Operations Management / Supply Chain Management

Module 03.02 Key Concept: Forecasting Methods

The Text presents several different forecasting options.

Qualitative techniques are generally based on judgment, intuition, and informed opinions.  Quantitative techniques are based on statistical calculations using collected data.  Qualitative techniques fall into two general categories:

  • Extrinsic – based on external indicators that relate to demand; that is we use data external to the demand that we want to forecast
  • Intrinsic – the use of historical data to create forecast; that is we use data internal to the demand

Extrinsic quantitative forecasts are sometimes called “Causal” forecasts and often require some type of regression analysis.

The text explores some of the more simple approaches and these will be presented below.  With all these different options (and more) available, what method is best to use?  Our favorite answer applies:

It Depends!

It depends on the product / product family / business portfolio.  Marketing sometimes groups products or families of products according to dollarized potential and probability of success (market growth).  Established products tend to have relatively stable demand on average whereas risky products, or products supplying highly dynamic markets, tend to have unstable demand.  Forecasting methodology must be selected that best fits current and future demand patterns.

It depends on the product / product family / business life cycle.   New products usually have very unpredictable and erratic demand then during the growth phase may show indication of an upward trend.  Once the business reaches saturation, demand is stable and then during the product phase-out or substitution phase there may be a negative trend and increased volatility depending on the withdrawal process and customer acceptance.  In early and late stages, qualitative methods may be employed because of the lack of consistent data.  In the growth phase, regression analysis techniques may be preferred and in the stable phase, averaging techniques may provide the best results.

It depends on the general patter of demand.  The APICS Dictionary definition of a forecast was presented earlier.  This definition considers four different, frequently found, demand patterns: random, trend, seasonal and cyclical.

Random demand is considered having no predictable pattern. For example, sales data may vary randomly about some forecast value with no specific pattern and no attendant ability to obtain a more accurate sales estimate than the forecast value.  We will later see that Safety Stock is usually carried to protect against stock-out due to this random variation.  In some cases random variation can be relatively small but in others it can be dramatic.  The dynamic nature of random variation is a major consideration in forecast model selection.

A demand trend usually describes a general upward or downward movement of a variable over time.  There are numerous types of trends that may be found.  However in Chapter 3 of the text, the authors only describe forecasting methodology for “linear” demand trends.

Seasonal demand describes the impact variations that occur because of the time of year (quarter, month, week).   Seasonal demand shows a recurring pattern, year-on-year.

Cyclical demand is usually thought of, or experience, changes in the business cycle – long-term fluctuations up and down.  We are now experiencing the dramatic impact of an economic downturn and slow recovery.

The best way to identify particular demand patterns is by graphing the data.

The basic principles of forecasting were presented earlier. We will now explore some of the specific techniques available for use with each of these different demand patterns.

Four Qualitative approaches are highlighted in the text:

  • Jury of Executive Opinion
  • Delphi Method
  • Sales Force Composite
  • Market Survey

Jury of Executive Opinion typically involves a small group of high-level experts and managers.  Group estimates of demand are established through collaborative analysis.  Thus, the process often combines managerial experience with statistical models.  An advantage is that the analysis is quick but it may result in a type of “Group-think” error.

The Delphi Method is an iterative group process that continues until consensus is reached.  Delphi Teams typically include decision makers, staff and external respondents.

With Sales Force Composite, each salesperson projects his / her customer sales.  The results are then consolidated at district and national levels.  In general, sales representatives should know their customer wants and thus be more able to predict future demand.  However, predictions may often be overly optimistic.

In the Market Survey approach, customers are asked about future purchasing plans.  This approach is useful for demand and product design planning.  These forecast may also tend to be overly optimistic.  What customers say may be different from what they actually do.

The text describes many different Quantitative approaches.  Among them are: moving average, exponential smoothing trend / linear regression and seasonal forecasting.  All of these techniques (and more) are available in modern ERP systems and add-ons.

These are often referred to as “Time-Series” forecasting methods.  A set of evenly spaced numerical data is collected over time.  Forecasts are based on past history with the expectation that the past results can be used to make an accurate projection of the future.  This approach also assumes that factors influencing the past and present will continue influence in the future.

The text presents several short-term statistical forecasting methods.  While they can be used for dynamic demand patterns, they work best when the demand is relatively consistent and variation in demand is due to random variation.  If the demand data is highly erratic, forecast error will be high as well.  Some popular methods are presented below:

  • Naive Approach: assumes that the demand in the next period will be the same as the demand in the most recent period.  For example, if January sales were 68 units, then February sales will also be 68 units.  This is a simple approach that can sometimes be cost effective and efficient.  This technique also represents a good starting point for other approaches.
  • Moving Average: un-weighted average of a given number of past periods is used to forecast the future.  A weighted moving average assigns weights to past month data in order to provide forecasts that are more – or less – responsive to current market changes.  The moving average forecast will lag the development of a rising or falling trend.  The farther back the moving average forecast reaches for data, the greater the lag.  The moving average forecast works best when demand is stable with random variation; it will “filter out” random variation
  • Weighted Moving Average: used when some trend is present.  In this case, older data may not be as relevant as recent historical data.  Weights are established based on experience and intuition.
  • Exponential Smoothing: weighted average of all past periods is used to forecast the future

As noted above, these approaches assume that there is an underlying pattern of demand that is consistent over some period of time.

The moving average forecast = average demand of a selected number of past periods.  It is calculated by:  (Σ Demand for “n” months)  / n;
n is known as the order of the Moving Average and it represents the # of periods of past data that you want to average from.

Examples are presented in the text.

The Exponential Smoothing forecast requires three pieces of data:

  • Previous Month’s Forecast
  • Previous Month’s Actual Sales
  • Value for the smoothing constant alpha (α).

Alpha can be any number between 0.0 and 1.0.  A low value for alpha gives more weight to the old forecast.   This is appropriate if demand is stable (not rising or falling.)  A high value of alpha gives more weight to recent actual demand.  It is more appropriate with dynamic demand as it results in forecasts more responsive to the changing environment.  It is best to run simulations with different α (alpha) values to see which one best fits the historical demand pattern

Exponential Smoothing Forecast =

Last Period’s Forecast + (α x (Last Period’s Actual Demand – Last Period’s Forecast))

 

Forecasts with Trends and Seasonality

The text presents more advanced techniques that are applicable to non-random demand data.  The calculations are more complicated than for the techniques described above.  However, they are easily handled with modern forecasting systems and in spreadsheets.   We will look at some formulas and examples for illustration purposes.  However, spreadsheets should be used in practice with the appropriate formulas entered.

Two methods of interest here is incorporating trend and/or seasonality into the forecasts.

  • A demand trend usually describes a general upward or downward movement of a variable over time.  There are numerous types of trends that may be found.
  • Seasonal demand describes the impact variations that occur because of the time of year (quarter, month, week or even time of day).   Seasonal demand shows a recurring pattern, year-on-year. In simple terms, seasonality in the demand data means the demand at some periods is larger than the average while at some other periods, it is smaller than the average.

Regression identifies a relationship between two or more correlated variables.  Linear regression is a special case where the relationship is defined by a straight line, used for both time series and causal forecasting.  The algebraic expression used to describe a straight line is as follows:

Y = a + bX

Where “Y” is value of dependent variable, “a” is the y-intercept of the line, “b” is the slope, and “X” is the value of the independent variable.   With linear regression, the objective is to find the line that minimizes the sum of the squares of the vertical distance between each data point and the line.  This requires calculation of several different elements from the data at hand.  This technique is described in the text but we are not going to explore in detail here.

With the Linear Regression approach:

  • We always plot the data to insure a linear relationship
  • We do not predict time periods far beyond the database
  • Deviations around the least squares line are assumed to be random

Seasonality may (or may not) be relative to the general demand trend.  Seasonality must be validated over several years to ensure that there is a true seasonal demand pattern.  Also, seasonality in general can be applied to anything.  For consideration, it is not just based on Fall, Winter, Spring, etc.  Just think of your local McDonald’s.  Are there certain periods of the day when demand is on average higher than other time periods?  At your bank, are there different days that are busier than others?  These provide practical views of the seasonal pattern of demand.

Seasonal forecasting employs a multi-step process:

  • Find average historical demand for each month
  • Compute the average demand over all months
  • Compute a seasonal index for each month
  • Estimate next year’s total demand
  • Divide this estimate of total demand by the number of months, then multiply it by the seasonal index for that month

 

 

The text brings in the concept of Trend Correction of Seasonal Forecasts.  It is possible to use statistical tools and the following process to calculate a reasonable forecast for the coming year.  However, the basic analysis process uses the same general steps:

  • Decompose the time series into its components
  • Find seasonal component
  • Deseasonalize the demand
  • Find trend component
  • Forecast future values for each component
  • Project trend component into future
  • Multiply trend component by seasonal component

The text also brings in the concept of Associative Forecasting.  This is very similar to the Linear Regression approach except that time series data is not used.  In this case changes in one or more independent variables can be used to predict the changes in a dependent variable.  The text authors provide an example that shows the relationship between store sales and store payroll.