Jonathan Thatcher, CSCP, CAE | March/April 2014 | 24 | 2
Strategies for a more perfect forecast
Reader D.H. writes, “Over the last couple of years, our forecast performance has been stable. Unfortunately, our numbers typically are overstated 8–10 percent from a macro perspective. We dug into the issue and have determined that the primary driver behind the overstated demand lies in sales order netting logic within our planning system. Is this a common problem? What steps might help improve our overall demand signals?”
There are issues to consider with forecasting practice and material requirements planning (MRP) system configurations. First, don’t schedule production based on the greater of the forecast or sales orders. Instead, make sure orders consume the forecast as they come in. Ideally, your enterprise resources planning system should display three rows: forecast, customer orders, and requirements summary.
Imagine that the forecast equals 100 units and customer orders equal 25 units for a given period. The orders reduce the forecast, leaving a remainder of 75 forecast units. The requirement summary row needs to reflect this by displaying total orders plus the remaining forecast—in this example, 25 ordered units plus 75 forecast remainder units. For further reading, John F. Proud, CFPIM, author of Master Scheduling: A practical guide to competitive manufacturing, covers this topic well.
This approach works because production need not care whether units are ordered or forecast; they are simply “units.” This is the agreement production makes with sales. Because no forecast is perfect, a later sales and operations planning meeting should discuss the accuracy of the forecast and how well production is meeting demand while avoiding unnecessary inventory.
The next issue is the MRP demand time fence (DTF). Set the DTF to equal production lead time, and make sure your MRP system shows forecasts to zero within the DTF. For example, if it takes one week to manufacture a unit, then the DTF would be one week. Forecasts made inside the DTF should be ignored, as it is too late to produce them, and those forecasts will overstate demand.
Also, it’s crucial to code or flag extraordinary, non-repeating, or non-representative orders. These outliers should not become part of the history on which ordinary forecasting is based.
Then, set MRP forecasting history. While many systems allow for forecast history to be based on either booked (expected) orders or actual (shipped) orders, most people choose the former in order to guard against situations where orders exceed inventory and customers are lost. For instance, suppose you have 50 units available, and an order of 100 units arrives. You want the history to show the booked order, or 100 units. If it shows the actual order, you will see only 50 actual shipped orders. MRP history will not recognize the other 50 orders, which the customer never received. And if that customer decides to look to a competitor, your history won’t reveal the problem; demand will just mysteriously fall. But the reason is clear: You did not satisfy the booked-order customer demand.
Finally, make sure the weekly forecasting period matches the periods used by sales, bearing in mind that sales rarely occur evenly week over week. Aggregate numbers, such as monthly or quarterly results, are your friends. Data based on long histories is easier to forecast than daily or weekly data—and has far less variability. (For more on this topic, see the September 2005 APICS magazine article “Exploring the Mystery of the Sales Forecast.”)
By definition, no forecast is completely correct. But we can get a little closer to perfection with a stronger forecasting practice.
Jonathan Thatcher, CSCP, CAE, is director of research for the APICS professional development division. He may be contacted at firstname.lastname@example.org.
Send APICS your supply chain or operations management questions at email@example.com.