New Forecast Engine Event Modelling method

The forecast engine is enhanced to include an additional event modeling method: Dynamic Events with Bayesian. This is added as an additional option in the forecasting engine for Event-based modelling, as an alternate to existing event processing.

The ’perform event modeling’" switch to control event modeling in the forecast engine in previous versions is migrated to a new property Event Modeling Method with a selection, from three options:

  • None: Equivalent of switch Off (False) in previous versions. All existing forecast engines with the switch Off (False) are upgraded to this value.
  • Least Squares: Equivalent of switch On (True) in previous versions. All existing forecast engines with the switch On (True) are upgraded to this value.
  • Dynamic Event with Bayesian: New method

A new parameter Minimum Events Length is added to the Least Squares method to define the number of history periods without an event to determine a reasonable baseline model.

These standard forecast engine definitions are updated:

  • Event based Holt Winters Months
  • Event based Holt Winters Weeks

Behavior of existing forecast engines is maintained on upgrade.

For the Dynamic Events with Bayesian all history time periods are used, there is no minimum events length. These are a few factors for converting historical events to independent variables:

  • There is one independent variable (IV) for each event type. The IV takes the values of 0 where there is no event magnitude on a time period for that event type. Otherwise, it uses the passed event magnitude.
  • A BATS model is fit to the history using a level component and IVs that represent each event type. These IV coefficients represent the event sizes for each event type.
  • Once the average event sizes are calculated the method returns to the normal event modelling process.

This new event method supports the case where events occur across many historical periods since the event does not require removal of event history periods to calculate a baseline. The baseline calculation considers periods with events + periods without. The regression component has a discount factor associated with it, wherein the event size can be weighted more to events closer to the present. This regression factor is currently set internally (hard-coded).

A typical use case for this method can be heavily promoted items where there is not enough non-event history periods to establish a reasonable baseline without events. The dynamic events do not have to be promotional in nature. The dynamic events can be used to measure correlation of other event or profile data on the demand and apply similar events to the future baseline forecast.

Event-based forecasting now supports multiple event types, however, two event types for the same historical period is not allowed. This restriction is planned to be lifted for the Dynamic Events with Bayesian method in a future release.

These standard forecast engine definitions are added, each passing a single event profile measure (Events):

  • Best ML Combination CMonths with Dynamic Events
  • Best ML Combination Months with Dynamic Events
  • Best ML Combination Weeks with Dynamic Events

Additional macros are added to the content for calling the individual Dynamic Events engines across all item-location combinations:

  • Generate Best Combination DynamicEvents CMonths (Demand)
  • Generate Best Combination DynamicEvents Months (Demand)
  • Generate Best Combination DynamicEvents Weeks (Demand)
Note: This feature is enabled on update, however related content changes are available only after loading the demand planning template (dpls.zip) for this version. You are not required a new role or privilege access to use this feature.