Demands of Forecasting

Knowing how much product to order is key to Avon.

Even so, Avon historically has relied largely on manual processes to collect and evaluate historical sales information.

Avon used this past data to create educated “guesstimates” of how much product the company would need on hand in both the short term (two to three weeks out), as well as in the longer term (up to six months in the future).

One key company goal as a result is to improve the accuracy of its forecasting by 30%—which can produce a huge impact on the bottom line of a $5.7 billion company.

Four years ago, the company began to ask supply-chain-management vendor Manugistics Group to help improve its acccuracy.

But this move fell short of expectations, former Avon executives say. For instance, they tell of a huge, missed opportunity two Christmases ago when Manugistics’ did not predict a vital new trend.

The result? Too few Pokémon watches for too many orders from sales reps in the field. What the experience said was Avon’s use of Manugistics in the U.S. was not ready for prime time.

Avon, though, is undertaking a second round of improving its key business processes and, under vice president of information technology Paul Krautheim, the company is still pointing to Manugistics as the secret weapon that will help the company get forecasting on target.

At the core of Avon’s attempts to forecast demand for lipstick, fragrances and even sundresses is a module in Manugistics’ Networks suite. The Demand module helps users cut buffer inventories and cycle times using product mix, promotion, and price analyses.

Here’s how it works: Manugistics collects data from Avon’s homegrown forecasting systems, as well as new orders from its sales representatives.

The data from the Avon ladies is collected in a variety of ways—paper forms, fax transmissions and phone calls—all of which must be fed into the program that forecasts demand for Avon products.

Soon, Manugistics’ applications will even take in data from online orders placed by Avon’s “e-reps,” as well as orders placed by consumers directly with Avon at, says Tim Murnin, Manugistics’ manager of consulting services.

But Avon isn’t pinning all its hopes on Manugistics any longer. In January, 2001, Avon decided to supplement Manugistics’ module with an application from Churchill Software, a Troy, Mich., company whose software is used by retailers such as Talbot’s, Sears, Wal-Mart and Kmart.

Supply chain automation software, such as that provided by Manugistics, SAP or E3 Corp., generally provides forecasts of future trends that are largely based on past trends.

By contrast, Churchill’s software accommodates data that doesn’t fit into neat categories.

Churchill’s Promotional Demand Forecasting software takes into account so-called “causal factors”—data such as how much one sale item will likely cannibalize another—that produce unusual changes.

Churchill’s software can handle as many as 35 such factors; in Avon’s case, it is computing 19 factors as part of the Avon Mexico demand-forecasting pilot that starts this Fall.

Both Churchill’s and Manugistics’ software help Avon determine “promotional lift,” that is, the likely promotional “bump” that will result from a variety of causal factors combined. If Avon were to place sale-priced Naturals Shower Gel next to a brand-new item—like Milk Made Wholesome Shower Gel—in one of its rep catalogs, the combination of Churchill’s and Manugistics’ software would help Avon predict how much cannibalization of the older item that would cause.

Some companies use multi-variate regression, a technique that involves coding custom algorithms to predict the impact of causal factors.

But Churchill, which got its start in the late 1980s as part of one of IBM’s labs experimenting with artificial intelligence techniques, is using neural network technology to do the same. Neural networks search for patterns, and use these patterns to predict trends.

“Neurals work better with non-complete data,” attests Churchill President and Avon liaison Harve Light, Sr. “When neural (networks) encounter missing data, they just mitigate it and give a slightly conservative answer.”

Light says Manugistics’ demand-forecasting system generally requires a consistent series of historical data in order to carry out its predictions.

But because Avon often removes items from its representatives’ catalogs for one of its regular two-week campaign cycles, only to return that same item to the catalog a month or more later, the data doesn’t lend itself to being captured in a historical sequence.

The complexities created when causal factors are in play simultaneously also can befuddle traditional forecasting systems, Light claims.

For example, “You can get a higher (causal) lift on a toy that’s discounted 10% before Christmas than you can get on the same toy that’s discounted 20% after Christmas,” according to Light.

If Avon were using Manugistics alone, this type of rule would have to be programmed manually; now, Churchill will override this type of input for Avon, he says, and provide a prediction based on recent selling patterns.