3. Demand Planning

It’s only been in the past 20 years or so that businesses have truly come to realize the importance of forecasting. If you think about it, forecasting is usually the first step in the planning and scheduling process for most goods and service organizations, and forecasts for demand drive everything in an organization: from longer-term decisions (3+ years out) as to new facilities and products, to medium-term decisions (months to years out) such as production planning and budgeting, and the short-term (months to a year at most), where we need to know what to produce (or purchase) and deploy (see Figure 3.1).

The function itself has evolved from being an almost “dreaded” responsibility of sales and marketing, to where operations took control to produce stable production requirements, and on to today where it is most typically part of the supply chain function, where it can rise to the level of importance in an organization to the point where there may be a director of forecasting or demand planning. In fact, there is now a professional organization dedicated to the profession: the Institute of Business Forecasting & Planning (www.ibf.org).

These limitations were due primarily to companies working in more of a “vacuum” because data was hard to get and limited to internal sources and storage space expensive and limited.

Many companies also operated under a two-number system, where sales and marketing budgeted one number (which might be changed only once per quarter) and manufacturing developed their own SKU (stock keeping unit) forecast based on more current sales. In some cases, there was even a third number used by those responsible for finished goods deployment to distribution centers, which in many cases was based on percentage allocations of one of the two aforementioned national or global forecasts.

As technology became more readily available (and more affordable) in the mid-1980s to early 1990s, many businesses were able to begin to get a better handle on the forecasting process. Using the pyramid approach to forecasting (see Figure 3.2), organizations were able to develop a bottom up/top down one-number forecast, which used various statistical methods as well as other sources of information at various levels of detail. These one-number forecasts were able to drive budgeting, production, and deployment simultaneously and be updated typically on a monthly (or more often basis).

The availability and sharing of point-of-sale (POS) data from either paid services or larger customers was also integrated into the process using collaborative programs between manufacturers and retailers such as quick response and efficient consumer response (to be discussed in more detail later) to help reduce the bullwhip effect.

All forecasts are wrong. It’s rare that a forecast is 100% accurate. The idea is to have an integrated, collaborative process that minimizes variance of actual versus target. You’ll learn later in this book about the process and importance of setting and measuring forecast accuracy targets.

The more “granular” the forecast, the less accurate it is. A national forecast for a family of items is likely to be more accurate than a weekly forecast for an SKU at a distribution center that handles a region of the country. We can compensate for some of the inaccuracy through proper inventory planning, factoring in scientific safety stock inventory based on desired service levels and reduced lot sizes and cycle times included in Lean, as covered later in the book.

It’s easier to forecast next month more accurately than next year. If we know what we sold yesterday, we typically have a better idea of what we’ll sell today; whereas 12 months from now, a lot of things can happen that can affect sales.

You will get a more accurate forecast using demand history rather than sales history. Years ago, when data storage costs were high and capacity lower, most companies only stored sales information. Today, most store order or demand information as well. Unless your company has a 100% service level, there will be occasions where you ship short, late, or from the “wrong” location. If you only use sales history, you would be forecasting to repeat yesterday’s failure. That’s why you should always use demand history to drive statistical forecasts.

Forecasting really is a blend of art and science. As we will discuss, there are both qualitative and quantitative methods of forecasting. Today, the best practice is a combination of both, in addition to collaboration with supply chain partners, providing better visibility downstream in the demand chain.

Marketing requires forecasts to determine which new products or services to introduce or discontinue, which markets to enter or exit, and which products to promote.

Salespeople use forecasts to make sales plans, because sales quotas are generally based on estimates of future sales.

Supply chain managers use forecasts to make production, procurement, and logistical plans.

Finance and accounting use forecasts to make financial plans (budgeting, capital expenditures, and so on). They also use them to report to Wall Street with regard to their earnings expectations.

For example, Procter and Gamble and Walmart have partnered in everyday low pricing (EDLP) to reduce costs and improve service. At one point, P&G went as far as stationing 200 employees at Walmart’s headquarters in Bentonville, Arkansas.

Previously, supply chain and operations were at the mercy of these internal drivers and had to live through the consequences. Of course, they will always exist to some degree.

1. Determine the use of the forecast. Varies by industry and company. In the case of manufacturing, it may be to drive production and deployment, in the case of retail it might be to determine purchasing requirements and for pure service companies might be used primarily for labor staffing.

2. Select the items to be forecasted. Will we be forecasting by individual item in various granulations, and at what levels and units of measure will be need to be able to aggregate forecasts and demand history?

3. Determine the time horizon of the forecast. Do we need to look at it in the short, medium, or long term (or all of the above), and what type of time planning buckets are appropriate (for example, 30 days or less: daily buckets; 1–3 months out: weekly buckets; and 4+ months: quarterly buckets).

4. Select the forecasting model(s) and methods. Based on a number of things we will be discussing, such as where a product is in its lifecycle, will we use qualitative, quantitative, or a blend of models? To what degree will we integrate externally supplied information (for example, customer forecasts, POS data, CPFR, and so on), and what weight will we give it?

5. Gather the data needed to make the forecas. When using forecasting software, the initial integration will consider much of this, such as using demand versus sales, as mentioned previously, eliminating data errors, and so on. Once this integration has been created and data validated, it becomes more of a maintenance issue for things such as new and discontinued items.

6. Generate forecasts. Typically, statistical methods are used to generate a baseline forecast, possibly at different levels of detail. The planner will then usually audit the results and, if needed, try other statistical models. They will then factor in management overrides based on their experience and knowledge as well as promotional plans, sales estimates, and externally supplied information mentioned earlier.

7. Validate and implement the results. During the demand part of the sales and operations planning (S&OP) process that we will discuss in the next chapter, forecasts are reviewed by cross-functional teams at various levels of detail and units of measure to ensure the highest level of accuracy possible. This will ultimately lead to a one-number system that was discussed earlier so that everyone is on the same page.

During this time, recent forecast accuracy will be evaluated, as well, to help target improvement. It is also during this step that the new forecasts are saved to be measured later for accuracy against predetermined variance/error targets, as discussed later in the chapter.

Many companies cycle through this process on a monthly basis (this again varies by industry and how the forecasts will be used), but forecasts are typically adjusted on an as-needed basis due to over/undersells, new demand information, changes to promotions and discounts, and so on.

Qualitative methods include knowledge of products, market surveys, jury of executive opinion, and the Delphi method.

However, you must always be aware of biases that may result from different individual’s motivation. For example, sale personnel may have an incentive to hit a high target to reach a bonus.

In my experience as a senior forecaster at Unilever, after awhile I was able to determine that sales estimates were typically 50% high, and once factoring that in, they were fairly useful (at least to start a dialogue). It is also important to be able to share forecast and historical data in units of measure and levels of aggregation that are meaningful to others. For example, the sales department thinks more in terms of revenue dollars, customers, and product categories. So, a good forecasting process and software system should be able to convert data back and forth to both present the data and receive feedback.

Another example are focus groups where people are asked about their perceptions, opinions, beliefs, and attitudes toward a product, service, concept, advertisement, idea, or packaging. Questions are sometimes asked in a group setting, and participants can talk with other group members.

Both the Delphi and jury of executive opinion forecasting methods are usually a bit more strategic in nature and used more in developing higher-level longer-terms forecasts.

Note that the product lifecycle also has an impact on the supply side, as covered in the next chapter.

The phases in the lifecycle of a product or service are introduction, growth, maturity, and decline (see Figure 3.3).

A general rule of thumb in forecasting is that to generate a decent statistical forecast, you need at least 12 months of history. So, during this growth phase, forecasting is truly a blend of art and science, as both quantitative and qualitative methods are both used to create a blended forecast.

During the growth phase, it can be very easy to over- or underestimate forecasts, which can have dramatic effects on cost and service. So, great care must be taken, and all lines of communication must be established and open both internally and externally, to avoid surprises where possible (which in some cases, such as new distribution, may be hard to avoid).

Eventually, the product may be discontinued. However, forecasts must still be generated to run out existing inventory. Therefore, similar to the introductory phase, the forecaster relies more on qualitative than on quantitative methods.

Trend: An ongoing overall upward or downward pattern with changes due to population, technology, age, culture, and so on that is usually of several years or more in duration (for example, a fashion trend toward smaller bikinis).

Cyclical: Repeating up and down movements typically affected by business cycle, political, and economic factors, which may vary in length and are usually 2 to 10 years in duration. There are often causal or associative relationships. Examples include economic recessions.

Seasonal: A regular pattern of fluctuations due to factors such as weather, customs, and so on that occur within 1 year. Examples include natural occurrences such as climatic seasons or artificially created such as the school year or seller promotional plans.

Random: Erratic fluctuations that are due to random variation or unplanned events such as union strikes and war, which are usually relatively short in nature and nonrepeatable.

Because these components can be combined in different ways, it is usually assumed that they are multiplicative or additive.

Naive approach: Last period’s actual demand is used as this period’s forecast, without adjusting them or attempting to establish causal factors. (For example, if January sales were 100, then February forecasted sales will be 100.) It is simple, yet cost-effective and efficient.

Moving average: The simple average of a demand over a defined number of time periods and is used if there is little or no trend because it tends to smooth historical data. Typically, more recent history is averaged to create the estimate. (For example, January–March sales are averaged to create an April forecast.)

Weighted moving average: An average that has multiplying factors to give different weights to data at different positions in the sample window. Typically used when some trend might be present because it treats older data as usually less important. The weights are based on experience and intuition and can be used to minimize the smoothing effect if desired.

For example:

Weighted moving average forecast for April = .6 * March sales + .3 * February sales + .1 * January sales

In this example, more weight has been given to March demand than to January and February to generate the April (and onward) forecast.

Exponential smoothing: A smoothing technique used to reduce irregularities. It is a type of the weighted moving average model where weights decline exponentially, with the most recent observations given relatively more weight in forecasting than the older observations. Exponential smoothing requires an alpha smoothing constant (ranges between 0 and 1 and denoted by the symbol α), which is subjectively chosen.

For example:

New forecast = Last period’s forecast + .7 * (Last period’s actual demand – last period’s forecast)

In this example, the smoothing constant used of .7 will give a relatively high weighting or smoothing factor to an over- or undersell during the most recent month of history when generating the new forecast.

In linear regression, the relationship is defined as: y = a + bx, where a = the y axis intercept and b = the slope of the regression line.

A simple example of linear regression would be to derive and use this equation to predict future sales (y) by plugging in the sales budget (x) that we plan on using (with n being the total number of observations). If we know that the two variables are strongly correlated, we can easily derive the equation using historical sales personnel employment numbers along with historical sales.

To come up with this equation, we must solve for b and then a. The formula used to derive each are shown here:

Once we have solved for a and b, we have our regression formula and can plug in future sales personnel employment estimates to predict future sales.

The range of correlation is 0 to 1. A perfect correlation between two variables would result in an r of +/–1. The lower the correlation between the two variables, the closer to 0 is the result.

A seasonality index is relatively easy to create and can be applied to any of the previously discussed forecasting methods to give the forecast more realistic peaks and valleys.

To create a seasonality index, you must do the following:

1. Calculate an average for all item history (for all years and periods).

2. Average each period’s historical data.

3. Divide each period’s average by the overall average.

4. Apply the period index to the existing time series or linear regression forecast.

Suppose, for example, that a snow shovel that we sell has historical quarterly sales, as shown here.

We can create a seasonality index to apply to a flat moving average quarterly forecast of 100, for example. To do this, we first calculate an average for each period, and then calculate an overall average of 268. From there, we can calculate indices for each quarter by dividing their period averages by the overall average of 268.

If we had a quarterly forecast for next year of 100/quarter, we could apply the seasonality index for each quarter to that forecast. The resulting quarterly forecasts would be: Q1 = 2.07 * 100, or 207 shovels; Q2 = .42 * 100, or 42 shovels; Q3 = .09 * 100, or 9 shovels; and Q4 = 1.42 * 100, or 142 shovels.

As mentioned previously, seasonality can be natural or induced. In either case, it can change over time and needs to be recalculated on an ongoing basis.

In the end, you will want to arrive at the best combination of independent variables for the best possible forecast. A statistic called an r-squared or coefficient of determination, which is the square of the correlation coefficient mentioned earlier and is a measure of the strength of the correlation between y and the various combination of x’s, is calculated. The closer to 1.0 the r-squared is, the better the correlation, and hopefully, the more accurate the forecast.

There are many other statistical methods used, ranging from simple to very complex. The best-in-class methods of forecasting use a blend of qualitative and quantitative methods that include collaboration both internally with staff from various departments, including sales, marketing, and finance, and externally with customers and suppliers.

To under the ABC method, you needs to understand a phenomenon known as the Pareto principle or the 80/20 rule. It states that a relatively small number of your items generate a fairly large percentage of your sales or profits and are referred to as A items (for example, Whopper, fries, and Coke are A items at Burger King).

In forecasting, these A items require more time and effort put into them and typically have better accuracy as a result. The slower movers, known as B and C items, are somewhat less important and require less forecasting time and effort and typically have more variability (Myerson, 2014).

A number of methods are used to measure and monitor accuracy. The main ones are described in the following subsections.

The actual calculation for the MAD is the sum of the absolute error in units divided by the number of occurrences and represented as follows:

Typically, targets are set by ABC code or other methods that highlight relative importance of items, with A type items generally having a smaller variance because they are major, everyday items and so more predictable. C items, because there are more of them with much smaller volume, tend to be more volatile and so tend to have greater forecast variance.

The MAPE is calculated as follows:

The tracking signal is used to determine the larger deviation (in both plus and minus) of error in forecast, and is calculated by the following formula:

Usually, upper and lower control limits (UCL and LCL) for the number of MADs that the tracking signal represents. There are no “magic” numbers for the UCLs and LCLs, because they are somewhat subjective, but keep in mind that 1 MAD = .8 standard deviations.

In a normal distribution, 3 standard deviations (or +/–4 MADs), should include 99.9% of the occurrences. So, if your tracking signal starts exceeding those levels, it is a good indication that something isn’t right.

Historically, at least in terms of best-of-breed forecasting functionality, the software has been licensed from a separate vendor and integrated with the accounting or enterprise resource planning (ERP) software system. These integrated applications were then used to manage business and automate back-office functions for all facets of an operation, including product planning, development, finance, human resources, manufacturing processes, sales, and marketing. More recently, through development and acquisition, accounting and ERP vendors are increasingly adding this and other nontraditional functionality to their systems (see Figure 3.6).

Yet, with all of this, according to a KPMG advisory global survey of 544 senior executives (KPMG, 2007), nearly all organizations still use spreadsheets for some parts of the process; more worryingly, 40 percent of them rely solely on spreadsheets to produce the forecast. As a result, this leaves major chances for losses in efficiency and for redundancies in work processes.

However, the KPMG advisory found that the following separated the best in class from the rest regarding forecasting:

Tend to take forecasting more seriously as they hold managers accountable for agreed-upon forecasts, incentivize managers for forecast accuracy, and use the forecast for ongoing performance management

Look to enhance quality beyond the basics by incorporating scenario planning and use external market reports and data more often

Work harder at it by updating and reviewing forecasts more often and more formally and tend to use packaged forecasting software systems more often rather than just spreadsheets

Now that we have answered the demand question, the next step in most goods-oriented planning processes is this: How much and when do we need to produce or purchase product?