IT and business leaders involved in demand forecasting and planning can learn best practices from Cisco's experiences in using analytics and a consensus process to improve its forecasts. An increase in forecast efficiency and accuracy can give enterprises more profits by making business operations more efficient.

• Cisco has highly variable demand and a complex mix of short (for example, Internet Protocol [IP] telephony) and long (for example, data center infrastructure) product life cycles. Therefore, Cisco's management could not rely only on the forecasting models in traditional demand-planning applications.
• Cisco employs a variety of forecasting models to improve forecast accuracy, including some that are Cisco-specific.
• At the end of the cycle, the demand-planning team compares forecasts and the output of the entire process with actual results, feeds that information back into the process, and uses it to assess everyone's performance.
• The consensus forecasting process, coupled with advanced analytics, improved Cisco's forecast accuracy, reduced forecast bias, improved supply/demand balancing, increased inventory turns, reduced excess and obsolete stock, and enabled Cisco to plan demand flexibly — in response to market run-ups as well as downturns.
• From 2001 to 2008, Cisco's forecast accuracy increased by 20% to 30%, and forecast bias fell by 60% to 80%.

• Enlist an executive who believes in analytics to champion your project, remove roadblocks and secure resources.
• Develop your analytical forecasting system and models through an iterative process, giving stakeholders visibility into the process and supporting system refinement.
• Build a tight relationship between the analytics team and the IT organization. A strong partnership ensures that the analytics platform integrates with workflow and other tools and fits enterprise architecture.
• Involve all key stakeholders in the development process. Have them collaborate — not just on creating the forecasting process, but also on executing it.

Cisco used analytics, domain expertise and a consensus process that increased the accuracy of its demand forecasts, enabled it to reduce the slack in its supply chain, and thereby improved management of its inventory-carrying costs. Cisco built a team with analytic expertise to create a statistical method to eliminate biases. The team implemented an analytic application that automated the generation of initial demand forecasts. The team also used some of the statistical models that came with the application, augmented some out-of-the-box models with data from multiple sources, and created new models itself. The initial forecasts fed a process that brought together people with domain expertise from marketing, finance, sales and other departments to develop final consensus forecasts. This process allowed for exception-handling. Cisco established performance measures, especially value-added metrics, that feed continuous improvement.

Errors in forecasting demand can send the managers of a supply chain scrambling to raise or cut production or change the mix of products to meet actual demand. These unplanned changes can cost a lot of money and could drive customers or partners away. Therefore, enterprises can gain a competitive advantage by improving forecasting and planning. Cisco, a network equipment vendor with about 300 major product families, carefully balanced domain expertise against statistical predictions to forecast future customer demand. Business and IT leaders involved with complex processes can improve their demand forecasts by following Cisco's best practices.

Cisco outsources 95% of its manufacturing and has over 1,000 suppliers. Historically, Cisco based its forecasting on sales and marketing projections. However, as Cisco's business expanded and the complexity and mix of products increased, this approach to demand forecasting did not suffice for supply chain execution. An increasing number of products and markets meant greater variability in customer demand. In addition, improved efforts in implementing lean processes across the supply chain reduced inventory buffers, typically used to hedge against forecast errors. Lean processes also meant that Cisco's contract manufacturers increasingly relied on Cisco's ability to communicate demand effectively and accurately.

In 2005, a newly appointed vice president of demand management and planning recognized the need for a more robust demand-forecasting and planning process. He challenged the organization to develop an integrated forecasting process that was efficient and analytically rigorous, yet balanced with domain expertise. It needed to be robust enough to scale to support Cisco's expanding breadth of products and cover a 15-month horizon, with reasonable consistency across all business units and product lines. Where appropriate, he preferred alignment to industry standards and best practices.

In an environment with highly variable demand and short product life cycles — the characteristics of Cisco's market — management knew it could not rely only on the forecasting models built into traditional demand-planning applications. Cisco concluded that those types of forecasting systems better suited more predictable demand patterns and longer product life cycles. Cisco turned toward a more customizable solution for forecasting that uses advanced analytics and automation to overcome the increasing complexity of demand.

In early 2006, Cisco's management created a statistical forecasting team. Over the subsequent 12 months, the team grew to six forecast practitioners, most with advanced degrees whose academic backgrounds included operations research, engineering and decision science. Team members also needed strong communication skills.

Through a series of prototypes, the team developed a statistical forecasting capability that would support the requirements of the vice president (see Note 1). The system went live in September 2007. The new statistical forecasting process (see Figure 1) uses a variety of objective inputs, including:

• Weekly demand data
• Cost
• Product hierarchy
• Sales region
• Market segment
• Product life cycle attributes
• Product class
• Product road maps

The inputs feed directly into an automated forecasting process. Some of the statistical models Cisco employed came from within the packaged analytics application it implemented. The application allowed the team to code processes to transform some data using various statistical models and to augment data with information such as leading indicators. In addition, the team developed Cisco-specific statistical models. The automated process uses all these models and data enhancements because different models work best with different kinds of data and in different situations. Data and models feed into the overall analytical forecasting engine to produce an automated forecast. Cisco can tune a model for a particular time frame or product. For a given month, the automated process takes the median of all forecasts appropriate for a particular data stream.

The analytics team then considers whether the automated forecast needs manual exception-handling, based on sample accuracy, the previous month's forecast accuracy and the inclusion of new products in the forecast. The team compares the manual and automated forecasts to create a final forecast, which feeds into a demand-planning application used by a consensus demand-planning process (see below).

To automate the statistical forecasting process, the team implemented an analytics application that was:

• Flexible enough to allow customization and configuration of the forecasting models.
• Scalable, to accommodate a large volume of data.
• Robust enough to operate in an enterprise environment.

Mathematically-derived forecasts help improve forecast accuracy. However, the best forecasting processes include all available data and domain expertise, and engage stakeholders to generate, review and sign off the final forecast. Therefore, Cisco implemented a consensus demand-planning process (see Note 2).

The process occurs on a monthly cadence. In addition to the forecast practitioners, several other critical players participate. Management reshaped traditional planning roles to facilitate this process, and representatives from sales, marketing and finance participate in the consensus process. The demand planner, a role in the supply chain organization, facilitates the meeting and is ultimately responsible for sending the demand plan up the supply chain to the contract manufacturers.

At the end of the cycle, the team compares forecasts and the output of the entire demand-planning process with actual results, feeds that information back into the process, and uses it to assess everyone's performance (see details below).

Like the statistical forecasting process, the consensus demand-planning process also needed automation to operate at the speed and efficiency that the business required. Cisco used process-oriented analytic tools, a business intelligence (BI) platform, a defined workflow, a data warehouse and demand-planning technology (see Figure 2). Inputs for the automated forecast are collected in the enterprise's data warehouse and operational data store, which had been created previously to feed BI applications. The team publishes its forecasts in the data warehouse, and people can access them via BI and demand-planning applications.

Cisco recognized the importance of accountability within the process. Cisco measures three major aspects of its demand forecasts: accuracy, bias and value added (see Table 1, and below for more on value-added metrics).

Cisco did a soft launch of its new forecasting process over six months, beginning in September 2007. By March 2008, the process was well advanced, with several business units fully engaged. Now, with processes wholly in operation, the new forecasts provide comprehensive week-by-week demand levels for all of Cisco's products across a 15-month horizon and guide the allocation of resources within Cisco.

From 2001 to 2008, Cisco's forecast accuracy increased by 20% to 30%, and forecast bias fell by 60% to 80% (based on a quarterly summary). Forecasts are created and measured at product-unit level so that Cisco can propagate a truer reflection of demand into its supply chain. The process reconciles various perspectives to provide a deeper understanding of demand and greater confidence in the end forecasts throughout the supply chain. The reliance on objective statistics in conjunction with domain expertise has also reduced the forecast bias. The BI and analytics applications give users greater visibility, allowing a more detailed look at the factors contributing to demand.

In addition, the consensus demand-planning process improved overall balancing of cost and resiliency across Cisco's supply chain, thereby improving supply/demand balancing, increasing inventory turns and enabling Cisco to plan demand flexibly in response to market run-ups as well as downturns.

The project has won more respect for analytics at Cisco. The company has expanded its interest in analytics so that the team can explore opportunities for using these techniques throughout the supply chain organization. The team's responsibilities have extended from statistical forecasting to include supply chain analytics as well as forecasting responsibilities in other areas, including distribution forecasting and call center forecasting. Cisco will encourage collaboration between stakeholders of the forecasting process to help continue its success.

Executive champion: The vice president of demand management and planning believed strongly in analytics. He became the project's champion, removed roadblocks and secured resources throughout execution of the initiative.

Iterative development process: The analytical forecasting system was developed through an iterative process, giving stakeholders visibility into the process, supporting system refinement and creating a sense of accomplishment within the analytics team.

Tight relationship with the IT organization: Analytical tools often conflict with enterprise IT strategies. A strong partnership with the IT organization ensured that the analytics platform would integrate with the workflow and BI tools and be accepted into the enterprise architecture.

Holistic approach: The project involved all key stakeholders in the development process. It brought everyone together so that they could collaborate — not just on creating the forecasting process, but also on executing it. This approach created a sense of ownership in the process.

Accountability: The project included accountability in the process itself, as well as for the people executing it. Forecasts are checked against actual results, and the team uses these comparisons to improve its models. The team established simple metrics to judge performance, the most important of which is value added (see Table 2). Cisco measures the value of each forecast by comparing it to a "naïve" forecast done in the easiest way, such as by taking a six-month moving average. Thus, an analyst with a higher forecast accuracy (Analyst A) may actually provide less value than someone with a lower forecast accuracy (Analyst C). Value-added metrics show the team where it should put resources and where a naïve forecast suffices. Value-added metrics appear on the dashboards of Cisco's senior management.

The forecasting initiative needed the right mix of talent, tools, processes and support. A consensus forecasting process needed to include all relevant sources of information and insight — not just best approximations of "the experts." For example, a blend of quantitative and qualitative expertise orchestrated as an analytical workflow within a business process produces the best results. Multiple inputs and participants also enable the process to adapt more quickly to changing business requirements or to the development of new models.

Cisco deliberately changed its culture of forecasting and demand planning. People, workflows and decision-making processes needed to incorporate the new analytic approach. For example, the analytics team had to overcome initial suspicion about the statistical models and the analysis performed — would these algorithms reflect reality? Analytical results are often counterintuitive, and the analytics team found it needed to balance clarity and detail in its communications. The team needed to explain how the analytic process works in clear terms so that business people could follow it, while providing enough detail to convince them to trust the process.

Cisco got its priorities and processes right, and technology was the least important component. In other words, the project might have failed had Cisco focused primarily on implementing analytic applications. Instead, the organization focused more on getting the right mix of people involved, creating a flexible, comprehensive process and defining the right metrics. Today, the analytics team can market its success and demonstrate its competency to bring analytic insights, skills and techniques to managing other business processes within Cisco.