Special days

Special days play a vital role in forecasting. They are used to forecast performance for scheduled events such as holidays or special sales. They allow for more accurate forecasting by indicating what days should not be used in the calculations for normal forecasting.

Skip special day

This option allows users to exclude low or high performance days from future forecast trends. This option is typically used on any day where the sales volume was extremely low (for example, store is closed or closed early) and on days where the sales volume was unusually high (for example, a special sporting event that will not occur yearly). By skipping these special days, the system can forecast more accurately.

When a day is skipped, linear regression will not include that day in the calculations, even if the day falls within the historical period that is selected for the forecast. Trend of Historic Averages goes back an extra week to find 8 weeks of non-skipped days for the trend calculation. If the “Same day of last year” happens to be skipped, the regression averages the weeks before and after to come up with a LY value for that day.

See Defining special days.

Skipped special days are considered in the machine learning forecasting process. When training a store, the machine learning model analyzes the extent to which special days produce lower sales than normal days. Special days are incorporated into the forecast model to prevent these lower results from affecting forecasts for normal days. For stores using machine learning forecasting, special days should only be marked for unexpected closures or other events with unusually low sales volume.

See Special days in machine learning.

Forecasting for special days

A future date can be marked as a special day when a known event is planned. An example of this might be Black Friday. The system has the option for a percentage adjustment to the regular forecast, or the system can be set up to forecast from special days in the past.

See Defining special days.

You can forecast volume for special events using the Trend of Historic Averages for all special days or from special days on the same day of the week. The special day may occur on different day of week/date from previous year, and the historical period referenced should be longer than a year since these events may only occur at most once a year.

Note: The special day must occur at least once per year and occurs relatively in the same seasonal period year to year. Generally it is not accurate to forecast for a sale in the fall using historical sales from last spring.
Note: This option is not supported by machine learning forecasting, which uses the Promotion Calendar to track and forecast sales events.

This solution is available by default for new installations (as of 6.2.3) but is not available by default for upgrades. The class for this core implementation (com.workbrain.util.callouts.scheduling.SpecialDayCalloutCommonImpl) is specified in the /system/callouts/SpecialDayCallout registry entry.