Interpretation of Demand Forecasting Hierarchy

Question

I am occupied with hierarchical demand forecasting (mainly about consumer packaged goods) and i have a question about the interpretation of its structure.

Let us assume that the hierarchy is as follows:

Total Demand for the Company --> Product Categories --> Products (SKU) --> Customers.

Let us assume that the product categories are chocolate, yogurt, and biscuits. The total demand for the company one time period could add up to e.g. 10,200 units of products that are not related e.g. 5000 units of yogurt 200gr 2%, plus 2000 units of yogurt 300gr 4%, plus 1000 chocolates 200gr, plus 1000 chocolates 80 gr, plus 500 units of chocolate biscuits 250 gr, plus 700 units of cocoa biscuits 300gr. In the top level of the hierarchy I end up adding irrelevant things and the aggregation does not make any sense. What is the business value of knowing that the demand is 10,200 units of irrelevant products?

In forecasting software the hierarchical facilities are very common and also a lot of vendors support reconciliation capabilities. What I don’t understand is the interpretation of aggregating different things in an upper level of the hierarchy.

In some cases aggregation makes sense. E.g. in a tourism forecasting application the hierarchy could be Europe --> Country --> Prefecture. If we assume that one unit=one tourist (person) then the aggregation makes sense. But what about in the case of consumer packaged goods?

Answer 1

You raise an extremely good point. Aggregation should not be done mindlessly. In fact, this applies both to forecasting aggregates and to aggregating historical data. I would say that a good test whether an aggregate forecast makes any kind of sense is whether or not aggregating the same time series historically is something you would ever want to do.

So now, in a way, we have moved away from forecasting and into accounting, business planning and so forth.

And some kinds of aggregation, both historically and for forecasting, do make sense. Let's concentrate on forecasting, since this is the context of the original question.

• Your products may all look very different, but they may all use one or more ingredients, e.g., sugar or milk, in varying quantities. In that case, you may want to have an aggregate forecast of the sugar or milk you will need in total to produce your end products.

• If you do not produce, but buy and sell your products (e.g., as a retailer or wholesaler), you will need to plan and budget your cash. If you source multiple products from a single CPG (consumer packaged goods) manufacturer, you may need to plan/budget your purchases from him on an aggregate level, especially in haggling over discounts. So you would need forecasts of total sales.

• The same applies conversely to the CPG manufacturer, who will sell multiple products to different retailers and will need a forecast for his total sales to each retailer.

• So far, we didn't really look at forecasting aggregate units (instead, we needed total weighted forecasts, where the weights would be sugar or milk content, or sales price). However, you may also be interested in forecasting total units, e.g., in workforce planning. When Amazon or others plan shifts in their warehouse, they will need to forecast the total number of so-called "picks", i.e., how many things will need to be physically taken from shelves and packaged. Similar questions arise if you think about building a new warehouse.

Of course, all these examples do not address the issue of whether the forecasts are best calculated at the top level (take the time series of total sugar consumption or total picks and simply forecast this total time series) or bottom-up (forecast each component separately and aggregate the forecasts, using weights as appropriate) or in yet another way.