Forecast Configuration — Promotion Model Settings

Algorithm: Available options are Decision Tree and Generalized Linear Models (GLM).

Minimum samples: Minimum number of data points for each leaf node of the decision tree. This is used to ensure there are enough data points for each node.

Maximum depth: Number of levels between the leaf node and the top of the decision tree. Start with a value between 4 and 6, depending on the number of features being used and increase the number gradually to see if the model performance improves.

Partition: The model can be partitioned, based on one or more of the features used for building the model. Selecting a partition builds a separate model for each value corresponding to the chosen partition. Multiple features can be selected to determine a partition. Partitioning by higher levels in the product hierarchy and the location hierarchy or based on relevant product/location attributes can help to determine the promotions lifts more accurately. Lower levels of product/location hierarchy may not have sufficient data to create reliable parameters.

Reliability metrics:

High percentile: A value of 0.9 removes the top 10 percent of the observed lift values for building the model. Adjust the setting to remove outlier values with very high lift.

Low percentile: A value of 0.1 remove the bottom 10 percent of the observed lift values for building the model. Adjust the setting to remove outlier values with very low lift.

Figure 11-32 Model Settings for Promotion EffectsThis image shows promotion effects output review.

Review the output of the promotion effects in Innovation Workbench (IW):
  • PMO_LLC_PL_MODEL table: Contains the details about the promotion model.
  • RUN_HDR_ID: Estimation run id.
  • CASE_TABLE, CASE_COLUMNS: Contains all the features available for the user to select.
  • VIEW_NAME, VIEW_COLUMN_NAMES: Based on the user selection in the UI.
  • Using IW, MODEL_NAME can be used to apply the model for a chosen time frame on the CASE_TABLE.
  • Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE): Error metrics for the promotion lift values.
  • Average Error (AVG_ERROR), Median Error (MEDIAN_ERROR), and Weighted Error (ERROR_WGT): % Error metrics for the promotion lift values.
  • ERROR_WGT: % Error weighted by sales units.

Figure 11-33 Promotion Effects Output Review in IWThis image shows the promotion effects output.