Every forecast answers a specific question about a specific time ahead, which is its horizon.
Choosing the horizon comes before choosing the method, because methods suit different ranges.
Short horizons rely on recent patterns, while long horizons rely more on trend and judgement.
A 1 step ahead forecast predicts only the next period, using all data up to now.
A multi-step forecast predicts several periods at once, such as each of the next 30 days.
Recursive forecasting feeds each forecast back in to make the next, which lets errors compound.
Direct forecasting builds a separate model for each step ahead, which avoids compounding at more cost.
For a random walk, the width of the prediction interval grows with the square root of the horizon.
Seasonal patterns stay forecastable far ahead, while short-term noise does not.
Planning decisions set the horizon: staffing needs days, stock orders need weeks, and new sites need years.
Lead time matters, since a forecast is useless if the decision had to be made before it arrived.
Different horizons may need different levels of detail, such as daily for next week and monthly for next year.
Judging a method should use the horizon it will serve, not just 1 step ahead.
In practice forecasts are updated as new data arrives, so long range forecasts are revised many times.
The usual mistake is judging a long range forecast by its 1 step ahead accuracy, which is always much better.