Best Practices for Predictive Planning: General Tips and Tricks
In SAP Analytics Cloud, Predictive Planning provides business analysts with the power to accelerate planning cycles and build more accurate and detailed business plans.
In the first part of our Best Practices for Predictive Planning blog series, explore our top tips and tricks for Predictive Planning in SAP Analytics Cloud. In the second part of the series, we’ll cover advanced techniques that will help you make the most out of your Predictive Planning experience.
Table of Contents
- Prepare: Filtering dimension members
- Prepare: Handling of 1000+ entities
- Prepare: Using hierarchies when defining the entity
- Predict: Dealing with calculated columns
- Predict: Choosing default or local currencies for your forecasts
- Review: Excluding entities with insufficient quality
- Apply: Choosing a location to save your forecasts
- Best practice: How to keep a private version after publishing predictive forecasts
- Entity: An entity is a set of variables (dimensions or attributes) you can select to create predictive forecasts. Entities define how you should split your training data source into subsets based on values of the combined variables forming the entity. A predictive model is generated for each entity to get the predictive forecasts adapted to the evolution of the predicted variable for this entity. For example, forecasting the revenue for each product family for each region in which your stores are located. An entity in this example corresponds to a product family combined with a region.
- HW-MAPE: The HW-MAPE (horizon-wide mean absolute percentage error) is the central quality criterion for evaluating the time-series forecast in predictive planning. It is the expected error for the forecast values in the selected forecast horizon
If you’re looking to forecast only for certain members of your dimension, this tip is for you! To effectively filter dimension members, you need to add a dimension attribute to the model. Here, you should only fill out the values for the specific members that you want to contain in the filter. Watch the video below for more details on how to filter dimension members when working with your predictive scenarios.
Predictive planning today has a restriction of a thousand entities. If you need to create a granular forecast beyond 1000 entities, check out this blog post and follow the requirements below:
- Several dimension attributes need to be defined. Each one is filled for only a subset of the dimension members so that together their values span all dimension members.
- Subsequently, predictive models can use the attributes in its entity definition and save respective forecasts to the same private version of the planning model.
- The final result is identical to a predictive model that had used the unfiltered dimension for its entity definition.
For a step by step guide on how to achieve this, refer to the video below:
When defining entities, only dimensions and their attributes can be selected. You can add attributes to the model that take the role of the hierarchy nodes and then use those in the entity definition.
- Level-based hierarchies: For level-based hierarchies, the approach is straight-forward. Instead of navigating the hierarchy itself, pick the relevant attributes that were used for defining the hierarchy. To filter certain elements of the hierarchy, create a derived attribute that contains a subset of the respective hierarchy members. You can then use this attribute to define the entity. Any members with blank attribute values won’t be considered when a forecast is computed.
- Parent-Child Hierarchies: In the case of parent-child hierarchies, you need to create one or several new attributes that reflect the different levels of the hierarchy. Be sure to only fill attributes for leaf members of the hierarchy and not for nodes. Once created, the attributes can then be used when defining the entity. As above, you can choose to filter elements by creating additional attributes and leaving their value blank for certain members.
Calculated columns cannot be selected as a target for the predictive model. Instead, you need to build independent predictive models for the different base values that feed into the calculation.
Example: In order to make a sales forecast where sales = price * quantity, you’d need to create independent predictive models for both price & quantity.
When forecasts are saved, they are stored in the related account. In the example above, this would be the accounts price and quantity. Since the formula for the calculated column “sales” is referencing back to those accounts, you’ll automatically have a forecast for sales once the forecasts for price and quantity are both saved.
Multinational companies generally plan in several currencies. Sometime they need to consolidate their planning in one currency. Sometimes they receive estimations in local currencies. When they wants to get predictive forecasts, they now have the choice to get them either in a default currency or in local currencies.
This blog post explains this best practice.
The central quality criterion for evaluating the time-series forecast in predictive planning is the horizon-wide mean absolute percentage error (HW-MAPE), which is the expected error for the forecast values in the selected forecast horizon. While this is the central metric for Smart Predict, other metrics are common to evaluate time-series forecasts like R2, Mean Absolute Error, etc. This means that rejecting an entity based on its MAPE is not always merited.
If you decide that the entity shouldn’t be used, follow one of the following methods:
- Filter on the respective dimension member as described in the previous section, Prepare: Filtering dimension members. If you retrain the predictive model, the entity will no longer be generated and written into the planning model.
- Alternatively, you can choose to save your forecasts into the planning model (i.e. the forecast of the unwanted entity) and manually delete the entity in the planning model.
Please note: If your entity definition comprises of more than one dimension/attribute, all entities that include that member will also be filtered out. Example: if you filter out member Spain of the COUNTRY dimension, then Road Bikes – Spain AND Mountain Bikes – Spain will both no longer be generated.
By default, forecasts are always saved to a prepared private version of the planning model and to the same account as the target.
Example: Forecasts for account sales are always saved to the account “Sales.” It is not possible to save them in another account like forecasted sales.
If the above does not fulfill your requirements, you should revert to data actions to reorganize the saved forecast data.
Watch the video below to learn how to consolidate independent forecasts into a single planning version:
Predictive forecasts are written back into a private version of the planning model. When a planner publishes a private version into a public version, a side effect is the deletion of the private version. This means that the planner cannot go back either to refine the settings of the predictive scenario to improve the predictive forecasts or to refine the planning model.
Discover Advanced Techniques for Predictive Planning
Check out Part 2 of this blog series to explore advanced best practices for navigating Predictive Planning in SAP Analytics Cloud.
Sign up for your free SAP Analytics Cloud trial and try out these Predictive features first-hand, today.