1. Bottom-Up Hierarchical Forecasting
Up until now we have focused a lot on forecasting one product. However, most of the time there are numerous product sales that need to be forecasted along with sales forecasts for whole regions and even sales for the entire company! That is where we need to consider hierarchical forecasting.
2. Hierarchical Forecasting
The beauty of hierarchical structures to data is that you can take advantage of those structures when it comes to forecasting. You can use hierarchical forecasting when your data aligns itself in a natural hierarchy. The hierarchy adds connections up and down your data. These connections need to be reconciled in the hierarchy. Let me show you what I mean.
3. Hierarchical Structure
In our data we have three regions all within a state. In each of these three regions we have products. In the videos we have been forecasting the high value product in the mountain region. We can also forecast the low value product in that region as well as the mountain region altogether. Obviously, we want our sales forecast for the whole mountain region to equal the sum of all the sales forecasts for both products in the mountain region.
4. Types of Hierarchical Forecasting
There are three different types of hierarchical forecasting that we can use.
First, bottom-up forecasting where we forecast every product then reconcile our forecast up.
Next, top-down forecasting where we forecast the top of the hierarchy and reconcile our forecast down.
Lastly, middle-out forecasting where we forecast a middle layer of the hierarchy and reconcile our forecast both up and down.
5. Bottom-up Forecasting
Bottom-up forecasting is by far the easiest to implement, but also the most time consuming. Quite simply, to perform a bottom-up forecast, you just need to forecast everything on the bottom of the hierarchy and then add them together up the hierarchy for reconciliation.
For example, we need to forecast our high value product sales and low value product sales, then add them together. That summation will be the total sales forecast for the mountain region.
6. Bottom-up Forecasting Example
I've done you a favor and went ahead and built a model for the low value product in the mountain region. Now all we need is to add up the forecasts for the high value and low value products' sales. Easy as that!
Let's see how good this forecast is in terms of MAPE. Looks like we were off by less than 8% on average.
7. Let's practice!
Time for you to start building the forecast for the metropolitan region from the bottom-up!