Business Forecasting

Introduction

Forecasting future events in the face of uncertainty is one of the major problems that
managers face to make good business decisions. For example, economic indicators such
as interest rates, energy prices etc. require periodic forecasts for proper financial
planning, forecasts of sales are needed to plan production and workforce capacity,
forecasts of trends in technological innovation are a must for long term strategic
planning. In what follows, we discuss some methods of such forecasting.

1. Qualitative and Judgmental Method: The Delphi Method

Such methods are useful when no historical data are available or when human expertise
and knowledge are the key inputs in the forecasting. A Strategic Planning Exercise is one
example of such a situation where identifying future opportunities and threats as part of a
SWOT (Strengths, Weaknesses, Opportunities and Threats) analysis is carried out.

A popular judgmental forecasting approach is called the Delphi Method. This method
uses a panel of experts to respond to a series of questionnaires. Identities of these experts
are typically kept confidential from one another to ensure anonymity. It is usually an
iterative process of typically two to three rounds of collecting responses from the experts
through the questionnaires. After each round, responses obtained from the experts are
suitably edited to ensure anonymity and shared among the experts. This allows each to
see what the other experts think. This helps experts with similar views to reinforce their
opinions. At the same time, it also allows them to get a feel for other factors that they did
not agree upon or considered relevant while giving their opinions. In the next round, the
experts revise their estimates, and this process is repeated two to three times to promote
unbiased exchanges of ideas and discussion, and it usually results in some convergence of
opinion. This approach is usually used to forecast long range trends and impacts.

2. Statistical Forecasting Models:

Many forecasts are based on analysis of historical time series data, and are predicted on
the assumption that the future is an extrapolation of the past. In what follows, we first
discuss briefly some salient features of a time series data.

A sequence of observations {Yt, t = 1,2,…, n, ….} is called a time-series data on some

variable of interest Y varying over time t. There are usually four components responsible
for the variation in the data, a combination of some or all of which may be present in any
such time series data. These components are:

(a) Trend (Tt): This is the smooth, regular, long-term movement of the time series if
observed long enough. Some series may exhibit an upward or a downward trend, or may
remain more or less at a constant level. Again, some series, after a period of growth
(decline) may reverse its path and enter into a period of decline (growth). But sudden or
frequent changes are incompatible with the idea of a trend.

(b) Seasonal Fluctuations (St): A periodic movement in a time series, with period of
movement less than one year. For example, passenger traffic during 24 hours of a day
show such a periodic movement which recurs or repeats after regular intervals of time.

(c) Cyclical Fluctuation (Ct): An oscillatory movement in a time series, with period of
oscillation more than one year. One complete period is called a cycle. The cyclical
fluctuations are not necessarily periodic since the length of the cycle as well as the
intensity of fluctuations may vary from one cycle to another. Every businessman is
familiar with the alternating period of prosperity (or, boom) and depression in business
which follow one another in an irregular manner.

(d) Irregular Fluctuations (It): These are either wholly unaccountable or caused by such
unforeseen events as wars, floods, famines etc. This is the random component in a time
series data.

Time Series Models: There are two types:

a) Additive Model: According to this model, Yt = Tt + St + Ct + It

b) Multiplicative Model: Here Yt = Tt x St x Ct x It

While the product model is found to be appropriate for most economic time series, one
may, of course, use the simpler additive model in cases where the data are subjected to
some transformation, say logarithmic transformation.

Time Series Forecasting: Decomposition Method

Nothing can be done about the random element since it is by definition unpredictable.
However, the first three elements (trend, seasonal and cyclical), once separated out, can
be re-assembled to make a forecast. The elements are isolated one by one.

a) Determination of Trend: We use a regression method here, with time (t) as the
independent variable and the time series data (Yt) as the dependent variable. The
least square method is employed. Thus if we have to fit a linear trend Tt = a + b.t
based on n data points, then the equations for determining the parameters ‘a’ and
‘b’ are:
Yt = n.a + b t and
t t

t.Yt = a t + b t2 .
t t t

If â and b̂ are the estimates of ‘a’ and ‘b’, then the fitted trend equation is
Tt = â + b̂ t
 b) Determination of Cyclical Component: The next step is to isolate any cycle in the
data. By choosing a suitable moving average (12 points for monthly data, four for
quarterly, etc.) the random and seasonal elements can be smoothed away, leaving
just the trend and the cycle. If Mt is such a moving average, the ratio between Mt
and Tt ( = a + bt) is the cycle. If the ratio is approximately 1 for all time periods
then there is no cycle. If it differs from 1 with any regular pattern then the ratio
should be inspected (usually graphed against time) to determine the nature of the
cycle. The idea is to look for an interval of time (more than one year) such that the
ratio appears to repeat its pattern. The size of the cyclical effect is measured by
calculating the average of the ratio for each point in the cycle.

c) Determination of Seasonal Component: Seasonality is isolated by an approach

similar to that used for cycles. Here we consider the ratio between the actual value
(time series data) and the moving average value, and it contains only seasonality
and the random effect. We now average these ratios for specific seasons (months,
quarters etc.) to eliminate the random fluctuation. These seasonal effects need to
be adjusted appropriately so that their overall effect (average taken over all the
seasons) becomes 1.

d) Forecast: In making a forecast, the three isolated elements are multiplied together.
Thus

Forecast = Trend x Cycle x Seasonality

Case Study: Forecasting Sales of Washing Machines

(Ref: Analytical Decision Making, by David Targett – Macmillan India )

A company makes a variety of white consumer goods. The products are of good quality
and have been successful in terms of sales volumes. Selling costs, however, have been
unacceptable to senor management, mainly because of the stock levels. There is now a
company-wide initiative to reduce stocks, starting with washing machines whose value,
and therefore stocking cost, is high.

As part of the initiative forecasting methods have to be improved and a major preliminary
task is to forecast sales of washing machines. The data is quarterly and sales volumes for
the last nine years, 1986-94, are shown in Table 1 below.

Table 1
Qtr\Yr 1986 1987 1988 1989 1990 1991 1992 1993 1994
1 14.4 19.8 21.0 30.9 45.3 40.8 58.5 69.3 70.8
2 20.7 28.5 32.4 55.2 56.4 58.2 85.2 93.0 103.8
3 17.7 23.7 27.9 47.4 46.2 53.1 78.0 78.0 93.9
4 24.3 31.8 44.1 59.7 59.1 78.3 103.2 103.2 110.1

The company requires the quarterly sales forecasts in 1996.

a) Calculate the Trend: A regression analysis is carried out with sales volume as Y
(having values 14.4, 20.7, 17.7, 24.3, 19.8, 28.5, ….) and time as the independent
variable (with corresponding values 1, 2, 3, 4, 5, 6, …). Plotting the data on a
graph paper exhibits a linear trend. The trend analysis thus gives the trend
equation as

Tt = 8.556 + 2.516.t (origin of time t = Qtr 1 of 1986, unit of t = 1 Qtr)

b) Calculate the cyclical effect: A four-point moving average smooth out the
seasonality in the data. The first moving average is

14.4 20.7 17.7 24.3

19.27
4
and this corresponds to mid-point of periods (quarters) 2 and 3 of the first year.
We arbitrarily assume it to correspond to period 3. The next moving average is
calculated by leaving out the first value 14.4 and incorporating the fifth value in
the series, 19.8, and this now corresponds to period 4 of the first year. We
continue this way, the last moving average is for period 35 (period 3 of the last
year) and includes the last four actual data points. These are given by column (4)
of Table 2.

The trend values are also calculated from the trend equation. For example, for
1989, quarter four (time period 16), the estimated trend is

Trend = 8.556 + 2.516 x 16 = 48.78.

These are given by column (5) of Table 2.

The cyclical ratio (moving average / trend = Col(4)/Col(5)) is shown in column

(6) of Table 2. Looking at these values, it appears that there is a pattern and the
length is 12 quarters since troughs and peaks both recur at this interval. For each
of the 12 periods within the cycle, this effect is calculated by averaging over all
such periods. The first period of the cycle is taken as the first period of the time
series. Thus, for example, the cyclical effect for period 8 of the cycle is calculated
as the average of the cyclical ratios for the original periods 8, 20 and 32:

0.92 0.86 0.97

Cyclical effect for period 8 of the cycle = 0.92 .
3
The effect for all 12 periods of a cycle is thus calculated, and these are shown in
Table 3.
 Table 2
Yr/Qtr Time(t) Sales Moving Trend Cycle Seasonality
vol.(yt) Average
(1) (2) (3) (4) (5) (6) (7)
1986/ 1 1 14.4 - 11.1 - -
2 2 20.7 - 13.6 - -
3 3 17.7 19.3 16.1 1.20 0.92
4 4 24.3 20.6 18.6 1.11 1.18
1987/ 1 5 19.8 22.6 21.1 1.07 0.88
2 6 28.5 24.1 23.6 1.02 1.18
3 7 23.7 26.0 26.2 0.99 0.91
4 8 31.8 26.3 28.7 0.92 1.21
1988/ 1 9 21.0 27.2 31.2 0.87 0.77
2 10 32.4 28.3 33.7 0.84 1.15
3 11 27.9 31.4 36.2 0.87 0.89
4 12 44.1 33.8 38.7 0.87 1.30
1989/ 1 13 30.9 39.5 41.2 0.96 0.78
2 14 55.2 44.4 43.8 1.01 1.24
3 15 47.4 48.3 46.3 1.04 0.98
4 16 59.7 51.9 48.8 1.06 1.15
1990/ 1 17 45.3 52.2 51.3 1.02 0.87
2 18 56.4 51.9 53.8 0.96 1.09
3 19 46.2 51.8 56.3 0.92 0.89
4 20 59.1 50.6 58.8 0.86 1.17
1991/ 1 21 40.8 51.1 61.3 0.83 0.80
2 22 58.2 52.8 63.9 0.83 1.10
3 23 53.1 57.6 66.4 0.87 0.92
4 24 78.3 62.0 68.9 0.90 1.26
1992/ 1 25 58.5 68.8 71.4 0.96 0.85
2 26 85.2 75.0 73.9 1.01 1.14
3 27 78.0 81.2 76.4 1.06 0.96
4 28 103.2 83.9 78.9 1.06 1.23
1993/ 1 29 69.3 85.9 81.5 1.05 0.81
2 30 93.0 85.9 84.0 1.02 1.08
3 31 78.0 85.9 86.5 0.99 0.91
4 32 103.2 86.3 89.0 0.97 1.20
1994/ 1 33 70.8 89.0 91.5 0.97 0.80
2 34 103.8 94.9 94.0 0.99 1.12
3 35 93.9 94.7 96.5 0.98 0.99
4 36 110.1 - 99.1 - -

Table 3
Time Period of 1 2 3 4 5 6 7 8 9 10 11 12
cycle
Cyclical effect 0.96 1.01 1.10 1.08 1.05 1.00 0.97 0.92 0.89 0.89 0.91 0.88
c) Calculate the seasonal effect: The seasonality is the ratio of actual to moving
average, averaged for each quarter. Column (7) of Table 2 shows these ratios,
calculated as column (3)/column (4). Average of these ratios for the first quarter is

0.88 0.77 0.78 0.87 0.80 0.85 0.81 0.80

Seasonality(Qtr 1) = = 0.82.
8
Seasonality for the other three quarters are similarly obtained as

Seasonality(Qtr 2) = 1.14, Seasonality(Qtr 3) = 0.93, Seasonality(Qtr 4) = 1.21.

Unfortunately there is a problem with these basic seasonal indices because their
average is different from 1:

0.82 1.14 0.93 1.21

Average Seasonal Effect = 1.025 .
4
Consequently the seasonal indices have to be adjusted so that their average is 1.
This is done by dividing each index by 1.025. The adjusted Seasonal indices are

Qtr 1 = 0.80, Qtr 2 = 1.11, Qtr 3 = 0.91, Qtr 4 = 1.18.

The average seasonal effect is now neutral since

0.80 1.11 0.91 1.18

1.00 .
4

d) Make the Forecast: The original series has been decomposed into trend, cycle and
seasonality. To make forecasts for the four quarters of 1996, the three components
are re-assembled.

For example, for the first quarter of 1996, the value of t is 41 and hence

Trend = 8.556 + 2.516 x 41 = 111.7.

Each cycle lasts 12 periods. Starting at the first quarter of 1986, the cycles are
1986-88, 1989-91, 1992-94. Thus the four quarters of 1996 are time periods 5-8
of a cycle. The cyclical effects for these periods are taken from Table 3 as 1.05,
1.00, 0.97 and 0.92 respectively. Thus cyclical effect for Quarter 1 of 1996 is

Cyclical effect = 1.05.

The (adjusted) seasonality for Quarter 1 is obtained as 0.80.

The forecast is the product of these three elements and hence

Forecast (Qtr 1, 1996) = 111.7 x 1.05 x 0.80 = 93.8

 Table 4 calculates forecasts for all the quarters of 1996.

Table 4
Quarter Trend Cycle Seasonality Forecast
1 111.7 1.05 0.80 93.8
2 114.2 1.00 1.11 126.8
3 116.7 0.97 0.91 103.0
4 119.3 0.92 1.18 129.5

3. Causal Modeling: Use of Regression Equation

Causal modeling is a group of methods which work on historical data to analyse how
variables influence one another. It considers two aspects of these influences: Correlation
addresses the issue of how strong the influences are, while regression analysis is
concerned with discovering the mathematical forms of the influences.

For example, suppose we have 10 years’ data on quarterly sales volume of a product (Y)
and the quarterly expenditures for advertising the product (X). Correlation can use this
data to measure the extent to which sales volume appears to have been influenced by
advertising expenditure. It can show whether a high level of sales has usually
corresponded with a high advertising expenditure in the same quarter (and low with low),
that is, it can measure the strength of the relationship between Y and X.

The regression can provide the formula linking sales with advertising. The formula might
look like

Sales Volume (‘000) = 3000 + 3.1 x Advertising Expenditure (000’ $)

Once such a regression equation can be obtained from the data (using the method of least
squares), the same equation can be used for forecasting the sales volume when the level
of advertising expenditure is known.

For further details, one is referred to the section on Analysis of Bivariate Data.