Subscribe to our Blog

Don’t Forecast Demand, Model Your Demand

By Jeff Bodenstab Stefan de Kok4 Nov 2014

Demand modeling is different than demand forecasting. Simply put, it doesn’t forecast demand, it models demand.

Forecasting typically starts with a time series of data—usually presented as a bar chart displaying demand one month or week after another. Based on what’s happened in aggregate over the last months or years, it makes a projection of what will occur in future months.

For example, say you are a consumer goods company that makes orange juice products. If you want a national forecast three months out—based on the last couple of years, and factoring in trends and seasonality—you can probably forecast within a close approximation.

The problem is that most supply chain decisions are not made at that level. The real question is, how many cartons of low-pulp, 16 ounce, SKU12345 are you going to need to ship from the Newark, NJ warehouse. So even if aggregate forecasts are only off a few percentage points—the error can easily translate to 40-50% for a specific week, distribution center, and SKU.

This is because so-called “splitting” algorithms take apart the total, apportioning 8% here, 12% there, but this doesn’t match activity at the SKU level. When you aggregate you reduce the “noise,” the data gets smoother and forecasting is easier. But when you aggregate you also lose signal; signal that can never be retrieved again at the aggregate level. You trade away accuracy for ease.

Demand modeling works the opposite; from the bottom-up, as opposed to top-down. It breaks the demand components into a series of internal and external factors—the demand stream—and looks at how each impacts demand to predict future demand. It looks at the specific factors driving demand at a granular and daily level for individual SKU-Locations. It considers external demand-shaping factors— like new product introductions, trade promotions, end-of-aisle displays, price reductions—that have an impact at the most detailed level, such as at the store, and incorporates them into the forecast.

Demand modeling also processes the data differently and generates a different kind of output. Forecasting looks for a best fit from all available algorithms, generating a single value output. Demand modeling creates an adaptive demand distribution that best fits the demand profile. It then produces a range of possible outcomes with probabilities assigned to all values within the range. It goes beyond the “demand forecast number” to the probability of demand.

Modeling demand even helps with fast-moving goods, where demand appears consistent. When you model these items at the granular level, the demand looks more intermittent, irregular, and volatile. Forecasting algorithms call this unforecastable—they either can’t do it or can’t add enough value. Demand modeling breaks down this demand into its constituent parts, to understand its rationale and make a complete forecast across all your items.

As for benefits, modeling demand usually greatly improves the forecast accuracy—measurably in the aggregate and very significantly at the detailed level. It also reduces the manual intervention needed to get everything to work. When ToolsGroup replaces forecasting with demand modeling, the planner workload is often cut in half as the computer handles more of the statistical workload and reduces the exception handling. For example, Costa Coffee reduced planner workload by nearly two thirds, refocusing the team on higher value-added activities.

Click here for next week’s blog, where we dig a little deeper into this topic, focusing on the issue of separating the signal from the noise. Here is a preview:

How to Model Demand

In modeling demand, the signal is determined by “decomposing” the data into a signal and a noise portion. It makes the granularity of this baseline demand as detailed as possible. The more detail, the more signal is preserved—and the clearer the signal can be identified from within the noise. The most commonly used granularity is individual sales order-line, daily by item and by ship-to location.

From this detail, all kinds of patterns can be identified. For example, each ship-to location may show clear ordering patterns favoring certain days of the week, and may exclude other days completely, like Saturdays, Sundays, and holidays. Similarly, there may be obvious patterns for weeks within the month, driven by sales targets in fiscal calendars. This detail allows the signal to be automatically detected.

Additional information—seasonality, promotions, market intelligence—further fine-tunes the separation of signal from noise for a “quieter,” more accurate forecast (See the diagram at the top of the page). As additional information is provided, every last bit of signal is isolated from the noise.

Click on the image below for a pdf of the infographic at the top of this page.

Subscribe to our Blog
Supply Chain Brief