THE DECISION PROCESS

FORECASTING DEMAND

1.1 Introduction

Forecasts are essential for the smooth operations of business organizations. They provide
information that can assist managers in guiding future activities toward organizational goals.

1.2 Forecasting Objectives and Uses

Forecasts are estimates of the occurrence, timing, or magnitude of uncertain future events.
Forecasts are essential for the smooth operations of business organizations. They provide
information that can assist managers in guiding future activities toward organizational goals.

Operations managers are primarily concerned with forecasts of demand—which are often made
by (or in conjunction with) marketing. However, managers also use forecasts to estimate raw
material prices, plan for appropriate levels of personnel, help decide how much inventory to
carry, and a host of other activities. This results in better use of capacity, more responsive service
to customers, and improved profitability.

1.3 Forecasting Decision Variables

Forecasting activities are a function of:

i. the type of forecast (e.g., demand, technological),

ii. the time horizon (short, medium, or long range),
iii. the database available, and
iv. the methodology employed (qualitative or quantitative).

Forecasts of demand are based primarily on non-random trends and relationships, with an
allowance for random components. Forecasts for groups of products tend to be more accurate
than those for single products, and short-term forecasts are more accurate than long-term
forecasts (greater than five years). Quantification also enhances the objectivity and precision of a
forecast.

1.4 Forecasting Methods

There are numerous methods to forecasting depending on the need of the decision-maker. These
can be categorized in two ways:

i. Opinion and Judgmental Methods or Qualitative Methods.

ii. Time Series or Quantitative Forecasting Methods.

1
1.4.1 Opinion and Judgmental Methods

Some opinion and judgment forecasts are largely intuitive, whereas others integrate data and
perhaps even mathematical or statistical techniques. Judgmental forecasts often consist of

i. forecasts by individual sales people,

ii. Forecasts by division or product-line managers, and
iii. combined estimates of the two.

Historical analogy relies on comparisons; Delphi relies on the best method from a group of
forecasts. All these methods can incorporate experiences and personal insights. However, results
may differ from one individual to the next and they are not all amenable to analysis. So there
may be little basis for improvement over time.

1.4.2 Time Series Methods

A time series is a set of observations of a variable at regular intervals over time. In

decomposition analysis, the components of a time series are generally classified as trend T,
cyclical C, seasonal S, and random or irregular R. (Note: Autocorrelation effects are sometimes
included as an additional factor).

Time series are tabulated or graphed to show the nature of the time dependence. The forecast
value (Yc) is commonly expressed as a multiplicative or additive function of its components;
examples here will be based upon the commonly used multiplicative model.

Yc = T. S. C. R multiplicative model (1.1)

Yc = T + S + C + R additive model (1.2)

where T is Trend, S is Seasonal, C is Cyclical, and R is Random components of a series.

Trend is a gradual long-term directional movement in the data (growth or decline).

Seasonal effects are similar variations occurring during corresponding periods, e.g., December
retail sales. Seasonal can be quarterly, monthly, weekly, daily, or even hourly indexes.

Cyclical factors are the long-term swings about the trend line. They are often associated with
business cycles and may extend out to several years in length.

Random component are sporadic (unpredictable) effects due to chance and unusual
occurrences. They are the residual after the trend, cyclical, and seasonal variations are removed.

1.5 Trend

Three methods for describing trend are: (1) Moving average, (2) Hand fitting, and (3) Least
squares.

2
1.5.1 Moving Average

A centered moving average (MA) is obtained by summing and averaging the values from a given
number of periods repetitively, each time deleting the oldest value and adding a new value.
Moving averages can smooth out fluctuations in any data, while preserving the general pattern of
the data (longer averages result in more smoothing). However, they do not yield a forecasting
equation, nor do they generate values for the ends of the data series.

MA=
∑x
Number of Period

A weighted moving average (MAwt) allows some values to be emphasized by varying the
weights assigned to each component of the average. Weights can be either percentages or a real
number.

MA=
∑ (Wt) x
∑ Wt
Example 1: Shipments (in tons) of welded tube by an aluminum producer are shown below:

Year 1 2 3 4 5 6 7 8 9 10 11
Tons 2 3 6 10 8 7 12 14 14 18 19

(a) Graph the data, and comment on the relationship. (b) Compute a 3-year moving average, plot
it as a dotted line, and use it to forecast shipments in year 12. (c) Using a weight of 3 for the most
recent data, 2 for the next, and 1 for the oldest, forecast shipments in year 12.

Fig 1.1

Table 1.1: 3 – year moving average

Year Shipments (tons) 3-year moving total 3-year moving average

1 2 - -
2 3 11 3.7
3 6 19 6.3
4 10 24 8.0

3
 5 8 25 8.3
6 7 28 9.3
7 13 34 11.3
8 14 41 13.7
9 14 46 15.3
10 18 51 17
11 19 - -

(a) The data points appear relatively linear. (b) See Table 1.1 for computations and Fig. 1.1 for
plot of the MA. The MA forecast for year 12 would be that of the latest average, 17.0 tons.

MA wt =
∑ ( Wt ) x = ( 1 ) (14 ) + ( 2 ) ( 18 ) + ( 3 ) (19) 17.8 tons
∑ Wt 1+2+3

1.5.2 Least Squares

Least squares are a mathematical technique of fitting a trend to data points. The resulting line of
best fit has the following properties:

(1) The summation of all vertical deviations about it is zero,

(2) The summation of all vertical deviations squared is a minimum, and
(3) The line goes through the means X and Y. For linear equations, the line of best fit is found
by the simultaneous solution for a and b of the following two normal equations:

∑ Y =¿ na+b ∑ x ¿
∑ XY =¿ a ∑ X+ b ∑ x 2 ¿
The above equations can be used in the form shown above and are used in that form for
regression. However, with time series, the data can also be coded so that ΣX = 0 . Two terms then
dropout, and the equations are simplified to:

∑ Y =¿ na ¿

a=
∑Y
n

∑ XY =¿ b ∑ x 2 ¿

b=
∑ XY
∑ x2

4
To code the time series data, designate the center of the time span as X = 0 and let each
successive period be ±1 more unit away. (For an even number of periods, use values of ±0.5, 1.5,
2.5, etc.).

Example 2: Use the least square method to develop a linear trend equation for the data from
example 1. State the equation and forecast a trend value for year 16.

Year X year coded Y shipments (tons) XY X2

1 -5 2 -10 25
2 -4 3 -12 16
3 -3 6 -18 9
4 -2 10 -20 4
5 -1 8 -8 1
6 0 7 0 0
7 1 12 12 1
8 2 14 28 4
9 3 14 42 9
10 4 18 72 16
11 5 19 95 25
0 113 181 110

We have

a=
∑ Y = 113 =10.30
n 11

b=
∑ X Y = 181 =1.6
∑ x 2 110
The forecasting linear equation is of the form Y = a + bX

Y =10.3+ 1.6 X

For year 16 forecast

Y =10.3+ 1.6 ×16=35.9 tons

1.6 Exponential Smoothing

Exponential smoothening is a moving-average forecasting technique that weights past data in an

Exponential manner so that most recent data carry more weight in the moving average.

With simple Exponential smoothening, the forecast Ft is made up of the last period forecast Ft–1
plus a portion, α, of the difference between the last periods actual demand At–1 and last period
forecast Ft–1.

5
 F t=F t −1 +∝ ( A t −1 −F t −1)

Example 3: A firm uses simple exponential smoothing with α = 0.1 to forecast demand. The
forecast for the week of February 1 was 500 units, whereas actual demand turned out to be 450
units.

(a) Forecast the demand for the week of February 8.

(b) Assume that the actual demand during the week of February 8 turned out to be
505units. Forecast the demand for the week of February 15, Continue forecasting
through March 15, assuming that subsequent demands were actually 516, 488, 467,
554 and 510 units.

Solution

(a) F t=F t −1 +∝ ( A t −1 −F t −1) =500+0.1 ( 450−500 )=495 units

(b) Arranging the procedure in tabular form we have

Week Actual demand Old forecast Forecast error Correction New forecast
At-1 Ft-1 At-1 - Ft-1 α(At-1 - Ft-1) Ft=Ft-1+α(At-1 - Ft-1)
Feb 1 450 500 - 50 -5 495
8 505 495 10 1 496
15 516 496 20 2 498
22 488 498 - 10 -1 497
Mar 1 467 497 -30 -3 494
8 554 494 60 6 500
15 510 500 10 1 501

The smoothing constant, α, is a number between 0 and 1 that enters multiplicatively into each
forecast but whose influence declines exponentially as the data become older. Typical values
range from 0.01 to 0.40. An α gives more weight to the past average and will effectively dampen
high random variation. High α values are more responsive to changes in demand (e.g., from new-
product introductions, promotional campaigns). An α of 1 would reflect total adjustment to
recent demand, and the forecast would be last period’s actual demand. A satisfactory α can
generally be determined by trial-and-error modeling (on computer) to see which value minimizes
forecast error.

Simple exponential smoothing yields only an average. It does not extrapolate for trend effects.
No α value will fully compensate for a trend in the data. An α value that yields an approximately
equivalent degree of smoothing as a moving average of n periods is:

2
∝=
n+¿ ¿

1.6.1 Adjusted Exponential Smoothing

6
Adjusted exponential smoothing models have all the features of simple exponential smoothing
models, plus they project into the future (for example, to time period t + 1) by adding a trend
correction increment, Tt, to the current period smoothed average, ^
F t.

^
F t+ 1= ^
Ft + T t

Figure below depicts the components of a trend-adjusted forecast that utilizes a second
smoothing coefficient β. The β value determines the extent to which the trend adjustment relies
on the latest difference in forecast amounts ( ^ ^ t −1 ) versus the previous trend Tt–1 Thus:
F t− F

F t=∝ A t −1 + ( 1−∝ ) ( F
^ Ft −1 )
^t−^

T t=β ( F ^ t −1) + ( 1−β ) T t −1

^ t− F

A low β gives more smoothing of the trend and may be useful if the trend is not well-established.
A high β will emphasize the latest trend and be more responsive to recent changes in trend. The
initial trend adjustment Tt–1 is sometimes assumed to be zero.

Self-adaptive models: Self-adjusting computer models that change the values of the smoothing
coefficients αs and βs in an adaptive fashion have been developed; these models help to
minimize the amount of forecast error.

1.7 Regression and Correlation Methods

Regression and correlation techniques quantify the statistical association between two or more
variables.

(a) Simple regression expresses the relationship between a dependent variable Y and a
independent variable X in terms of the slope and intercept of the line of best fit relating
the two variables.
(b) Simple correlation expresses the degree or closeness of the relationship between two
variables in terms of a correlation coefficient that provides an indirect measure of the

7
 variability of points from the line of best fit. Neither regression nor correlation gives
proof of a cause-effect relationship.

1.7.1 Regression

The simple linear regression model takes the form Yc = a + bX, where Yc is the dependent
variable and X the independent variable. Values for the slope b and intercept α are obtained by
using the normal equations written in the convenient form:

b=
∑ XY −n XY 1
∑ X 2−n X 2
a=Y −b X 2

In Equations. (1) and (2), X = (ΣX ) / n and Y = (ΣY) / n Y are the means of the independent and
dependent variables respectively, and n is the number of pairs of observations made.