Statistics Glossary → Forecast horizon and forecast interval
Forecast horizon and forecast intervalconcept
How many periods beyond the last observation a time series model forecasts (the horizon), and the interval around each forecast within which the future value is expected to fall with a given probability.
The forecast horizon h is the number of periods to be forecast after the last observation: with monthly data, h = 12 means monthly forecasts for the next year. Each forecast comes with a forecast (prediction) interval: a 95% interval is the range within which the future observation should fall with about 95% probability if the model's assumptions hold. In models such as ARIMA and VAR, the variance of the forecast error grows with the horizon because it reflects the accumulation of future shocks, so the intervals widen further ahead.
A forecast interval is wider than a confidence interval for the mean, because it also includes the random variability of a single future value. Standard intervals often account only for future shocks, not for uncertainty in the parameter estimates or in model selection, so in reality they may be somewhat too narrow. For multiplicative models or non-normal errors, intervals can be built by simulation. Long-horizon forecasts rest on the assumption that the structure the model captured in the past will continue; a structural break invalidates it.
An everyday example
An energy company forecasts monthly electricity consumption. Next month's forecast is 500 GWh, with a 95% forecast interval of 480-520 GWh. The same model may again forecast about 500 GWh twelve months ahead, but the interval widens to 430-570 GWh. This says the company can plan confidently for next month but needs more flexible capacity a year out.
In MerQur
It appears as 'Forecast horizon (1-100) and Confidence Level' in ARIMA Forecast, 'Forecast horizon and Confidence Level' in Exponential Smoothing and VAR, and 'Forecast horizon (default 12; 0 = no forecast)' in the Unobserved Components Model. In VAR, the 'IRF / FEVD horizon' also sets the number of steps for impulse responses and variance decomposition.
Analyses where this concept appears
ARIMA ForecastExponential Smoothing (Holt-Winters)VAR (Vector Autoregression)GARCH (Volatility Model)Unobserved Components Model (Structural Time Series)