, required lead time, and whether the target is a single period or an aggregate. Report accuracy and interval coverage by horizon, and consider the cost of errors at the horizons where commitments become difficult to reverse.

Check your understanding

Answer: A forecast uses data through March and predicts June monthly demand.

What is the horizon?

Answer: Three monthly steps, assuming April is the first forecast month.

In the worked example, why is four-week variance 64?

Answer: Four independent innovation variances of 16 are added.

Would the four-week interval also cover total demand over four weeks?

Answer: Not by the calculation shown; the target quantity is different.

Related terms

Prediction interval
forecast origin
lead time
multi-step forecast
rolling-origin evaluation
forecast path
training window
temporal aggregation
forecast error

Learn next

evaluate a simple benchmark at several horizons using the same historical issue dates. Then compare forecasts for an individual future period with forecasts of a cumulative total, keeping the decision target explicit.

Sources

Sources: Glossary of Forecasting Terms, Forecast horizon: https://robjhyndman.com/mwh3/FG4.pdf . Hyndman and Athanasopoulos, Forecasting: Principles and Practice, time-series cross-validation and prediction intervals: https://otexts.com/fpp3/tscv.html and https://otexts.com/fpp3/prediction-intervals.html . Primary glossary and indexed textbook sections consulted. The known-parameter drift example and fictional case are original.

Previous: Forecast bias Next: Forecast skill