Since 2001, we have been studying the release of current and historical products. We have tracked the evolution of forecasts, factory production, and inventory for many major product releases and studied how the signals flow through teams across Intel's organization. Our methods have included both quantitative analysis of our data sets and interviews with personnel in groups that work with these data sets to understand policies, strategies, and perspectives on the product transitions.

We learned that calling demand correctly for new products–and the products the new products replace–is a formidable task. Four fundamental sources of noise cause difficulty in determining true market demand: current data, such as orders and inventory; market assessment, such as intelligence and consensus on how appealing products and promotions (and competing products) might be to the market; market objectives, the goals Intel has for its products, such as unit sales, average price, market segment share, and technology leadership; and, strategic plans, such as the decisions about which products and stock keeping units (SKUs) to sell, how to price them, and how to take advantage of technological and manufacturing capabilities. Nearly all the pitfalls we have discovered in forecasting demand can be linked–with the benefit of hindsight–to one or to a combination of these factors. The question, of course, is how to account for these factors in advance to systematically and repeatedly do the best possible job of forecasting and planning.

The fundamental problem in managing forecasts is twofold. First, the hard data being created, judged, and passed from group to group lack credibility based on past performance, so each group feels the need to adjust the information based on any number of experiences and heuristics. Groups preparing to publish data are aware of how other groups will likely judge the data and are therefore prone to gaming the system, i.e., adjusting numbers in anticipation of future judgment.

Second, data sets themselves do not really convey any specific meaning. Meaning can be inferred from how the data compare to expectations or previously published data, but numbers in enterprise applications or spreadsheets cannot explain the strategies Intel and its customers are employing or the uncertainties they are facing. Decentralized organizations must find a means of transmitting business context; in other words, instead of transmitting mere data sets, they must transmit information and intelligence from employees who have it to employees who need it to make decisions and plans. We learned that Intel has many informal networks that attempt to move that knowledge across the organization, but these networks have many failure modes: turnover of employees in key positions, limited bandwidth of each individual and team, and difficulty systematically discovering the important information to be learned (stated differently, whom to include in the network).

Our research has led to three methods that are being used at Intel today. One focuses on market assessment and uses data from across the organization to score factors affecting ramp rates (Product Transition Index). The second ties market objectives, strategic plans, and market assessment, identifying risks and developing contingency strategies to improve coordination and cooperation (Transition Playbook) [4]. And, the third (IAM) paradoxically uses the most structured of the three methods to promote transmission of the most unstructured information, i.e., any and all information participants feel is relevant to developing a forecast.

The source of demand uncertainty for Intel begins with biased signals from the interaction with our customers. Customers typically signal strong demand for popular upcoming products. In fact, if Intel fulfilled all bookings (advance orders) for all customers, the result would most often be substantial oversupply. Customers want to assure supply and be certain that a competitor does not procure an unfair share, so the condition of "phantom demand" develops. Orders are inflated to keep the playing field level across customers and so that each can lock in as much supply as possible in the event of a shortage. Conversely, orders for new products are sometimes deflated, signaling that customers do not want to go to the new product too quickly. Perhaps they prefer the prior product for any number of reasons, or they believe low demand might lead to price reduction, or in some cases, new technologies and supporting components are relatively scarce and will increase in supply (decrease in price) over time. Whatever the exact cause, a study of orders and forecasts developed by Intel's geographical sales organizations shows that the volatility of these signals is large, and it is not unusual for the forecasts to be over 20% high or low.

Once geographical ("geo") forecasts are published, a central business planning group is responsible for publishing official demand forecasts that guide the supply network. The geo forecasts are one input considered by this team, but many other factors including models of worldwide sales growth, Intel's share of the market segment, product mix by any number of attributes, sales versus price point, historical product ramp rates, and various inventory data (for instance, work in process, finished goods, customer stocks) are used to produce official forecasts. While historical results show that the central business planning team does reduce the volatility of the geo data and often achieves improved accuracy, their track record shows that overcalling or undercalling sales, especially during product ramps, is not as rare as we might hope. These missed calls can lead to significant surpluses or shortages that take money right off the bottom line. Intel's factories, keen not to get caught in these situations, do not always build to the official forecasts. They also use models to help maintain proper inventories, smooth production, and achieve high operating efficiency, but our research has found no evidence to date that this final judgment improves demand fulfillment systematically or repeatedly.

The challenge of demand forecasting is real and costly. Demand risk is among the greatest threats facing Intel and other manufacturing firms day to day. To demonstrate how formidable demand risk can be, the following are actual situations we have discovered:

• Various groups across Intel estimated sales of a new product over an initial period after launch to be anywhere from one million to four million units.
• Two similar products (common architecture) were released within a quarter of one another in different (essentially non-competing) market segments. One resulted in a shortage, the other in a surplus.
• Geo forecasts for one new product were as low as 13.5 million units for a fixed period, while official forecasts were guiding the factories to build 26.5 million and sales targets were 27 to 28 million.
• Two products were projected to sell over 10 million and below 7.5 million units during a future period. In three months the forecasts flipped to under 5 million and over 11 million, respectively.
• Sustained growth in the mobile PC segment beginning in 2004 caught the whole industry by surprise. Four quarters of year-over-year growth, roughly double what was expected, made for a tough supply picture.

Certainly not all misses are this extreme. Intel's forecasting teams routinely perform quite well given the challenge of their task, often achieving forecasts with less than +/- 5% error. However, sustained high performance does not make up for each isolated miss that costs the company millions to, conceivably, hundreds of millions of dollars. Everyone involved in forecasting at Intel continually strives to achieve better performance across the board, and we are always exploring new approaches that might bring improvement.