One of the questions my students sometimes ask is how to set the forecast horizon. The answer depends largely on the task at hand, but there are still some guidelines.

First, the forecast horizon depends on data granularity. A “year ahead” forecast on monthly data means forecasting 12 steps ahead, while for daily data, it means 365 (or 366) steps. When deciding on the length of the horizon, think about what you’ll do with the forecast. Are there decisions to be made daily for the next year? If not, maybe it makes sense to aggregate the data to monthly, quarterly, or yearly frequency.

Second the forecast horizon should align with the decisions you make. There’s no point in forecasting a year ahead if you’re making marketing decisions based on three month ahead. You will just spend resources on something not important and potentially distracting.

On the other hand, if your horizon is too short, you risk missing important information. If your lead time is three days, a one-step-ahead forecast won’t give you the full picture. You need to forecast for the entire lead time to be fully prepared.

All of this might seem obvious, but for those unfamiliar with forecasting, it can be a revelation.

We once had a client who wanted half-hour forecasts of demand for a year ahead, i.e they wanted a forecast horizon length of 17,520. There was no business decision aligned with that level of detail; it was simply a misunderstanding of the forecasting process.

Third, some students and even experts in time series analysis focus only on one-step-ahead forecasts. In my experience, a horizon of one is rarely useful. Yes, if your lead time is only one day or one week, then a one-step horizon makes sense (for daily/weekly data respectively). But in many other cases, the decisions will be made based on longer horizon. Furthermore, as the horizon increases, so does the uncertainty due to unpredictable factors building upon each other over time. If your research focuses only on h=1, it might not generalize to longer horizons. Make sure your method performs well for h>1 as well.

Finally, if you do forecasting just for fun or a pure theoretical research, there are no strict rules – you can do whatever makes sense. Still, I’d recommend setting the horizon at least as long as the time series’ periodicity. For example, if you work with daily data and have day-of-week seasonality, set the horizon to at least seven steps ahead. This helps you see how the model performs over a full seasonal cycle.