The goal of an organization should be to build closer customer relationships to better understand demand patterns rather than rely only on forecasting models. Closer customer relationships reduce the need for developing models. In the last part of this chapter, non-mathematical tools and methods will be discussed relative to demand management. But if forecasting needs to be done, then the accuracy of the forecast should be measured and improved by investigating the root causes for poor demand management practices.

To predict future demand, organizations use a combination of methods, both qualitative and quantitative. Qualitative methods include gathering opinions by bringing together people who are knowledgeable of customer demand (i.e., a jury of executive opinion) and asking the sales teams using interviews or surveys (i.e., build a sales force composite forecast by market segment or vertical and other stratification factors). Quantitative methods include time series analysis, regression-based models, and specialty models. In time series analysis, a forecasting model is built with lagging values of the dependent variable to predict its future demand. An example would be predicting sales next month based on sales from previous months. In contrast, regression models forecast future sales using lagging, leading, and coincident values of a dependent variable, as well as one or more independent variables. Leading variables enable a model to predict future demand based on variables known to predict future demand. An example would be real disposable income in this month as a leading indicator for retail sales next month.

Forecasts are used by different teams across an organization to plan their work. As an example, marketing uses forecasts to plan advertising and other promotional activities for direct and indirect sales. These forecasts are made at a product group level monthly and into a future time horizon. Sales teams use forecasts to measure performance to sales targets. Sales forecasts are made by product group and other demographic factors, such as region, on a monthly basis and into a future time. Logistics needs forecasts to know where, when, and how much capacity is needed (e.g., inventory) to satisfy demand, and to plan capital expenditures to efficiently move materials and information across its global supply chain. Forecasts are required for items at stocking locations at the appropriate time and into the future. The common time interval for forecasting was monthly, but now the lead time to meet demand is measured in weeks, days, and even hours. Manufacturing requires forecasts for purchased materials and components or dependent demand items aligned to its manufacturing schedules and offset by component lead times. Finance needs accurate forecasts to set sales, cost, profit, and cash flow projections and targets at business, divisional, and product line levels.

Selection of a forecasting time horizon is an important requirement for all types of forecasting methods. The forecasting time horizon is the period over which a forecast is estimated and used. It is also called a time fence, within which schedules should not be changed. A forecast also has a time interval. This is the actual length of time for building the forecasting model (e.g., hour, day, week, month, or another convenient interval). When a product’s production schedule arrives at its cumulative lead time or time fence, it becomes firm (or “frozen”) and orders are placed for its dependent demand items (i.e., the components based on cumulative lead time and the bill-of-material dependencies). Orders become firm and scheduled once they are within their time fence or cumulative lead time. A demand forecast not in the frozen window and beyond the cumulative lead time is not fixed and can be modified as shown in Figure 13.4.

Forecasting models also require differing levels sophistication. The simpler time series models such as moving average or exponential smoothing require little skill. In fact, they can be automated. If seasonality is modeled, then 36 to 60 months of data is required to accurately estimate the seasonal indices month-over-month. However, if seasonality estimates are not required, then only 12 to 24 months of data is required to build a non- seasonal model. More complex econometric models, based on multiple linear regression, require advanced analytical skills. The data required to build these types of models is usually a minimum of 10 observations


Time fence.

or time periods per estimated parameter. If a model has three independent variables, then approximately 36 months of demand is required. Digitization over the past decades has enabled the immediate acquisition of sales information by supply chain participants. Highly flexible organizations are moving away from a heavy reliance on forecasting models based on statistics.

A heavy reliance on statistics-based forecasting models with high error rates is detrimental to operations. Forecasting error rates only increase the further out a forecast is made. Forecasts made at a lead time of 30 days or less will be more accurate than those made on a quarterly or an annual basis, all other things equal. Forecasts are also more accurate when aggregated to a product group than at lower levels such as a specific item at a specific location. Studies have shown that, at an organizational level, statistical forecasting can be successfully applied to forecast product revenue with error rates between 1% and 5%. In contrast, at an item and location level, error rates can easily exceed 25% from month to month. Because forecasting error is difficult to reduce, organizations should develop strategies that do not rely exclusively on statistics-based forecasting.

Figure 13.5 shows how error statistics are calculated for forecasting models with an example. Forecasting errors are calculated after actual demand is realized and then compared to the original forecasted quantity by time interval. Although there are several forecasting error statistics, two are commonly used by most organizations. The first is percentage error, which is calculated by subtracting actual demand from its forecast and dividing by actual demand. The advantage in this calculation is the error statistic is a percentage and adjusted for volume. This makes it useful for comparing error percentages across different products having differed demand patterns. A modification of the percentage error statistic is the mean absolute percentage error statistic, which is calculated as the average of absolute percentage errors over several time periods.

The root mean square error or deviation (RMSE or RMSD) is the second error statistic. It is calculated on a per-unit basis rather than a percentage basis. It can be substituted for the unit standard deviation in safety-stock calculations. As an example, if a forecast is very accurate as measured by its RMSD, its required safety-stock inventory quantity will be less if the RMSD is used rather than its standard deviation. The concept is this: if I can forecast perfectly, I do not need safety stock. This situation is very unlikely because the levels of safety stock are calculated based on average


Calculating forecasting error.

demand over lead time, service level, as well as the standard deviation of demand.

Forecasting error statistics are an important basis for continuous improvement and can be used to identify beneficial projects. Table 13.5 lists several common process issues caused by low forecasting accuracy. Major impacts include customer service, manufacturing, operations, finance, transportation, marketing and sales, inventory management, logistics, and others. Each issue is a major project focus area and creates a portfolio of several projects to implement a solution. Business benefits are easy to calculate for these types of projects because the root-cause problems increase cost and lead time and lower customer satisfaction.

Forecasting models rely on underlying non-random patterns such as trends, periodicity, seasonality, or cycles. These should also be relatively stable and repeatable over several time intervals. Several years of historical data are required to build a seasonal model using a period of one month

TABLE 13.5

Poor Forecasting Impact





  • • Schedule changes
  • • Overtime expense
  • • Schedule changes
  • • Overtime expense



  • • Increased inventory carrying expense
  • • Lower cash flow
  • • Increased labor expense
  • • Lower cash flow



  • • Unnecessary product transfer
  • • Higher premium freight expense
  • • Unnecessary information transfer
  • • Higher premium freight expense



• Excess and obsolete inventory

• Excess and obsolete promotional materials

Inventory Management

Inventory Management

  • • Lower inventory turns
  • • Higher inventory obsolescence
  • • Lower transaction expense
  • • Employee skill obsolescence



  • • Excess warehousing space and expense
  • • Damaged product expense

• Excess warehousing space and expense

Customer Service

Customer Service

  • • Poor line item availability
  • • Backorders
  • • Customer complaints

• Customer complaints

to estimate the monthly seasonal indices. Regression models require incorporating forward-looking information such as leading economic indicators as well as extensive historical records when long cycles exist at a macroeconomic level. There are circumstances where standard forecasting methods cannot be used to build a model. These include unusual events including a loss of market share, recessions, and other major events. Forecasting models are also ineffective for one-off events without underlying patterns. Examples include disruptive technology that threatens an industry.

The sources of independent demand for end items (as opposed to dependent demand items that comprise the end item) are aggregated by the demand management module within a forecasting system to build the models. The two common types of forecasting models are time series, which use exponential smoothing algorithms, and multiple linear regression models. Exponential smoothing models are built using historical demand, with smoothing parameters to fit the model to the historical pattern, a time interval, and for a specific length of time. A forecasting system also creates management reports of various types that break forecasts down by time period, geographical location, organizational level, product group, and other variables. The forecasts are calculated using a unit basis but are converted into monetary units using standard cost data. Forecasting accuracy metrics are also created by these systems to aid in continuous improvement efforts.

An effective forecasting system must have certain attributes to create efficient and accurate forecasts. First, it is important that a product’s actual historical demand be used to develop models. Some organizations use shipment history to forecast demand. The problem is that shipments are influenced by inventory availability. If inventory is not available for shipment, then demand may appear lower than actual. Forecasts based on shipment histories create chronic backorders because inventory will always be set lower than actual demand. Basing forecasts on actual customer demand is the proper method to build a forecasting model. If items are not available for shipment, the forecast will still be based on original customer demand and the product’s target inventory level will be constant except as the shortage is made up. It is very useful to simultaneously record shipment history, actual demand, and forecasted demand. Differences between any of these histories enables an analysis to eliminate ordering discrepancies due to forecasting errors.

Time series with periodic patterns such as seasonality require at least three to five years of month-over-month observations to ensure the monthly seasonal indices are estimated accurately in the model. The time horizon should set at least as far out as the annual operating plan and even further to plan future capacity. It is also important that forecasting accuracy be very high based on the time fence concept that was shown in Figure 13.4. This ensures manufacturing schedules are accurately estimated.

Non-random patterns caused by product promotions or other atypical demand patterns must also be analyzed and incorporated into a product’s forecast. In some industries, quarterly or seasonal adjustments are made. In effect, there would be four similar forecasting models. Forecasting analysts should also specify the periodicity of a forecasting model and truncate the time series’ historical basis if there are recent changes to its underlying pattern. This might occur if a product is at the end of its life cycle and demand exhibits irregular patterns. Analysts would use the most recent demand history to ensure the model is current. Depending on the application, additional information may be needed, such as underlying macroeconomic trends (e.g., recessions or expansionary cycles).

Effective forecasting systems should automatically create and track product forecasts. This is useful to organizations that sell thousands of products. Automatic tracking of a model enables analysts to focus attention on products with unusual demand patterns or are critical to the organization’s operation. Aggregation and desegregation of product forecasts are also made, from an item’s location up to the product group level. A forecasting system should also enable continuous improvement by providing accuracy metrics for each model.

The strategy of an effective demand management system is to make an increasingly larger proportion of its product demand firm while simultaneously decreasing the percentage of forecasted products. Expanding customer relationships and leveraging digital technology facilitates less reliance on product forecasting by capturing demand at its source. Demand estimation accuracy is especially important for industries that sell expensive products, such as capital equipment. Alternatively, the application of advanced technology can be used to measure customer demand at its POS in some industries. If POS information is available, manufacturing schedules will be accurately updated.

An effective demand management strategy optimizes resource effectiveness and utilization. This enables analysts to focus on creating the right forecast models for different demand patterns. Table 13.6 shows how products are stratified by unit volume and variation. It may be possible to place low-volume products with low demand variation on a minimum/maximum (min/max) system and managed using a simple rule. This rule states that when inventory reaches a certain level, more of the product is scheduled for production. The set level is the product’s or component’s reorder point. This stratification strategy enables forecasting analysts to focus on higher- volume or critical products. On the other hand, high-volume products

TABLE 13.6

Strategic Forecasting

Low Demand Variation

High Demand Variation

Low Volume

Standard forecasting models

Special teams to investigate

High Volume

Min/max systems

Special models used to predict unusual demand patterns

having low demand variation can be automatically forecast using time series models. For products with low volume but high demand variation, special forecasting methods are needed to estimate their future demand.

Reconciliation of forecasted demand across an organization helps ensure operational plans and schedules are aligned to a “one number” consensus forecast. The strategic forecast is reconciled by a sales and operations planning process based on a cross-functional consensus. This reconciliation is developed as marketing and sales plans are broken down to a product group level and demand forecasts are verified by each team. The allocation is applied linearly from the one number down through product groups and locations to an item location level. Discrepancies related to capital, labor, and available capacity to meet production schedules are reconciled through the sales and operations planning process. Updated forecasted quantities, once verified by stakeholder groups, are aggregated upward in reverse. The updated consensus forecasts are represented in several forms, including unit quantities and revenue calculated using quantity and standard cost. The strategy for managing demand by progressive organizations is that it needs to be managed rather than pushing unintelligent forecasting models to drive scheduling. Forecasting models are needed and used by almost every major corporation to estimate demand, but these should only be used when more accurate methods cannot be cost effectively utilized.

The common time series forecasting models are listed in Table 13.7. These have several components that describe the time series average level, a trend if present, seasonal variation if present, and longer-term cyclical patterns. These components are modeled by using the appropriate model listed in Table 13.7. Models forecast demand into the future at time periods f+1, where t is the current period. They use actual demand from the current period and previous time periods (i.e., t, f-1, t-2, t-3, etc.). These previous time periods are the lagging dependent variables. As an example, the sales forecast in units next month is estimated using sales from this month and from previous months. For this example, sales last month lag the forecasted sales of the dependent variable of “sales next month.” Time series models also use smoothing parameters to weight the impact of various lagging dependent variables. By modifying the parameters, lagging dependent variables can be weighted to make more recent history more impactful or not.

Figure 13.6 shows monthly sales data collected over eleven years. The graph shows a periodic pattern in the sales data and an upward trend.

TABLE 13.7

Time Series Models



Trend plot

Plots time series data versus time without creating a mathematical model.

Time series decomposition

Breaks a time series into its level, trend, seasonal, and irregular components. It models both trend and seasonal patterns using constants calculated from the decomposition.

Moving average models

A time series model created by taking the average of observations from the time series to smooth out seasonal or other data patterns.

Simple exponential smoothing

Models a level (stationary) time series (i.e., no trend or seasonality) using one smoothing parameter.

Double exponential smoothing (Holt’s method)

Models a level (stationary) time series with a trend but no seasonality using two smoothing parameters.

Triple exponential smoothing (Winters method)

Models a level (stationary) time series with a trend and seasonality using three smoothing parameters.

Autoregressive integrated moving average models (ARIMA)

Statistically based time series models that model level, trend, and seasonal components of a time series.

It will be obvious from Table 13.7 that some time series models fit this historical pattern better than others. Best model fit can be verified using the mean absolute percentage error statistic. The accuracy of a time series model requires fitting a model to historical data and measuring the deviations of forecasted to actual values at each forecasted period. As a side note, when statisticians compare a new forecasting method against previous ones, they use a standardized database that is available to all researchers. It consists of several hundred time series having differing patterns. In these evaluations, forecasting models are compared by fitting them to earlier parts of each reference time series (historical basis) and forecasting demand for the most recent part of the time series called the holdout period. The error statistic is calculated by comparing the forecast to the data in the holdout period. Either Winter’s method or a time series decomposition model will be suitable for modeling the level, trend, and seasonal components.

The decomposition method breaks this time series’ pattern into components. Table 13.8 shows the logic for doing a decomposition. The first step


Decomposition model. This model adequately captures the seasonal pattern: seasonal length 12 > multiplicative model > trend plus seasonal > first observation is in seasonal period 1 > number of forecasts 12 > starting from origin 132. MAPE = mean absolute percentage error; MAD = Mean Absolute Deviation; MSD =Mean Squared Deviation.

is to fit a moving average model to the time series with the same interval and periodicity (i.e., days, weeks, months, quarters, or years). In the current example, the interval is monthly sales, which has an annual seasonal pattern that repeats every twelve months. Creating a moving average model eliminates the seasonality of the original time series. When the moving average time series is subtracted from the original time series, the seasonal indices can be calculated. If the data set is quarterly, then there will be four seasonal indices. The decomposition method continues until all components have been isolated. The irregular component is the variation of sales for which the model cannot account (i.e., the models error).

Figure 13.6 shows the decomposition model is a good fit to historical sales and the extrapolated forecast exhibits a pattern like the original time series. The mean absolute percent error is 5% (rounded), which indicates that 95% of the sales variation is explained by the model.

Demand management drives global supply chain planning. Problems with demand estimation, whether from forecasting models or consensus models, cause misalignment of scarce resources and process issue of many types throughout the supply chain. Customer satisfaction and operational

TABLE 13.8

Time Series Decomposition

1. Start with a time series.

2. Time series = Trend + Cycle + Seasonality + Irregular

3. Calculate a new time series using a moving average model having the same order as the seasonality of the time series to create (A).

4. (A) Time Series = Trend + Cycle

5. Determine the seasonal indices. Deseasonalize the original time series to create (B).

6. (B) Time Series = Trend + Cycle + Irregular

  • 7. Subtract (A) from (B) to obtain the irregular component to identify outliers that may need to be adjusted.
  • 2. Toriginai Time series = Trend + Cycle + Seasonal + Irregular 4- TTrend and cyde only Trend + Cycle Toeseasonaiized Time series Trend + Cycle + + Irregular
  • 2- Tjrreguiaj Component *" Irregular

productivity are reduced by inaccurate demand because capacity planning is incorrect and schedules are missed. Forecasts need to be strategically aligned at all levels, and a consensus forecast must be developed by the sales and operations planning team. To the extent that forecasts are necessary, it is important to measure and continually improve their accuracy. With digitization, organizations should emphasize customer relationships and developing systems to gather customer demand automatically through digitization methods. This will provide visibility to demand across the global supply chain using POS data and technologies.

Strategic forecasts are made at long time horizons to ensure capacity is available to produce products or services in the future. They also identify whether there is a future need for a new facility, enhancements of current locations, or facilities need to be relocated. These future needs may be regional, national, or global, and may be either temporary or permanent. Strategic capacity planning is important to an organization because a failure to adequately plan for the necessary capacity will result in lower sales, higher supply chain costs, and longer lead times.

Supply chain capacity planning also supports expansion into existing and new markets. Important components of the planning process are estimates of facility capacity, location, available equipment, materials, and needed employee skills. Design and available capacities are estimated using strategic forecasts period-by-period, over the forecasting time horizon, for each location in the supply chain. They are calculated using the expected throughputs across the supply chain’s participants. Allowances are made that reduce the design capacities to available capacities. Available capacity is estimated at an aggregated level, by quarter and month, to ensure it is available where it is needed to meet forecasted demand. Business unit forecasts are made by year and broken into quarters and months by facility. At a facility level, forecasts are broken into months and weeks by product group. Capacity is planned on an aggregated basis for each facility and for the processes within them. Supply chain productivity is directly tied to how well capacity is utilized by its participants to meet external customer demand. Digitization is enabling creative ways to effectively make capacity available across supply chains.

Table 13.9 list other ideas to increase available capacity to meet demand using the tools, methods, and concepts discussed in this book. Implementation of these ideas may depend on preceding improvement actions. Efficiently matching available resources to demand through accurate forecasting and scheduling methods ensures a system is not idle or producing the wrong products, services, or information. Scheduling should be managed to the system’s bottleneck, which must be matched to the takt time to prevent a buildup of excess WIP inventory. Excess inventory wastes capacity that could have been used to produce other products or left idle. Increasing scheduling flexibility using methods such as mixed- model scheduling enables systems to respond dynamically to variations in external demand and optimizes capacity. Transfer batches using unit-flow also optimize capacity. In other words, scheduling rules impact available capacity.

Supply chain capacity is increased by accurate forecasting as well as product and service design modifications that impact supporting processes by reductions of components or process steps with standardization. There are other strategies to match capacity to demand. Higher process yields using Lean, Six Sigma, and total productive maintenance increase available capacity because scrap and rework are reduced and production does not need to be replaced. Some facility layouts are more efficient than others for bringing together people, equipment, and materials to reduce waiting and unnecessary movement. Highly efficient layouts increase capacity over less efficient ones. The more highly cross-trained people are, the better they can match resources to meet customer demand. The same concept applies to the selection of equipment. To the extent that organizations can deploy multipurpose and simple machines, the greater the organization’s operational flexibility will be. Finally, an organization can work with other organizations to plan and allocate capacity across a global

TABLE 13.9

Ideas to Increase Capacity


1. Match resources to demand through effective demand management.

2. Work with customers to level load demand.

3. Use price and promotional policies to level load demand.

4. Enable customers to self-service.

5. Hire temporary or part-time workers.

6. Use overtime to handle demand.

7. Overlap work shifts.

8. Lease facilities.

9. Lease equipment.

10. Modify' product design.

11. Modify' process design.

12. Modify' the facility layout.

13. Move to mixed-model production strategies.

14. Use transfer-batch processing.

15. Change the service discipline and its ty'pe (e.g., first come, first serve).

16. Improve quality.

17. Reduce operational cycle time or process lead time.

18. Increase equipment and worker availability.

19. Use single-minute exchange of dies to reduce setup time and cost.

20. Transfer lessons learned across the supply chain.

21. Increase the skill flexibility' of workers through cross-training.

22. Deploy simple and standardized redundant equipment that can be utilized at different rates or levels at a bottleneck.

23. Create virtual capacity using technology and agreements and alliances.

24. Outsource non-core operations.

25. Insource related work to core operations.

supply chain through a strategy of outsourcing and insourcing of work to improve asset utilization efficiencies and productivity.

Another effective strategy to efficiently meet customer demand without large investments in infrastructure is to employ virtual capacity by leveraging IT solutions across supply chain participants at high resource activation. Virtual capacity implies capacity can be created using technology deployed within processes to help manage resources efficiently. Virtual capacity is also integrated using agreements, contracts, and other alliances. Examples include joint ventures, partnerships, sales of non-core assets, the sharing of information and resources, as well as modifications to organizational structures such as decentralized management. These enable supply chains to increase capacity by building non-core and redundant infrastructure with other participants and avoid large capital expenditures. This increases flexibility, and capacity can be matched to demand more easily.

Asset utilization will also be higher with virtualization of capacity because flow can be balanced across a dispersed network. In some systems, such as global call centers, activation and utilization rates may exceed 98% using technology to move calls to facilities around the world based on volume and time of day or week. Bottlenecks and capacity-constrained resources are virtualized by redundancies in people and systems. As an example, geographically dispersed call centers can transfer incoming and outgoing calls immediately from those without available capacity to others with capacity. Virtual capacity is also enabled through partnerships that reduce investment and operation costs. This flexibility enables supply chains to easily enter and leave markets without high costs. In effect, technology, information, collaboration, and new organizational systems replace physical facilities and objects.

The extraordinarily high productivity and operational capability of today’s global supply chains is enabled by the rapidly evolving evolution of IT systems discussed in Chapter 8, which enables information to be collected, analyzed, and used to change a process based on software rules and algorithms such as those that control scheduling. As an example, scheduling algorithms enable changes in the production sequences based on changes to a system’s status to reduce lead time to meet customer service targets while optimizing capacity. Scheduling rules and algorithms were discussed in Chapter 6. In summary, developments in IT can eliminate manual interventions and intermediaries, and it provides information status and reporting, enables adaptable sequencing and scheduling, helps coordinate work across the global supply chain, enables rules for asset sharing amongst participants, and promotes the production of differentiated products and services.

< Prev   CONTENTS   Source   Next >