Forecasting Process
Running a Forecast
Once the forecast hierarchy and settings established, click "Run Forecast" to commence the forecasting process.
Upon opening the popup, the last time the forecast was run, the hierarchy selected and period settings from the last forecast run are applied. If no date is present, it means that a forecast has not been run with the set hierarchy.
These settings may be modified as follows:
- Aggregation Type: Select either top-down or middle out to select the type of aggregate forecasting process. See below to understand how the different options produce a forecast. Additional information regarding the aggregation type is provided in the article Top Down, Bottom Up, or Middle Out.
- Bucket Type: Select monthly or weekly buckets - we recommend choosing a single bucketing method consistently.
-
Forecasting Hierarchy:
- The forecast levels included in the hierarchy are presented. If a level is removed or change, the selections will in turn change.
- If top down is selected as the aggregation type, only one level is allowed to be selected as the forecast level.
- If middle-out is selected, two levels are selected as the forecast levels.
Forecast settings are shared for both manual forecast analysis and automated forecast analysis. Automated forecast uses the latest saved forecast settings (used during last manual run).

If an incorrect selection is made, an error will be displayed and the forecast will not be able to be executed. In the example below, a second level is required to fulfill the requirements of the Middle-Out aggregation type.

Running a Forecast
Once the options are selected, click "Run Forecast" in the pop-up to commence the forecasting process.
The process runs the statistical forecasts and calculates the disaggregation percentages necessary to drive the forecasts down to the base (PCL) entity level, sums up forecasts at each level of the hierarchy, calculates sales dollars and average sales prices.
The process of how the system selects the appropriate algorithms to apply are covered below.
Forecast Overview
CoV and ADI help the system apply the proper forecasting algorithm to the level and nodes where the forecasts are generated. The CoV and ADI threshold settings are applied in system settings during the forecast setup.
The first step is to establish the different types of demand pattern categories to forecast.
To determine the characteristics of a forecast entities demand history, the ADI and CoV coefficients are used:
- ADI: Average Distribution Interval measures the regularity of a demand in time by computing the average interval between two demands.
- CoV: Coefficient of Variation measures the variation in the demand quantities.
ADI
To calculate ADI, we measure the average interval between two demands over the entire history. Per the following example:

The calculation for ADI is calculated as:

Per above, demand occurs on average every 1.83 periods.
Demand Profile
To compute the coefficient of variation, we will only consider the non-zero values of the demand history. In the example above, the average quantity equals to 10.57 while the standard deviation equals to 6.43.

Based on these 2 dimensions and the thresholds for these values set within system settings, we classify the demand profiles into 4 different categories. This will run with each new planning period as a prior bucket period passes. This analysis will be calculated in the forecast analytic page.
- ADI Threshold: 1.32 (this will be the default setting that will be overwritten)
- CoV Threshold: 0.70 (this will be the default setting that will be overwritten)
The above will fill in the thresholds to categorize the history that will be the basis for the forecast algorithm family selected:
- Smooth demand (ADI <= (ADI Threshold Value) and CoV <= (CoV Threshold Value)): The demand is very regular in time and in quantity. Apply exponential smoothing models - single, double, triple
- Intermittent demand (ADI > (ADI Threshold Value) and CoV <= (CoV Threshold Value)): The demand history shows very little variation in demand quantity but a high variation in the interval between two demands. Use Croston's (chosen best from Croston/Croston Standard/ Croston Syntetos Boylan) if demand does not exceed 25 units per period or Croston Syntetos Boylan model if greater than 25.
- Erratic demand (ADI <= (ADI Threshold Value) and CoV > (CoV Threshold Value)): The demand has regular occurrences in time with high quantity variations. Use Syntetos Boylan model only.
- Lumpy demand (ADI > (ADI Threshold Value) and CoV > (CoV Threshold Value)): The demand is characterized by a large variation in the quantity of demand and in the interval between two demands. This is quite impossible demand to reliable forecast, no matter what forecasting tools and methods are being used. Use bootstrapping.
In addition, if the forecasting history contains <= 2 buckets with values > 0, Simple Moving Average will be applied.
Top-Down Forecast
If Top-Down is selected, the forecast at an upper level of the hierarchy will superimposes the shape of the aggregate forecast down to the item by the distribution of the items history within the context of the top level. In the example below, a top-down customer forecast is generated. As can be seen the first item below the Customer level will take on 23.01% of the customer forecast across all future periods. This % allocation remains unless there are any user defined edits in any forward period that override the statistical value.
To view the distribution percentages, in the Demand Plan view, replicate the forecast hierarchy and then in the statistics tab, select the percent view as documented in the article Demand Planning Stacked Grid Editing.
Below is quick video that shows how the top level forecast is translated down to lower levels of the hierarchy when a top-down forecast is generated. Please note that the image below is from an earlier version of DemandCaster.

Middle-Out Forecast
The Middle-Out forecast approach combines two forecast levels. It takes the shape of the aggregate forecast and combines with the shape and trend of the item level forecast. This means that the percent distribution in each period will vary to take into account the statistical value of the lower level forecast relative to the upper level forecast. Though the middle-out approach takes longer to process since it is running twice the number of forecasts, we recommend it in most instances.
The image below shows how the distribution percentages vary period by period to account for the two forecasts that are generated and combined to form a composite forecast.
Below is quick video that shows how the top level forecast is translated down to lower levels of the hierarchy when a middle-out forecast is generated. Please note that the image below is from an earlier version of DemandCaster.

Helpful Information
The time required to generate the forecast is dependent on numerous factors including the number of customers, levels, and items being forecasted.
Please note that the forecast UI only presents the forecast results and the measures related to forecast. Any overrides to the forecast values will be performed in the Demand Planning section.
0 Comments
Add your comment