Matching Demand Forecast Periods to Production Lead Times

It is common to speak of a “one number” demand forecast. And a “one number” forecast works fine if your demand forecast and your manufacturing planning are based on the same time period. For example, if you do a monthly forecast and the plant uses it to plan production for the entire month, then there is no inconsistency.

But a lot can happen in a month.

I was briefed a while back by Nitin Goyal, product manager for JDA Software’s Sales & Operational Planning (S&OP) solution. I asked him how his customers with long manufacturing lead times were dealing with this issue. He told me that some of JDA’s clients in the semiconductor industry are using banded forecasts that include a high-confidence forecast (for products the manufacturer is very certain the market will need); a medium-confidence forecast (often the “one number” consensus forecast); and a low-confidence forecast (a high demand scenario).

In this case, a manufacturer can safely make the high confidence forecast units early in the month, then produce the medium-confidence forecast units later in the month, unless they see a drastic drop-off in orders. Only if they see an unexpected surge of demand does the manufacturer ramp up the factory in the last week of the month. In the semiconductor industry, adding capacity is expensive. These banded probabilistic (“stochastic”) forecasts make a lot of sense and can improve the S&OP process.

Many companies have found that if they move from monthly to weekly forecasts their forecasts become more accurate. What if you take it a step further and produced forecasts every day? That is what Land O’Lakes did. The company moved to doing a daily rolling forecast that covers the next 28 days.

Jeanine Viani, the Director of Supply Chain Planning for Dairy Foods at Land O’Lakes, spoke at Oracle OpenWorld last month about the company’s implementation of Oracle’s Demantra Advanced Forecasting and Demand Modeling application, an optional add-on to Oracle’s Demantra Demand Management solution. This application makes use of an individual customer’s order history and compares it to the orders that have dropped in the last few days. Land O’Lakes is the first Oracle customer to use Demantra for open order correction.

You might be wondering, why forecast daily unless the lead time for your factories is also just a day?

At Land O’Lakes, the production lead time to make butter is one week and the lead time for cheeses 3 to 4 weeks. The company’s traditional weekly forecast was 45 percent accurate, captured at the item/location level. But when they do daily forecasts, the following week’s accuracy is better than 70 percent. The forecast for two weeks out drops to about 55 percent, and the accuracy for weeks three and four remains at 45 percent. So, in effect, for the one-week lead time products, accuracy is now at 70 percent rather than 45 percent. For cheeses with a lead time of three weeks, the accuracy is now 56 percent (70 percent week one, plus 55 percent week two, plus 45 percent week three, divided by three weeks) rather than 45 percent.

So, why not forecast weekly? Because the plant can base its production schedule on the most current and accurate forecast. Demand is not jumping around so dramatically – partly because many customers want their shipments on the same day every week – that they are having a hard time balancing demand and supply.

Land O’Lakes is on track to receive the payback several months ahead of the original estimate, based on reductions in inventory, coupled with a reduced need to redeploy inventory from one distribution center to another.

Finally, some companies are using Demand Signal Repositories to take downstream data and create accurate, short time forecasts and in this way become “demand driven” (see, for example, “Del Monte Foods and Shelf-Level Collaboration”). But accomplishing what Del Monte did is difficult. For many companies, starting with stochastic forecasts, or moving to weekly or even daily forecasting, will be a better intermediate step.