You’re reviewing several forecasts from various departments. The production report numbers state that they plan to procure automotive parts to manufacture X number of passenger cars.
The sales report says something else.
Imagine having a single source of (data) truth with which you can make better decisions. It’s no longer sufficient to simply add up all the bottom-level forecasts and "hope and pray" that they’ll add up.
Traditionally organizations use bottom-up forecasting - a projection of micro-level inputs to assess sales for a given set of months. However, bottom-up forecasting is not optimal because at less disaggregated levels (i.e. cars sold in a small region contrasted to worldwide sales), the volatility is usually higher, implying that it will be harder to make an accurate prediction.
If you only make a lot of “bottom-level” forecasts, and then sum them up to create the higher levels, i.e. regions and finally worldwide sales, you will aggregate not only the total number of sales, but also the total error if say that all these forecasts are a bit optimistic.
What about organizations employing only top-down forecasting?
Say that we only forecast worldwide sales and have decided beforehand that Europe accounts for 30% of total sales, the US for 40%, and RoW for 30% of the total.
In this scenario, the organization might not have factored in market trend shifts in these individual markets. These trend shifts might have predicted a downturn in some markets and the opposite in others.
Another disadvantage of only using a top-down approach is that it typically does not account for the #SKU-level forecasts, but rather for the total level. For example, an automotive manufacturer might have a figure accounting for the passenger car demand (we’ll take an arbitrary figure of 30%). However, there would be some difficulty in pinpointing what the 30% constitutes -- eg: how many cars will have big engines as supply chain planning forecasts traditionally provide this information.
This results in an inaccurate top-down forecast.
How does hierarchical forecasting help with aligning the results of these forecasts?
By using optimal forecast reconciliation, forecasting every possible level in the hierarchy is possible.
This reconciliation benefits forecasts at every level of the hierarchy in terms of accuracy.
How is this done?
Let's begin by taking the individual forecasts and specifying the respective hierarchy and the residuals of the involved models.
These forecasts are then calibrated using a method that weights them based on their historical accuracy; a step called reconciliation. The residuals are used as a measure of certainty, giving higher weight to certain forecasts and lesser weight to uncertain ones.
This means that an accurate worldwide forecast will cause less precise regional forecasts to be adjusted more than the other way around. This reconciliation benefits forecasts at every level of the hierarchy in terms of accuracy.
Interested in trying out hierarchical forecasting for your organization? Book a free demo here.