Forecasting: Adjusting the Pyramid Strategy

This article is part of the Forecasting series, split from the original Forecasting article.

Adjusting the Pyramid Strategy

Arkieva uses machine learning to select methods based on accuracy and robustness over time.

Adjust Strategy

There are several ways of running a forecast with Arkieva, ranging from easy to use (Simplified), to more complex models (Segmented). Arkieva provides an easy-to-use Forecast Strategy interface to define your forecast parameters.

Click the Adjust Strategy button to launch the Edit Forecast Strategy drawer

1. Forecast Planning Bucket

Under the Forecast Bucket section, select whether your forecast is calculated on a weekly or monthly basis.

2. Outlier Handling

Under the Outlier Handling section, you can select whether or not to calculate the forecast using published adjusted values by checking or unchecking the checkbox. Using published adjusted values assumes you have adjusted flagged outliers or made other types of adjustments.

3. Global Filter

Select a defined global filter (from the Settings Manage Filter Preset) to filter data based on a combination of attribute values by selecting items or locations. The Global Filter applies to any filter type except the "Segmented Forecast".

Click the Manage Filters button to create a Global Filter if the Global Filter dropdown is empty. Click +Add Filter to create a Global Filter.

4. Forecast Error measure and Historical Error weight

Select an Error measure from the dropdown, then choose an Error Weight.

HWGHTD is a variant of WMAPE that evaluates forecast methods using a broader set of information:

WMAPE evaluates forecast methods primarily based on performance during a 4-period comparison (test comparison) window.
HWGHTD considers both

Performance within the 4-period test window, and historical fit quality across a longer horizon (typically 32 periods called as training window).

By balancing historical fit quality and recent forecast performance through a user-defined Error Weight, HWGHTD provides a more robust approach to forecast method selection and can improve overall forecast accuracy and forecast quality.

The Historical Error Weight is a business/user-defined parameter that determines how much importance is assigned to the historical training window versus the test comparison window. The specified weight is applied to the historical training-window error, and the remaining weight (100% minus the Error Weight) is applied to the test-window.

Error Weight Setting	Emphasis	Effect
Higher test weight (e.g., 80% test / 20% train)	Recent forecast performance	More responsive to recent changes
Balanced (e.g., 50% / 50%)	Both recent and historical performance	Stable compromise
Higher train weight (e.g., 20% test / 80% train)	Long-term fit quality	More stable, less sensitive to short-term fluctuations

5. Measurement Level

The Measurement Level section in Edit Forecast Strategy lets you choose the attribute level at which pyramid forecast quality is measured, instead of always measuring at the base Item, Location, and Customer level.

Each selector lets you choose the attribute that pyramid quality is measured against for that dimension. By default, this is Item, Location, and Customer themselves, the base level. Some data models need quality measured at a different attribute instead. For example, if an item represents one packaging variant of a broader product, the attribute of real interest may be the Item Group that variant belongs to, not the individual packaging variant itself. Measurement Level lets you select Item Group (or another configured attribute, such as Location Region) in place of the base attribute.

The Attribute Must Belong to the Pyramid. The attribute you choose for Measurement Level must exist somewhere in the strategy's own pyramid. If you select an attribute that isn't part of the pyramid you've built, a warning appears and the Apply button is blocked until the mismatch is corrected. Pyramid quality is normally compared level to level, for example, comparing the accuracy of a Level 2 aggregate against the base Level 1. Measurement Level lets that comparison happen at the attribute that actually matters for your data model, rather than always defaulting to Item, Location, and Customer.

6. Forecast Type

Under the Forecast Pyramid section, select attributes to build your strategy pyramid.

Auto-Learning

Given a user selected set of data attributes, Auto-Learning automatically creates and selects the best pyramid for forecasting for you. Independent attributes are always included in the run; this can only be modified with the appropriate access. You may add dependent attributes if so desired.

Select your demand ILC attributes and click Apply. Arkieva's machine learning will scan through multiple pyramids and select the optimal Pyramid and the forecast using the selected Pyramid.

The Auto-Learning output shows the Optimal Pyramid with forecast accuracy measure. Each level of the pyramid is measured based on the forecast accuracy of an item.

A drilldown function starting with Item is available after the run is complete. Expanding these items provide details of the remaining attributes such as Location and Customer.

Manual Auto

Manual Auto or Manual Learning assumes you understand the concept of a pyramid forecast and you have a specific pyramid in mind. The lowest level, Level 1, uses independent attributes and is already set because that is the level we wish to implement the forecast.

To add additional levels, click +Add Level to add a new Level, then click Add Attribute.
Click Apply to run to build your strategy pyramid.

The Manual Auto output shows the defined pyramid with forecast accuracy measure.

A drilldown function starting with Item level is available after the run. Expanding items provide details of the remaining attributes such as Location and customer.

Simplified

Simplified forecasting is Arkieva's basic forecast module and runs the forecast for the tuple level.

You will need to use the Global filter to restrict the data, and select a method from the dropdown (Average MAD w/ Sales Adjustments, Removed Method). Click Apply to run the Simplified forecast.

Simplified forecast displays the forecast at the tuple level, and displays only the selected global filter.

Segmented

Segmented forecast allows you to construct a forecast structure one segment at a time. The segment consists of a partition which defines the attribute values allowed, a selected method, and an aggregation level (forecast level). You can create several segments that would run sequentially, and the segment partitions can overlap.

For each segment, you have complete control over the following:

Control of data that is processed in each partition. Select any combination of attribute values for the segment; the forecast will be dependent on the defined partition.
Control over Methods is selected. Each segment requires a method, and you can select from a list of custom methods; new methods can be added.
Control the level of aggregation of the data. This defines the level of attribute combination to forecast. For example, you may want to forecast the Item, Location and Customer level or just Item and Location.
Control to define the relationship if any between segments. Relationships between segments are possible but not required.

Each Segment is saved and disaggregated to the lower levels, so be careful of how you construct each level.

Segmented Example:

A basic segmented forecast may be a single segment at the ILC level of aggregation with a partition for a specific item. The forecast is only a process for the partition.

A Segmented forecast can be set up in any way you choose. Let’s say you want to create pyramid to forecast 3 items: A, B and C.

Segmented set up example 1

Seg-1: we can partition such that only item A is selected, and data aggregation is at the item, location and customer level.

Seq-2: we can partition such that only item B is selected, and data aggregation is at the item, location and customer level.

Seq-3: we can partition such that only item C is selected, and data aggregation is at the item, location and customer level.

Seq-4: we can partition for items A, B, and C are selected, and data aggregation is at the item, location and customer level.

Partition 4 overlaps with all other partitions. This requires that each segment has a dedicated method.

Segmented set up example 2

Seg-1 we can partition that items A, B and C are selected, and data aggregation is at the Item, location and customer level.

Seq-2 we can partition such that items A, B and C are selected and data aggregation is at the Item level.

Partition 2 overlaps with partition 1, both are restricted to items A, B and C.

The difference between the 2 examples above is that we are forced to use a dedicated method for segment. Until a run is published, it is not available in the collaboration environment.

Create a Segmented forecast

To create a Segmented forecast, first click +Add Segment.

Next, click Create data partition filter. The Filter drawer will slide out; create a filter for the partition. Click Apply.

Next, select a method from the dropdown. If no methods have been created, the dropdown will be blank; you will need to create a method.

Next, select a combination of attributes for the aggregation level. for example we will select Item, Location, and Customer.

When a segment is completed, you can click the +Add Segment to add another level to the forecast pyramid.

When finished created segments, click Apply to generate the segmented forecast strategy.

Segmented forecast will output:

All data at the tuple level is displayed, both defined and undefined partition.
The forecast data is disaggregated only for the defined partition to the tuple level.
Any unpartitioned data would show the results of the previous run unless deleted.

Forecasting series