Challenge:


Inability to cleanse the history adversely affects the forecast accuracy.

There was a need to cleanse the history of products which are highly affected by one-off events such as promotions (or projects), to get a better baseline forecast.

It is not always possible to get the past data for one-off events for history cleansing. This means either we use automatic outlier corrections, or an alternative method to smooth the history, to achieve a better baseline forecast.
 

Solution:

Use of outlier corrections affected the seasonal peaks and client was not able to get the right seasonality in the baseline.

Forecast in monthly buckets smoothed the history, but introduced a forecast lag and the weekly profile was lost.

For these highly promoted and seasonal products, a new method of smoothing was developed, based on advanced macros. It addressed the issues mentioned above by using ‘centred moving average'. This resulted in a much better baseline forecast.

Graphic for a group of seasonal products:


Brown:
Adjusted History
Black: Smoothed history
Purple: Earlier Forecast using Adjusted history
Blue: Forecast using smoothed history

This problem was reviewed and the solution was developed in the production environment. A short training course was conducted to explain various forecast models in APO and the limitations, and the advantages of the new smoothing technique.

Benefit
This method of history smoothing was adopted to achieve forecast accuracy improvements.
 
Sitemap