Forecast Reconciliation

Dynamic Proportions

Dynamic Proportions method is time-dependant and forward looking. So it creates proportions that are different each month so it incorporates seasonality. This method always uses some version of the forecast so it is forward looking. The moment you start using future periods, you also incorporate trend influences in your proportioning logic.

This method is also called pro-rata logic since this is using proportions on the basis of forecasts that already exist for both SKU and the Brand. Generally you will proceed using the following steps to create dynamic proportions:

1. Let’s assume you have the same five SKUs in each brand. You will first build models at the SKU level and create SKU level forecasts for each month in the horizon.
   
2. Then build a model at the brand level to derive the higher level forecasts also by month.
3. Dividing 1 above by the brand level forecast by month will give you a dynamic proportion that varies by SKU by month. These proportions will then be adjusted for the forecast at the lower level every time the forecast is changed. That proportion is then used to disaggregate the model at the higher level.

This logic has a variety of advantages:

1. If the SKU is declining, the declining trend will be built into the SKU level forecast, affecting the higher level model as well as influencing the proportions.
2. If there is seasonality, it will be factored into the SKU-to-Brand proportions as well.
3. When the total brand level forecast is disaggregated, you should get a clean disaggregation, because the proportions now incorporate both trend and seasonality using the latest forecast models.

However, there are some criticisms against this technique as well. The main argument is that lower level forecasts are inferior to begin with. So why use that to develop the proportions?

There is some merit to this argument since the seasonality and some of the patterns at the lower level may not be very clean due to additional noise in the data. However in reality, persistent trend would be reflected in the proportions fairly well. As in the case of most CPG/FMCG companies faced with phase-in/phase-out issues, this method would work well even if more complex models cannot be created at the lower level.

Effect of Changing Forecast Models and Proportions

Since the business world is constantly facing changes, we will experience changes to the forecasts very often either induced by the models or by management over-rides. So let us look at the effect of changes you make to the forecasts or to the proportions. Most forecasting systems that have proportioning logic try to keep the hierarchy of forecasts internally consistent.

1. Re-model one of the five SKU’s and change the forecast.
   1. This will trigger a change in the higher level forecast based on that one changed forecast
   2. Changes the sum of the forecasts at the Brand level.
   3. Creates new set of proportions for each SKU. So all SKU level proportions are now changed.
   4. Lower level forecasts will stay the same except for the SKU with the new model until you re-model the higher level forecast. If you now remodel the brand level forecast, that would trigger a change in the forecasts for all five SKU’s, because the system will assume the proportions should stay the same.
2. Manually change the proportion of one of the SKUs
   1. When you change one SKUs proportions, this will lock that proportion to the total.
   2. The system will re-allocate the remaining forecast to the other four SKUs based on the old proportions within those four skus.
3. Manually Change the forecast at the Brand level
   1. The System will disaggregate the forecast to the SKU level based on existing proportions.
   2. The System will aggregate the new forecasts to the category level and create new proportions at the brand level with other brands that roll up into that category.

Conclusion

In summary, it is important to understand the different approaches to keep the hierarchy of forecasts internally consistent. Although different methods are available, time dependent dynamic proportions comes out as a clear winner.

<<Previous Page: Using Forecast Proportions