- Outlier correction is in the pre-processing phrase: All subsequent processing, for example, forecasting, is impacted.
- Outlier correction and forecasting are interdependent: Different forecasting algorithms work best with different outlier correction.
- Proper outlier handling requires expertise:
- Which detection algorithm to use?
- Which parameter set?
- Which correction method?
- High complexity and risk: Great opportunity to apply AI.
In the voting mechanism, we consider a data point an outlier if the majority of voters say that it is an outlier. Each detection algorithm has edge cases, false positive or false negative, but combining different outlier detection algorithms we can make outliers detection more robust and reduce the false positives and negatives.
The following are tested methods for detecting outliers:
- Probabilistic methods like MAD.
- Proximity-Based Methods DBSCAN.
- Outlier Ensembles like Isolation Forest.
Agenda
Automated outlier correction
Product Lifecycle Management enhancements
- Additional use cases
- Aggregated product assignments
Curve-based forecasting enhancement
A new solution for product lifecycle management is being phased in with IBP 2405, bringing great new functions.
The new PLM Engine can be enabled for a group of users. Accordingly, it can run together with the legacy solutions: some users can keep using the old solution, while others can experiment with the new.
- To activate the new PLM Engine, set Global Configuration Parameter "USE_NEW_PLM_ENGINE."
- Assign the user to the group specified in global parameter "USER_GROUP_NEW_PLM_ENGINE."
The new PLM Engine comes with great new features, and some temporary limitations.
Out of the various forecast dates used in PLM, Segmentation, Forecast automation, and Curve clustering only considers Phase-in End date to construct the final history.
New features
- Available for ABC/XYZ Segmentation
- Available for Forecast Automation
- Available for Curve Clustering
- Aggregated handling of product assignments is available
Restrictions in IBP 2405
- Planning objects are not generated automatically: You can use the Copy operator to generate missing planning objects in advance.
- Forecast start dates in the future are not considered.
- Planning objects without forecast date entries (no * entry either) are not dismissed.
To ensure data consistency, PLM must either be used by all connected processes, or by none of them.
Examples
- PLM-enabled forecast models will only consider the results of PLM-enabled forecast automation profiles
- XYZ segmentation with PLM disabled can only consider results of forecast automation profiles where PLM is disabled
Exception
PLM enablement does not have to match between forecast models and curve clustering profiles.
Product Lifecycle Management: aggregated product assignments.
With aggregated product assignments, applications can consider product assignments, even if those are maintained on a different level.
- Aggregated product assignments can be enabled for planning areas.
- To use aggregated product assignments, the new PLM engine must be enabled for the user.
- On planning areas where aggregated product assignments are enabled, all applications will consider product assignments in the aggregated way.
- Aggregated product assignment is a powerful feature of PLM, but understanding the results can be significantly more difficult.
- Aggregated product assignment is independent from Aggregated Phase-in and Phase-out planning:
- Aggregated Phase-in and Phase-out planning is a function of the forecast engine. Other applications don't have access to this feature.
- Aggregated Phase-in and Phase-out planning can be used even if Aggregated product assignment is disabled.
Simple Assignments: All PLM attributes must be present in calculation level of the consuming job.
Aggregated assignments: All PLM attributes must be present in the Base Planning Level of the input key figure of the consuming job.
* PLM Attributes: Level 1, 2, 3 and Launch dimension 1, 2, according to Planning Area PLM settings.
Simple and Aggregated Product Assignments and Dates: Examples
| Request Level | Assignment Level | Launch Dimensions | Assignments | Dates |
|---|---|---|---|---|
| Product ID | Product ID | No dates for Level 1 | Simple | None |
| Product ID | Product ID | Product ID | Simple | Simple |
| Product ID - Location ID | Product ID | Product ID | Simple | Simple |
| Product ID - Location ID | Product ID | Location ID | Simple | Simple |
| Product Group | Product ID | Location ID | Aggregated | Aggregated |
| Product ID | Product Group | Location ID | None | None |
| Product ID | Product ID - Location ID | Location ID | Aggregated | Aggregated |
| Product ID | Product ID - Location ID | Product ID | Aggregated | Simple |
| Product Group - Location ID | Product ID | Location ID | Aggregated | Aggregated |
| Product Family | Product Group | Location ID | Aggregated | None |
| Product ID - Location ID | Product ID - Location region | Customer ID | Simple | Aggregated |
* All following attributes are part of the base planning level of the PLM input key figure.
Automated Outlier Correction
Product Lifecycle Management Enhancements
- Additional use cases.
- Aggregated product assignments.
Curve-Based Forecasting Enhancement
- Curve-based forecasting: specify curve clustering profile to use.
- The options for selecting curve clustering profiles in forecasting have been extended.
Now, you can manually specify a clustering profile. Note the following points:
- The periodicity and input key figure of the selected clustering profile must match those of the forecast model.
- Since the calculation level of forecasting is determined during runtime, validating the calculation level upon saving the forecast model is not possible.
- If the forecast model is executed on a calculation level that is not equivalent with the one on the clustering profile, then forecast will not be calculated.
- With this option, you can compare forecast accuracy using different clustering settings. For example, you can add two cluster-based forecasting algorithms to the forecast model, referring to different clustering profiles, with different number of clusters specified. Forecast engine will calculate forecast using the results of both clustering profiles, and pick the one where forecast is more accurate.
- General accuracy improvements are also delivered for curve-based forecasting in IBP 2405.
When the option "Consider Trend" is selected, a short-term trend feature is added to the forecast model. This feature is derived from sales data and handles the short-term trends (the next four weeks) of a sales cycle.
The "Consider Trend" feature uses feature engineering to create a additional independent variables derived from Sales data to better handle short-term (the next four weeks) trend of Sales cycle:
- Slope (the first order derivative or the rate of change between two consecutive Sales points.)
- Slope_of_slope (the second order derivative of how the slope changes.)
- Cycle_length (the counter of the short-term Sales cycle.)