Use the History Offset for Out-of-Sample Forecast Tuning

This topic explains how you can use the history offset for a forecasting profile for performing out-of-sample forecast tuning or back testing.

You set the history offset in the Forecast Profiles section on the Demand tab of the Plan Options page for a demand plan, demand and supply plan (in the Demand and Supply Planning work area), or replenishment plan. When you set the history offset, the forecast start date is one bucket after the history end date.

Benefits of Using the History Offset

The history offset enables you to shift the history end date backward so that you can generate a forecast for the period for which you have actual data. By comparing the forecast with your actual data in the period between the history end date and the current date, you can accurately fine-tune the forecasting methods, causal factors, and forecasting parameters used in your forecasting profile, thereby increasing the reliability of the statistical forecast.

By using this feature, you don't have to wait for several months to collect data and compare it with your forecast so that you can fine-tune your forecasting profile.

Points to Note When Using the History Offset

Some points to note when using the history offset are as follows:

Preferably, create a forecasting profile for using the history offset. Ensure that the output measure isn't used by other forecasting profiles.

This forecasting profile should contain the forecasting methods, causal factors, and forecasting parameters that are used by the other forecasting profiles that you use for the statistical forecast.
The history offset uses the time level you specify in the Forecasting Time Level field.
When you set or change the history offset, the history start date, history end date, and forecast end date are automatically updated.
When you use the history offset, you should ensure that the tables and graphs in your plan display historical data for enough days to cover the new history start and end dates.

You can control the number of days for displaying historical demand before the plan start date in tables and graphs by using the Demand History Days field in the Demand: Advanced Options dialog box that you open from the Demand tab on the Plan Options page. Enter the number of days to display historical demand before the plan start date in tables and graphs when a time filter isn't specified. The default is 182.
If your data is seasonal, you preferably should have at least two years of history before the new history end date so that trends are captured and an accurate statistical forecast is generated. You can't generate an accurate statistical forecast if you have less than one year of history before the new history end date.
The history offset isn't considered in the calculation of the measure that stores the history average for the forecasting profile. Instead, the original history end date is used in the calculation of the history average.
If the plan is used as a demand schedule for a replenishment plan or a supply plan, the history offset isn't considered.

Refine Your Statistical Forecast by Using the History Offset

These are the high-level steps for using the history offset for refining your statistical forecast:

Run your plan with the forecasting profile that you normally use.

On the Parameters tab of the Run Plan dialog box, under Details, select Refresh with current data.
Create a forecasting profile and output measure for using the history offset.
Attach the new forecasting profile to the plan, and set the history offset.
Run the plan with the new forecasting profile.

On the Parameters tab of the Run Plan dialog box, under Details, select Do not refresh with current data. For as long as you're fine-tuning the statistical forecast, you don't need to select Refresh with current data.
Compare the forecast with the actual data, and fine-tune the forecasting methods, causal factors, and forecasting parameters.
Make changes to the forecasting methods, causal factors, and forecasting parameters in the forecasting profile that you want to use for the statistical forecast, and run the plan with the forecasting profile.