FORECASTING

Actual 3-Month Weighted Moving 3-Month

Month
Demand Average Moving

January 60 Average
February 52
March 40
April 26 [(60*1)+(52*2)+(40*3)]/6 = 47 51
May 18 [(52*1)+(40*2)+(26*3)]/6 = 35 39
June 15 [(40*1)+(26*2)+(18*3)]/6 = 24 28
July 10 [(26*1)+(18*2)+(15*3)]/6 = 18 20
August 18 [(18*1)+(15*2)+(10*3)]/6 = 13 14
September 26 [(15*1)+(10*2)+(18*3)]/6 = 15 14
October 33 [(10*1)+(18*2)+(26*3)]/6 = 21 18
November 46 [(18*1)+(26*2)+(33*3)]/6 = 28 26
December 60 [(26*1)+(33*2)+(46*3)]/6 = 38 35
Table 1 : 3 Month Moving Average and weighted moving average

Weighted moving average improves results only a little. In June and July, for instance, the
weighted averages (24 and 18) are closer than simple moving averages (28 and 20) to the real demand (15
and 10). In the fall and peak months, the weighted averages also track the upward trend somewhat better than the
simple moving averages. For more beneficial forecasting of such rapidly changing trends, still more refined
approach is required for correcting the equation for seasonal effects.

Exponential smoothing
Exponential smoothing is a more refined version of the weighted moving average that is
relatively easy to use and requires very little record keeping. Exponential smoothing software is used
widely in many organizations in their demand planning.
The exponential smoothing equation requires only three basic terms:
 Last period's forecast,
 Last period's actual demand, and
 Smoothing constant

Smoothing constant is a number greater than 0 and less than 1 represented as alpha (a, Greek letter),
which is basically a percentage weighting between 1 to 100 %. There are two ways to present the same
equation, and either way produces the same answer. The first method helps call out how this forecast is
built: It starts with a naive forecast and adds (or subtracts if negative) a definite amount (weight) of the
prior period's measured forecast error:

New Forecast = Last Period's Forecast + a (Last Period's Demand – Last Period's Forecast)

Naive forecast Weight placed on Forecast error Measure of forecast error

 FORECASTING
The second method is easier to calculate and is more commonly used:

New Forecast = (a x Last Period's Demand) + [(1 – a) x Last Period's Forecast]

As you can see in the formula above, the new forecast includes the previous forecast and actual
demand from the previous period, adjusted by the constant. Thus the equation automatically produces a
weighted average of previous results. This disregards the need to maintain data from a number of
preceding periods for use in recalculating a moving average for the new period's forecast.

Consider an example. Let's say you're using a smoothing constant of 0.3 and your March demand
for Safety gloves was for 200 pairs against a forecast of 180. The April forecast, using exponential
smoothing, is calculated as follows:

April Forecast = (0.3 x 200) + [(1 – 0.3) x 180]

= (0.3 x 200) + (0.7 x 180) = 186

The smoothing constant of 0.3 ensures that your April forecast takes account of only 30 percent
(0.3) of the excess of March demand over your March forecast, placing 70 percent (0.7) of the weight on
the earlier forecast. Thus the formula "smoothes out" any random variations in demand.

What if April demand is 180 against your forecast of 186? The May forecast will again reflect 30
percent of the change, this time in a negative direction.

May Forecast = (0.3 x 180) + (0.7 x 186) = 184.2

Although the April demand of 180 is six pairs of safety gloves under your forecast of 186, your
forecast for May drops only 1.8, to 184.2 (or 184, assuming that you don't want to sell one-fifth of a pair
of safety gloves). This reflects the fact that your forecast overvalued actual demand. Your series of
forecasts-180, 186, 184 varies much less than the series of actual demand numbers. The table compares
the results of exponential smoothing forecasts with the forecasts based on weighted moving averages in
the preceding example. The smoothing constant used here is 0.4.

Smoothing can make your forecasts more responsive to trends than simple moving averages.
Like any weighted moving average, however exponential smoothing lags behind changes in the data. In
the series of forecasts in the table, for example, exponential smoothing actually does a worse job of
forecasting than does the weighted moving average, effectively smoothing out the trends as well as
random variations.
 FORECASTING
Generally speaking, firms use exponential smoothing constants that fall between 0.05 and 0.5.
The higher the constant, the more weight your forecast gives to the actual demand data from the preceding
period. A constant of .05 would give minimal weight to the preceding period. .A constant of 1.0 would
yield the same result as a naive forecast, because it would include the entire last period's demand (100
percent) and none of the last period's forecast amount (0 percent).

Using the best possible smoothing constant is critical. Like the numbers used in simple weighted
averages, the exponential smoothing constant is not a given. This will have to be determined and the same
reflects the best judgment of experts from the series of previous forecast and demand numbers, or a
combination of both. Like the constants used in weighted average, the smoothing constant can be selected
based on various constants on the historical data to view the best results.

Weighted Moving Average Exponential Smoothing

Month Demand
Forecast Forecast
January 32
February 26
March 12
April 5 20 20 (previously calculated forecast)
May 4 11 14 (0.4 x 5) + (0.6 x 20)
June 3 6 10 (0.4 x 4) + (0.6 x 14)
July 2 4 7.2 (0.4 x 3) + (0.6 x 10)
August 5 3 5.12 (0.4 x 2) + (0.6 x 7.2)
September 10 4 5.07 (0.4 x 5) + (0.6 x 5.12)
October 15 7 7.04 (0.4 x 10) + (0.6 x 5.07)
November 25 12 10.22 (0.4 x 15) + (0.6 x 7.04)
December 32 19 16.13 (0.4 x 25) + (0.6 x 10.22)

Comparison of forecast types

Below table compares the results of forecasts for one month using the naive, three-month moving
average, weighted moving average, and exponential smoothing methods.

Naive forecast (Previous month's demand)

(Average of previous 3 months)

Moving average
(975 + 1,125 + 950)+ 3 = 1,016.67 (1,017)

(Weighted average of previous 3 months)

Weighted moving average
(3 x 975+ 2 x 1,125 +1 x 950)+ (3 + 2 + 1-) = 1,020.83

(Alpha** x Last demand) +

Exponential 1 smoothing
[(1 — Alpha) x Last forecast]: (0.2 x 975) + (0.8 x 1,040) = 1,027
 FORECASTING
Modifying Quantitative Methods

Quantitative methods are usually modified to account for trends (including cycles) and seasonal
effects. As part of the forecasting process, forecasters decompose data to remove trend and seasonality and
then proceed with the forecast projection. Once the qualitative forecast of intrinsic data is completed, the
forecasters recompose the forecast by applying any trend, cycle, and seasonality adjustments to the
forecasted data. Adjustments include modifying forecasts for internal trends (such as a historical growth
rate in demand of five percent for a product) or external trends (such as a recession suppressing demand to
historically low levels).

Fundamentals of Supply Chain Forecasting:

One such process of decomposing and recomposing forecasts is demonstrated in the discussion
that follows regarding the seasonal index.

Seasonal index

Seasonal variations, such as those appearing in the gas stations demand numbers we've been
using as an example. Since the moving average always holdups behind an upward or downward trend,
forecasters have to advance other quantitative methods to take account of these seasonal movements. As
mentioned earlier, "seasonal" in this context can refer to any forecasting period, from actual seasons, to
months, to days, or even to parts of days. (Some restaurants track demand periods as small as quarters of
an hour.)

A common way of doing this is to look at past demand data for each season to determine how
much, on average, the seasonal demand departs from the average demand. For example, if you are a Sales
Executive who sees 320 customers each month on average but you typically treat 360 customers in the
average July, you have a seasonal index for July of 1.125 (360/320). If you think your average monthly
customer load is going up to 340 in the next year, then you would expect to see about 382 customers in
July of next year (340 multiplied by the seasonal index of 1.125).

Computing the seasonal index for our homeowner market segment for Gas station requires the
following calculations. We'll use three years' worth of monthly demand data.

 What is the seasonal average of monthly demand for each of the 12 months in the past
three years?
o To define this, divide the sum of the demand in the three Januarys by 3, divide the sum of
all February demand by 3, and so on, through December.
 FORECASTING

 What is the deseasonalized average monthly demand during those three years?
o Calculate this by adding the 12 monthly averages from the preceding step and dividing
the total by 12. (This monthly average is said to be "deseasonalized," because it ignores
the effect of seasonal variation. Note that using deseasonalized demand is an example of
decomposing data prior to proceeding to the quantitative forecast computation.)
 What is the seasonal (monthly) index?
o This is the average for a particular month (such as January) divided by the overall
deseasonalized monthly average demand.

Seasonal Index:

 Average Demand for Period (e.g., Month)

 Deseasonalized Average Demand for all Periods (e.g., Year)

In the Gas station example above, the deseasonalized monthly average demand for the years 2009
through 2011 is 14 stations per month and monthly demand for the market segment fluctuates throughout
the year.

With very few sold in the summer months and many sold during the winter, the actual demand in any
given month varies very far in both directions from the monthly mean. For instance in January, the
average monthly demand for the three years has been for 31 gas station. Compared to the average of 14,
this gives you a seasonal index for January of 2.214. Multiplying the decomposed forecast data by this
index is an example of the process of recomposing a forecast to include (in this case) the seasonality.

How could the partners in the supply chain use this index? Let's say the retailer, the distributor, and
the manufacturer get together to plan for 2012. They might begin by scheming a monthly average demand
for 15 gas station. (There is a slight upward trend for the three years, with the average for 2011 being
about 14.25 units per month.)

They could then plan for January demand of 15 units multiplied by the January index of 2.214, which
yields a (rounded) forecast of 33 units in the demand plan.

After combining this seasonalized forecast with the deseasonalized forecasts for other market
segments such as home builders (who place most of their orders in the summer months), they could plan
their staffing, warehousing, shipping, and sums to arrive at 330. If the forecast demand were 110, 120, and
130, the cumulative forecast demand is 360. Calculating bias can use a variation on the forecast error
calculation, but it doesn't use absolute amounts as the positive or negative sign can show the direction of
the bias:
 FORECASTING
Cumulative Forecast Error = Cumulative Actual Demand — Cumulative Forecast Demand

= 330 — 360 = —30

Any answer that does not result in zero reflects a bias. The size of the number reflects the relative
amount of bias that is present. A negative result shows that actual demand was consistently less than the
forecast, while a positive result shows that actual demand was greater than forecast demand.

Bias could be the result of a temporary situation or an unaccounted change in the trend or
seasonal effect. Tracking the circumstances surrounding each bias can help distinguish between the two.

Random variation:

In terms of measuring errors, random variation is an amount of variation in which the cumulative
actual demand equals the cumulative forecast demand.

For example, if actual demand is 100 for three periods, cumulative actual demand is 300 and if
forecast demand is 90, 110, and 100, the cumulative forecast demand is 300. Because of the zero net
difference, the error over this period can be said to be the result of random variation. Note that wide
swings in either direction that just happen to balance out would still be difficult to plan around.

Mean Absolute Period Deviation:

A common way of tracking the extent of forecast error is to add the absolute period errors for a
series of periods and divide by the number of periods. This gives you the mean absolute deviation (MAD).
Note: In the formula below, the Greek uppercase letter stands for "the sum of."

An alternative is to calculate absolute -deviations of actual sales data minus forecast data. These
data can be averaged in the usual arithmetic way or with exponential smoothing.

With absolute values, the forecast falls short of demand or exceeds demand doesn't matter; only
the magnitude of the deviation counts in MAD. We can understand how this works by looking at the gas
station demand and forecasts used previously and illustrated below.
 FORECASTING

Month Demand Exponential Error (Deviation) Absolute

Forecast with Deviation

January 32 0.4 Smoothing

February 26
March 12
April 5
May 4 14.00 -10.00 (4 - 14) 10.00
June 3 10.00 -7.00 (3 -10) 7.00
July 2 7.20 -5.20 (2 - 7.2) 5.20
August 5 5.12 -0.12 (5 - 5.12) 0.12
September 10 5.07 +4.93 (10 - 5.07) 4.93
October 15 7.04 +7.96 (15 - 7.04) 7.96
November 25 10.22 +14.78 (25 - 10.22) 14.78
December 32 16.13 +15.87 (32 - 16.13) 15.87
Total Deviation 65.86
MAD (65.8618 = 8.23) 8.23
MAD of 8.23 units implies that forecasts are off on average for the review period by about plus or
minus 8.23 units.

Standard Deviation
Mean squared error (MSE)
MSE and MAD comparison
In addition to MAD, another way to calculate forecast error would be to
use standard deviation, which is commonly provided in most software
programs. An Approximation for standard deviation when you know
the MAD follows:
Standard Deviation (approximate) = MAD x 1.25
Another method of calculating error rates, the mean squared error
(MSE), magnifies the errors by squaring each one before adding them
up and dividing by the number of forecast periods. Squaring errors
effectively makes them absolute since multiplying two negative
numbers always results in a positive number. The formula for mean
squared error follows:
Sum of (error of each period) 2 MSE – Number of Forecast periods
Note that the process of squaring each error gives a wider range of
numbers much larger numbers on one hand (251.86 as opposed to 15.87
in December on the high side and 0.01 instead of 0.12 in August on the
low side). The greater range gives you a more complex measure of the
error rate, which is especially useful if the absolute error numbers are
relatively close together and reduction of errors is important.
 FORECASTING
Exponential
Absolute Squared
Month Demand Forecast with Error (deviation)
Deviation Errors
0.4 Smoothing
January 32
Februa 26
March 12
April 5

May 4 14 -10.00 (4 -14) 10 100.00

June 3 10 -7.00 (3 -10) 7 49.00

July 2 7.2 -5.20 (2 - 7.2) 5.2 27.04

August 5 5.12 -0.12 (5 - 5.12) 0.12 0.01
September 10 5.07 +4.93 (10 - 5.07) 4.93 24.30
October 15 7.04 +7.96 (15 - 7.04) 7.96 63.36
November 25 10.22 +14.78 (25 - 10.22) 14.78 218.45
December 32 16.13 +15.87 (32 - 16.13) 15.87 251.86
Total Deviation 65.86
MAD (65.86/8 = 8.23) 8.23
MSE (734.02/8 = 91.75) 91.75
Measuring the extent of deviation helps determine the need to improve forecasting or rely on
safety stock to meet customer service objectives.

There is a drawback to the MAD calculation in that it is an absolute number

that is not meaningful unless compared to the forecast. Mean absolute
Mean absolute percentage
percentage error (MAPE) is a useful variant of the MAD calculation because it
error (MAPE)
shows the ratio, or percentage, of the absolute errors to the actual demand for a
given number of periods.

Continuing the previous example, Table below shows how the absolute percent error (APE) is
first determined for each period by taking the absolute error divided by the actual demand (A). Then, the
sum of the APE (percent) for periods 1 through 8, which is 956.8 in the example, is divided by the number
of periods, which is 8 in the example, to calculate the MAPE. On average, MAPE is 119.6 percent.
 FORECASTING

Exponential
Demand Absolute
Month Forecast with 0.4 Error (Deviation) APE
(A) Deviation
Smoothing

January 32
February 26
March 12
April 5
May 4 14.00 -10.00 (4 -14) 10.00 250.00
June 3 10.00 -7.00 (3 - 10) 7.00 233.33
July 2 7.20 -5.20 (2 - 7.2) 5.20 260.00
August 5 5.12 -0.12 (5 - 5.12) 0.12 2.40
September 10 5.07 +4.93 (10 - 5.07) 4.93 49.30
October 15 7.04 +7.96 (15 - 7.04) 7.96 53.06
November 25 10.22 +14.78 (25 - 10.22) 14.78 59.12
December 32 16.13 +15.87 (32 - 16.13) 15.87 49.59
Total 65.86 956.80
MAD (65.86/8 = 8.23) 8.23
MAPE (956.8/8 = 119.6) 119.60