Top Down, Bottom Up, or Middle Out?

This article provides an overview of the two aggregate forecasting methods available in Advanced Planning Forecasting.

Forecast Aggregation Methods

There are two primary forecast aggregation methods available in DemandCaster Advanced Planning.

  • Top-down aggregates demand at an upper level of the hierarchy, generates a forecast at that level, and distributes it down to the lower levels of the hierarchy with a weighted distribution of the entities within the level. The level chosen can be at any level including the lowest level.
  • Bottom Up is a top down but running the forecast at the lowest level of the hierarchy. In this case, there is no percent distribution down but rather a bottom up aggregation of the forecast values.
  • Middle Out is an expansion on top down in that the forecast is generated at two levels, an upper and lower, and then combined in a proprietary manner to form a composite forecast.
    • The upper level forms the shape of the overall demand pattern
    • The bottom level forecast establishes the basis for the period by distribution of the overall demand patterns through the lower levels
    • The upper level forecast does not change with the bottom distribution – it establishes the overall demand plan for the category.

The Forecast Settings are accessed by clicking "Run Forecast"

Top Down

When this option is selected, only one level may be selected to generate the forecast. In the example below, the choices are Company, Customer, or Item.

Forecast Aggregation Methods


When this option is selected, two levels may be selected to generate the forecast. In the example below, we have selected Customer and Item. This process creates a composite forecast that combines the top level forecast with the lower level forecast. The lower level forecast serves as the basis for the weighted distribution of the top-level forecast.

Top Down versus Middle Out

The examples below are an example of a top level forecast using the Top Down method at the Customer level and Middle-Out at the Customer and Item levels.

Top-Down Forecast Result at the Customer Level

The forecast is generated at the top-level via the top-levels history.

Top-Down Forecast Result at the Item Level

Example 1: You can see that the pattern of the item level is a replica of the top down forecast but at a smaller volume. The volume is dictated by the historical participation of the item at the overall level.

Example 2: Same as above but at a smaller overall volume.

Middle-Out Forecast Result at the Customer Level

The middle-out statistical forecast result at the top "Customer" level is most often identical for both the top-down and middle-out methods for the customer "Salamanca." This is because the top level sets the forecast basis for all the levels below. The top level statistical forecast is only slightly influenced by the bottom levels in most instances. The middle-out method primarily affects the forecast results of the bottom forecast level.

Middle-Out Forecast Result Item Level

Example 1: The forecast at this level is informed not only by the pattern at the top-level but the items own forecast. The pattern does not have flattened peaks like the top-down method. Its historical pattern follows more closely to the historical pattern of its history based on the forecasting algorithm chosen which in this case is triple exponential smoothing.

Example 2

What level to forecast at? (Video)

Please find below a short video explaining the difference between top-down and middle-out, and suggesting levels to forecast at.


The negative of the middle out method is that it takes a lot longer to run since the number of statistical forecasts that need to be generated, increases exponentially. We do recommend running middle out when there is a clear seasonality at the upper level that is not as defined as the lower level. This way, you are able to generate a hybrid item level forecasts that is influence by both patterns.

We recommend bottom-up when the item and customer/channel/market relationship is 1 to 1 and each item has its own discrete pattern.

For most companies, however, running a top down forecast at a low enough level where the demand pattern is captured, is the best approach particularly for those companies that sell products with many attributes such as clothing.


Add your comment

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.