It’s been more than two years since Nike Chairman Phil Knight owned up to the sneaker giant’s disastrous $400 million experiment with demand forecasting software. The headlines are well known: Nike went live with its much-vaunted i2 system in June 2000, and nine months later, its executives acknowledged that they would be taking a major inventory write-off because the forecasts from the automated system had been so inaccurate. With that announcement in February 2001, Nike’s stock value plummeted, along with its reputation as an innovative user of technology.
But what has since trickled out in court documents from shareholder lawsuits may be even more disturbing because it shines a harsh light on the inherent limitations of demand forecasting software. According to the documents, i2’s supposedly state-of-the-art forecasting system couldn’t communicate with Nike’s existing systems, which impaired its ability to analyze large amounts of product information. At some point, the data even had to be entered in by hand, greatly increasing the chance for mistakes. And the forecasts themselves were way off. Relying exclusively on the automated projections, Nike ended up ordering $90 million worth of shoes, such as the Air Garnett II, that turned out to be very poor sellers. The company also came up with an $80 million to $100 million shortfall on popular models, such as the Air Force One.
Nike isn’t the only company with a forecasting horror story. Corporate America is littered with companies that invested heavily in demand software but have little or nothing to show for it. Goodyear, for example, implemented a demand forecasting system in mid-2000 but hasn’t shown significant improvement in managing its inventory, and last year the tire company lost more money than the year before.
Yet vendors and academics are still pushing forecasting software. In 2002 alone, companies spent $19 billion on demand forecasting software and other supply chain solutions, according to IDC (a sister company to CIO’s publisher). And in a speech in February, Stanford University supply chain guru Hau Lee extolled the virtues of harnessing software to extract customer knowledge in order to forecast demand.
Many CIOs, however, remain skeptical. Privately, members of Lee’s audience complained to a reporter present that the ability to accurately forecast could hardly be taken for granted. And according to a recent Booz, Allen & Hamilton survey of 196 senior executives, 45 percent said that supply chain technology in general had failed to meet their expectations. More than half?56 percent?blamed the shortcoming squarely on demand forecasting software. From hard experience, a growing number of CIOs now realize that computer systems alone are incapable of producing accurate forecasts.
There are a number of reasons why. To begin with, forecasting systems are only as good as the data put in them and, due to the complexity of modern supply chains?where a company wants to collect information about multiple products from multiple customers and suppliers?more often than not the data isn’t accurate enough. Furthermore, software can’t predict the future, particularly sudden, unexpected shifts in economic or market conditions. Nor can it exercise the kind of rational analysis or judgement that human beings excel at. Hence, demand forecasting technology is inherently limited, and companies such as Nike and Cisco that rely on it without an institutionalized set of human checks and balances will invariably end up in trouble.
“Demand forecasting sounds scientific,” says Sumantra Sengupta, CIO for the Scotts Co., the world’s largest supplier of consumer lawn and garden care products. “But I would say that if you looked at the split between people, science and process, people are half the equation. Algorithms are algorithms. That is not what will win the game for me.”
Good demand forecasting requires a combination of accurate data and smart people. Up-to-date sales data and point-of-sale (POS) information will almost always improve a forecast. So will having the pro-cesses and people in place to make sense of anomalous results or simply to check computer-generated predictions against the pulse on the street.
“Anyone who thinks you can do it with just mathematics and statistics is only partly right,” says Doug Richardson, CIO of electronic products maker Vicor. “Human intelligence is also required.”
The Whiplash Effect
Before vendors began selling demand planning software, forecasting was essentially a balancing act between competing factions within the enterprise. The marketing department would set a high target because it wanted the product to be a success, says Tom Burns, CFO for the enterprise network division of telecommunications equipment maker Alcatel. Salespeople, on the other hand, would come in with conservative forecasts since they wanted to keep their sales quotas low and manageable. “If marketing says we are going to sell $150 million, and the salesguy says we are going to sell $75 million, what do we tell the supply chain guys to build?” Burns asks.
One can see the appeal of a computerized system that could provide an objective answer to that question. Furthermore, the math needed to build these systems has been around for nearly 75 years. It was Ronald Fisher, a British mathematician working after World War I, who first conceived of a system that could take numbers, look for patterns and then make predictions based on those patterns. The result, the classic regression model, is still used in 90 percent of demand planning software today.
Regression essentially takes multiple variables, makes inferences about the relationships between them, and ultimately charts the result as a curve showing upward or downward trends. The curve can be extended to predict future results. For example, a regression study of the rate of death among people between the ages of 20 and 80 would, despite numerous exceptions, find a general trend that as people got older, the rate of death increased. You could then predict that an 81-, 85- or 90-year-old person would be even more likely to die than someone who is 80.
The problem is that regression analysis?and any other statistical model a demand forecasting system may use?requires clean data and a potential relationship among the variables, says Rob Cooley, U.S. technical director at KXEN, a demand forecasting vendor. In Fisher’s day there wasn’t the computational power to consider more than a few variables, which made it easier to focus on the accuracy of a few data points, like in the rate of death example. But today, computerized systems make it possible to consider hundreds, if not thousands, of variables?anything from weather to time of day?and a correspondingly vast number of data points.
Most of these data points are inaccurate, or more specifically, are only an estimate of what actually happened. The most common example is guessing what consumers bought based on what the company itself sold. While a retail store knows how much of a product it sells, the manufacturer only knows how much the retailer orders?and more often than not there are distributors acting as middlemen to further muddy the transfer of information about sales.
Logistics executives at Procter & Gamble studied how this dynamic affected demand planning and found that the further away your data is from the point of sale, the more data accuracy decreased and forecasting errors increased. For example, P&G found that consumers bought its Pampers diapers at a fairly steady clip and that retailers’ orders reflected this: Orders had moderate swings one would associate with relatively flat demand. The distributors, however, would react to moderate increases by not only increasing their orders, but by upwardly adjusting their reserve stock, signaling a much larger increase in demand back to P&G. The manufacturer, in turn, would ramp up its Pampers production and continue the bullwhip effect down the supply chain. Ultimately, everyone would be left holding excess inventory.
Garbage In, Garbage Out
The best way to avoid the trap of overforecasting demand is to use point-of-sale information directly from the retailer. Since POS data is an accurate gauge of consumption, it improves the reliability of a forecast. That is how Scotts CIO Sengupta was able to improve his company’s forecasting results. By using point-of-sale data, Scotts increased its forecast accuracy by more than 30 percent in one year.
Before it started using POS data, Scotts would forecast the demand for its products at the national level?but not at the individual store level?for each of its customers, such as Wal-Mart and Home Depot. Each forecast would take into account how much each customer ordered of, say, a particular type of fertilizer in the past and combine that with other factors such as expected weather patterns. Since orders were simply an estimate of retail sales, the process left Scotts susceptible to the bullwhip effect. Furthermore, the sheer volume of the orders greatly inflated the forecasts’ margin of error.
Now that Scotts has point-of-sale information from each retail outlet, its forecasts are more accurate and the risk of bullwhips is all but eliminated. Furthermore, the POS data lets Sengupta produce smaller, more detailed forecasts for each individual retail outlet if he wants (Sengupta says that Scotts actually forecasts in groups of stores to help reduce the impact of a one-time event in one store that wouldn’t be replicated in another). Having many smaller, more accurate forecasts further reduces the overall margin of error.
In improving its forecasting process, Scotts has an advantage: The retail and consumer packaged goods industry is well ahead of the game when it comes to sharing data such as point-of-sale information. Most companies in this industry follow the blueprint laid out by the Voluntary Interindustry Commerce Standards Association subcommittee on collaborative planning, forecasting and replenishment.
In the rest of the world, however, most companies aren’t in a position to get POS information from their customers. In the first place, few companies collect product-level data at the point of sale. And second, many aren’t willing to share data that has traditionally been viewed as a closely guarded competitive secret.
But that doesn’t mean you can’t get better data and use it to improve your forecasts. In Europe, for example, Scotts gets POS information from only its three biggest customers or 20 percent of its business there?the others either aren’t able or willing to share it. “In Europe we understand that we can’t get point of sale, but we still try to get as close to the point of final consumption as we can,” Sengupta says. In this case, that’s when sale items leave the distribution centers. While it isn’t the same as point of sale, Scotts at least knows where its products are going and how much has actually been sold to retail outlets, which is more accurate than traditional order information. Furthermore, the distribution center is several days closer to the eventual point of sale than Scotts’ own warehouse, making that data a better indicator of current market trends.
There is other data available that can help CIOs improve the data in their forecasts. Imperial Sugar Vice President and CIO George Muller says that his company combines the order information it receives from its customers with market intelligence reports from Information Resources Inc. data about what actually gets sold in stores. The IRI sales data, which will show, for example, how much sugar was sold in Atlanta area supermarkets, can serve as a surrogate for point-of-sale data. At the very least, it gives Imperial Sugar something to check its order data against.
Taking the Market’s Pulse
Even if a demand forecasting system had 100 percent accurate information, there is another problem: The past can’t predict the future. Computer-generated forecasts use historical data to make assumptions about what will happen, but there is no way for them to anticipate major market changes. For example, Belvedere International, which is based in Ontario, Canada, makes skin-care products. When SARS broke out in Toronto, Belvedere sold more than a year’s worth of its One Step hand disinfectant in a month. No forecasting system could have predicted that. Belvedere has kept its assembly line running 16 hours a day, six days a week?modifying production of other goods in the process?just to keep pace with demand. “It’s no different from forecasting the weather,” says Gene Alvarez, Meta Group’s vice president of technology re-search services. “Once in a while something the model couldn’t figure out catches them off guard. Same thing happens with consumer taste and demand.”
Even forecasts that are made with a limited number of variables and with accurate data will be off. They still make the fundamental assumption that what was true yesterday will be true tomorrow. But because the data about a change lags behind the change itself, it takes human market watchers to note business climate alterations.
Vicor, which manufactures power converters for electronic circuit boards, found this out at the beginning of the recent economic downturn. Until a year ago, the company had used a homegrown forecasting system that it had built in 1993. CIO Richardson describes it as a straight-line forecast based on sales history. Company executives relied solely on the automated forecasts to predict demand for their products.
“It was reasonably good in the ’90s when demand was increasing at a nice steady rate,” he says. “Where it broke down was when the product mix increased and the business downturn started.”
Vicor didn’t see it coming. In a conference call in April 2001, CEO Patrizio Vinciarelli said that shipments “fell off the table,” and the company was left with a massive inventory glut. “When the future doesn’t resemble the past, none of this forecasting software works well,” Richardson says.
The mishap taught Vicor the necessity of factoring human intelligence into its forecasts. In order to make sure that it isn’t caught off guard again, the company set up a dual forecasting process in which the sales department comes up with a forecast and the computer system, which was upgraded a year ago, makes another. The two are complementary; the sales department is too conservative with its forecasts (Richardson thinks the salespeople are merely cautious; a cynic might point out that they are compensated for selling above quota). On the other hand, the computer system won’t necessarily pick up on changes in the market that salespeople can often see.
For example, month after month, one telecom customer of Vicor kept placing the same order, and month after month the computer spit out a flat forecast. But a sales manager in the field found out that the telecom increased its order with another supplier whose parts are used in the same product as Vicor’s. The sales manager talked to the telecom company and found out that the company had indeed decided to ramp up production, and armed with that information, Vicor increased its production as well, and was thus prepared when the telecom placed a bigger order.
While Vicor uses computer-generated demand forecasts as a check and balance for the human-generated forecasts, Scotts takes a slightly different approach. It takes its computer-generated forecast and distributes it to designated forecast planners for feedback. The planners, who are experts for the store and area they represent, make changes based on their expertise. For example, a planner in the Northeast might lower a forecast due to bad weather that limits gardening, or another might increase a forecast if he knows that a particular store is planning a promotion. Scotts takes one unusual step to ensure that the planners have access to the most up-to-date information: The planners actually work in the office of the customer company’s buyer. If, for instance, the planner represents Home Depot in a particular region, he works in the office of Home Depot’s buyer for that region. Sengupta says that the close proximity fosters collaboration between the planners and Scotts’ customers.
In the end, the demand forecasting failure at Nike and other companies can be laid squarely on the shoulders of executives who put too much faith in technology. Court records in the lawsuits by shareholders against Nike reveal that executives for the sneaker company didn’t even hold meetings to review and discuss the computerized forecasts that turned out to be so disastrously wrong. In other words, Nike management neglected to put in place a high-level process of human checks and balances for the computerized forecast. While that negligence actually enabled Nike executives to successfully argue that they were initially unaware of the flawed forecast that was generating such a huge inventory glut, it was a Pyrrhic victory. The company still lost $180 million in sneaker sales and a third of its stock market value.
The Nike case powerfully illustrates that forecasting, no matter how advanced vendors claim their technology is, has to be an executive-level process. Executives need to review the computerized forecast and analyze how it squares with information from their sales and marketing reps, and then sign off on a number that the whole company can live with. At Alcatel, executives meet on a regular basis to dissect and discuss forecasts, which are produced by combining computer readouts with human intelligence.
“The final meeting here is attended by me, the head of supply chain and the heads of marketing and sales,” says Alcatel CFO Burns. “We all have to approve the decision. We live and die by it.”