Chapter 4 Exponential Smoothing

4.1 Simple exponential smoothing, nonseasonal data 25

(SIMPLE)

4.2 Simple exponential smoothing with seasonality 31

(SIMPLEADDSEAS and SIMPLEMULSEAS)

4.3 Trend smoothing, nonseasonal data 34

(TREND)

4.4 Trend smoothing with seasonality 38

(TRENDADDSEAS and TRENDMULSEAS)

Simple exponential smoothing is a short-range forecasting tool that assumes a

reasonably stable mean in the data with no trend (consistent growth or decline).
If the data are nonseasonal, use the SIMPLE model. If seasonal, you have two
choices: SIMPLEADDSEAS for additive seasonality and SIMPLEMULSEAS for
multiplicative seasonality. The seasonal models perform seasonal adjustment,
forecast the adjusted data with simple smoothing, then reverse the seasonal
adjustment in the forecasts.

To forecast a trend in nonseasonal data, try the TREND model, which allows you
to compare several different types of trend before committing to a forecast. If the
data contain both a trend and seasonality, again there are two choices, depending
on the type of seasonality: TRENDADDSEAS and TRENDMULSEAS.

24 Forecasting
4.1 Simple exponential smoothing, nonseasonal data (SIMPLE)

Exponential smoothing models are ubiquitous in business forecasting, especially in budgeting

and inventory control applications. Such models are relatively simple, easy to understand, and
easy to implement. Smoothing models also compare quite favorably in accuracy to complex
forecasting models. One of the surprising things scientists have learned about forecasting in
recent years is that complex models are not necessarily more accurate than simple models.

The simplest form of exponential smoothing is called, appropriately enough, simple smoothing.
Simple smoothing is used for short-range forecasting, usually just one month into the future.
The model assumes that the data fluctuate around a reasonably stable mean (no trend or
consistent pattern of growth).

Figure 4-1 illustrates an application of simple exponential smoothing at the International Airport
in Victoria, Texas. The airport has been open for a year and the data are the monthly numbers of
passengers embarked. There is not yet enough data to estimate a seasonal pattern, so we will use
SIMPLE, the nonseasonal version of simple exponential smoothing.

To get the forecasting process started, SIMPLE automatically sets the first forecast (E26) equal
to the average of the number of warm-up data specified in cell D9. The number of warm-up data
is 6, so the first forecast of 30.0 is the average of the data for months 1-6. If you don't like the
first forecast, replace the formula in E26 with a value. Thereafter the forecasts are updated as
follows: In column F, each forecast error is equal to actual data minus the forecast for that
period. In column E, each forecast is equal to the previous forecast plus a fraction of the
previous error. This fraction is found in cell D8 and is called the smoothing weight. The model
works much like an automatic pilot, a cruise control on an automobile, or a thermostat. If a
given forecast is too low, the forecast error is positive, and the next forecast is increased by a
fraction of the error. If a given forecast is too high, the forecast error is negative, and the next
forecast is reduced by a fraction of the error. If we get lucky and a forecast is perfect, the error is
zero and there is no change in the next forecast.

A total of 12 data observations are entered in Figure 4-1. The model automatically makes
forecasts through the last period specified in cell D10. For months 13-24, the forecasts are
constant as shown in Figure 2-2. Remember that the model assumes no trend, so the only option
is to project the last forecast for every period in the future.

The model computes two mean forecast error measures. The MSE is the mean-squared-error and
the MAD is the mean of the absolute errors (called the mean absolute deviation). Both are
commonly used in practice. The MSE gives more weight to large errors, while the MAD is
easier to interpret.

Forecasting 25
Figure 4-1
A B C D E F G H I J K L M
1 SIMPLE.XLS INTERMEDIATE CALCULATIONS: DATA TABLE:
2 Simple exponential smoothing. RMSE (Square root of MSE) 3.91 Select K6..L16.
3 3 X RMSE 11.74 Select Data Table.
4 Enter column input cell = D8.
5 Warm-up SSE 91.81 Weight MSE
6 Warm-up MSE 15.30 11.87
7 INPUT: Forecasting SSE 163.05 0.10 11.41
8 Smoothing weight 0.30 Fcst SSE - Warm-up SSE 71.24 0.20 11.58
9 Nbr. Of warm-up data 6 Nbr. of forecast periods 6 0.30 11.87
10 Last period to forecast 24 Forecasting MSE 11.87 0.40 12.24
11 0.50 12.74
12 OUTPUT: Warm-up Sum abs. err. 22.07 0.60 13.41
13 Number of data 12 Warm-up MAD 3.68 0.70 14.29
14 Nbr. of outliers 0 Forecasting Sum abs. err. 40.76 0.80 15.39
15 Warm-up MSE 15.30 Fcst Sum abs. - Warm-up Sum 18.69 0.90 16.73
16 Forecasting MSE 11.87 Nbr. of forecast periods 6 1.00 18.33
17 Warm-up MAD 3.68 Forecasting MAD 3.11 Note: MSE values are based
18 Forecasting MAD 3.11 Last forecast at period 24 on forecasting periods.
19
20
21 Sum
22 Avg. absolute
23 Period data errors
24 Month Month Period Actual Outlier Sum
25 & year nbr. nbr. data Fcst Error Indicator nbr out
26 Jan-00 1 1 28 30.00 -2.00 0 0 1 28.00 2
27 Feb-00 2 2 27 29.40 -2.40 0 0 2 27.50 4.4
28 Mar-00 3 3 33 28.68 4.32 0 0 3 29.33 8.72
29 Apr-00 4 4 25 29.98 -4.98 0 0 4 28.25 13.696
30 May-00 5 5 34 28.48 5.52 0 0 5 29.40 19.2128
31 Jun-00 6 6 33 30.14 2.86 0 0 6 30.00 22.07456
32 Jul-00 7 7 35 31.00 4.00 0 0 7 30.71 26.07779
33 Aug-00 8 8 30 32.20 -2.20 0 0 8 30.63 28.27553
34 Sep-00 9 9 33 31.54 1.46 0 0 9 30.89 29.73711
35 Oct-00 10 10 35 31.98 3.02 0 0 10 31.30 32.76022
36 Nov-00 11 11 27 32.88 -5.88 0 0 11 30.91 38.64405
37 Dec-00 12 12 29 31.12 -2.12 0 0 12 30.75 40.76272
38 Jan-01 13 #N/A 30.48 #N/A 0 0 13 #N/A 0
39 Feb-01 14 #N/A 30.48 #N/A 0 0 14 #N/A 0
40 Mar-01 15 #N/A 30.48 #N/A 0 0 15 #N/A 0

26 Forecasting
 24
26
28
30
32
34
36
Jan-00

Feb-00
Figure 4-2

Mar-00

Apr-00

May-00

Jun-00

Jul-00

Aug-00

Sep-00

Oct-00

Forecasting
Nov-00
Simple Smoothing

Dec-00

Jan-01

Feb-01

Mar-01

Apr-01
Actual
Forecast

May-01

Jun-01

Jul-01

27
Both MSE and MAD are computed for two samples of the data. The first sample (periods 1-6) is
called the warm-up sample. This sample is used to "fit" the forecasting model, that is to get the
model started by computing the first forecast and running for a while to get "warmed up." The
second part of the data (periods 7-12) is used to test the model and is called the forecasting
sample. Accuracy in the warm-up sample is really irrelevant. Accuracy in the forecasting
sample is more important because the pattern of the data often changes over time. The
forecasting sample is used to evaluate how well the model tracks such changes. There are no
statistical rules on where to divide the data into warm-up and forecasting samples. There may
not be enough data to have two samples. A good rule of thumb is to put at least six nonseasonal
data points or two complete seasons of seasonal data in the warm-up. If there is less data than
this, there is no need to bother with two samples. In a long time series, it is common in practice
to simply divide the data in half. If you don't want to bother with a warm-up sample, set the
number of warm-up data equal to the total number of data. The forecasting MSE and MAD will
then be set to zero.

How do you choose the weight in cell D8? A range of trial values must be tested. The
best-fitting weight is the one that gives the best MSE or MAD in the warm-up sample. There are
two factors that interact to determine the best-fitting weight. One is the amount of noise or
randomness in the series. The greater the noise, the smaller the weight must be to avoid
overreaction to purely random fluctuations in the time series. The second factor is the stability
of the mean. If the mean is relatively constant, the weight must be small. If the mean is
changing, the weight must be large to keep up with the changes. Weights can be selected from
the range 0 - 1 although we recommend a minimum weight of 0.3 in practice. Smaller values
result in a very sluggish response to changes in the mean of the time series. In practice,
smoothing weights in the range of 0.30 to 0.50 are commonly used. Weights in this range offer a
reasonable compromise between sensitivity to noise and responsiveness.

An Excel data table is available in columns K and L to assist in selecting smoothing weights.
Column K displays smoothing weights from 0.10 to 1.00 in increments of 0.10 while column L
displays the corresponding MSE for the forecasting sample. Follow the instructions at the top of
the data table to update MSE values. The weights in column K can be changed. You can also
edit the formula in L6 to compute MAD rather than MSE results. The formula in Cell L6 is
“=I10”. If you want MAD results, change this to “=I17”.

Two other graphs in the SIMPLE workbook assist in evaluation of the forecast model. The error
graph (Figure 4-3) compares individual forecast errors to control limits. These limits are
established at plus and minus three standard deviations from zero. The standard deviation is an
estimate based on the square root of the MSE, called the RMSE for root-mean-squared-error.
The probability is less than 1% that individual errors will exceed the control limits if the mean of
the data is unchanged. The “outlier” count in cell D14 of Figure 4-1 is the number of errors that
went outside control limits. Finally, the MSE graph (Figure 4-4) is a bar chart of MSE values for
alternative smoothing weights. Since the data are not changing, the best weight is relatively
small.

28 Forecasting
Figure 4-3

15 Control chart for forecast errors

10

-5

-10

-15
Apr-00

Aug-00

Apr-01
May-00

Nov-00

May-01
Oct-00
Feb-00

Mar-00

Jun-00

Jul-00

Sep-00

Dec-00

Feb-01

Mar-01

Jun-01

Jul-01
Jan-00

Jan-01

Figure 4-4

SM OO THING W EIGHT VS. FO RECASTING M SE

20.0

18.0

16.0

14.0
FORECASTING MSE

12.0

10.0

8.0

6.0

4.0

2.0

0.0
0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
SM OO THING W EIG HT

Forecasting 29
The forecasting model in SIMPLE is based on two equations that are updated at the end of each
time period:

Forecast error = actual data - current forecast

Next forecast = current forecast + (weight x error)

A little algebra shows that this model is equivalent to another model found in many textbooks
and in practice:

Next forecast = (weight x actual data) + [(1 - weight) x current forecast]

The mathematical formulation of the simple smoothing model is:

^
et = Xt - Xt-1(1) (4-1)

St = St-1 + h1et (4-2)

^
Xt(m) = St (4-3)

We compute the forecast error et, defined as actual data Xt minus the forecast. The letter t refers
to the time period. The current period, for example the month, is t, while last period is
designated t-1.
^
Xt-1(1) is the forecast made at the end of t-1 for 1 step ahead, that is the forecast made last month
for this month. St is the new level or the weighted average of the time series at the end of t. The
new level is equal to the level at the end of t-1 plus a fraction of the error. The fraction h1 is the
smoothing weight or parameter. Finally, in equation (4-3), the forecast at the end of t for m (any
number) of steps ahead is equal to the smoothed level at the end of t.

Many companies forecast with moving averages. It may be surprising that simple exponential
smoothing is equivalent to a weighted moving average in which the weights decline
exponentially with time. Here are the effective weights assigned by the model to past data, with
the smoothing weight denoted by h1:

Most recent data observation h1

Data in previous period h1(1- h1)
Data two periods ago h1(1- h1)2
Data three periods ago h1(1- h1)3

In the Victoria data, the weight assigned to December, 2000, is 0.30, the weight for November,
2000, is 0.30(1-0.30) = 0.21, and so on. Every data observation gets some weight although the
weights decline rapidly with time.

30 Forecasting
4.2 Simple exponential smoothing with seasonality (SIMPLEADDSEAS and
SIMPLEMULSEAS)

If the data are seasonal, with no trend, use SIMPLEADDSEAS for additive seasonality or
SIMPLEMULSEAS for multiplicative seasonality. In many cases, the type of seasonality is not
obvious, even after testing, so it may be wise to experiment with both models before committing
to a forecast.

SIMPLEADDSEAS is a combination of the ADDITMON and SIMPLE models. Data entry is

done in the first tab, named Additive (identical to ADDITMON and not shown here).
Seasonally-adjusted data from the first tab is automatically used in the second tab, named
Simple, shown in Figure 4-5. The data are highly seasonal sales of chlorine tablets at a supply
store for home swimming pools. Column D repeats seasonally-adjusted data from the Additive
tab, while column E repeats the seasonal indices. In column F, simple smoothing is applied to
the seasonally-adjusted data. Errors in column G are equal to seasonally-adjusted data minus
seasonally-adjusted forecasts. In column H, seasonally-adjusted forecasts are converted to final
forecasts by adding back the appropriate seasonal index—that is, seasonal adjustment in the
forecasts is reversed.

Two forecast graphs are included in the SIMPLEADDSEAS workbook (Figures 4-6 and 4-7).
The first graph shows seasonally-adjusted data and seasonally-adjusted forecasts, while the
second shows original data and final forecasts. The graphs demonstrate that seasonal adjustment
made an enormous reduction in variance in the chlorine-tablet data. The workbook also includes
an error graph (not shown here) like the SIMPLE model. Which errors are plotted, seasonally-
adjusted or errors based on the final forecasts? The answer is both. Seasonally-adjusted errors
and errors based on the final forecasts are identical.

SIMPLEMULSEAS is a combination of the MULTIMON and SIMPLE models and is not

shown here. All calculations and worksheet operations are similar to SIMPLEADDSEAS except
that seasonality is multiplicative.

Forecasting 31
Figure 4-5
A B C D E F G H I J K L M
1 SIMPLEADDSEAS INTERMEDIATE CALCULATIONS: DATA TABLE:
2 Simple smoothing with additive seasonality RMSE (Square root of MSE) 10.87 Select L6..M16.
3 Note: data entry is done in the additive tab. 3 X RMSE 32.61 Select Data Table.
4 Enter column input cell
5 Warm-up SSE 1417.94 Weight MSE
6 Warm-up MSE 118.16 55.93
7 INPUT: Forecasting SSE 2760.19 0.10 58.89
8 Smoothing weight 0.30 Fcst SSE - Warm-up SSE 1342.25 0.20 55.82
9 Nbr. Of warm-up data 12 Nbr. of forecast periods 24 0.30 55.93
10 Last period to forecast 48 Forecasting MSE 55.93 0.40 57.83
11 0.50 60.87
12 OUTPUT: Warm-up Sum abs. err. 113.10 0.60 64.75
13 Number of data 36 Warm-up MAD 9.42 0.70 69.36
14 Nbr. of outliers 0 Forecasting Sum abs. err. 264.54 0.80 74.71
15 Warm-up MSE 118.16 Fcst Sum abs. - Warm-up Sum abs. 151.45 0.90 80.89
16 Forecasting MSE 55.93 Nbr. of forecast periods 24 1.00 88.11
17 Warm-up MAD 9.42 Forecasting MAD 6.31 Note: MSE values are
18 Forecasting MAD 6.31 Last forecast at period 48 on forecasting periods.
19
20
21
22 Avg.
23 Period data
24 Month Month Period S. Adjuste Seas S. Adjusted Final Original Outlier Sum
25 & year nbr. nbr. Data Index Fcst Error Fcst Data Indicator nbr out
26 Jan-99 1 1 87.11 -55.11 79.75 7.36 24.64 32.00 0 0 1 87.11
27 Feb-99 2 2 79.48 -55.48 81.96 -2.48 26.48 24.00 0 0 2 83.29
28 Mar-99 3 3 65.65 -20.65 81.21 -15.57 60.57 45.00 0 0 3 77.41
29 Apr-99 4 4 63.94 -3.94 76.54 -12.61 72.61 60.00 0 0 4 74.04
30 May-99 5 5 90.48 -36.48 72.76 17.72 36.28 54.00 0 0 5 77.33
31 Jun-99 6 6 71.23 78.77 78.08 -6.85 156.85 150.00 0 0 6 76.31
32 Jul-99 7 7 93.47 80.53 76.02 17.44 156.56 174.00 0 0 7 78.76
33 Aug-99 8 8 91.02 112.98 81.26 9.77 194.23 204.00 0 0 8 80.30
34 Sep-99 9 9 72.61 -17.61 84.19 -11.58 66.58 55.00 0 0 9 79.44
35 Oct-99 10 10 76.06 -9.06 80.71 -4.65 71.65 67.00 0 0 10 79.10
36 Nov-99 11 11 80.11 -22.11 79.32 0.79 57.21 58.00 0 0 11 79.19
37 Dec-99 12 12 85.86 -51.86 79.55 6.30 27.70 34.00 0 0 12 79.75
38 Jan-00 13 13 84.11 -55.11 81.44 2.66 26.34 29.00 0 0 13 80.09
39 Feb-00 14 14 87.48 -55.48 82.24 5.24 26.76 32.00 0 0 14 80.61
40 Mar-00 15 15 79.65 -20.65 83.81 -4.17 63.17 59.00 0 0 15 80.55

32 Forecasting
 0
50
100
150
200
250

0
50
100
150
200
250
Jan-99 Jan-99

Figure 4-7
Figure 4-6

Apr-99 Apr-99

Jul-99 Jul-99
Forecast

Forecast
Oct-99 Oct-99

Original data
Jan-00 Jan-00
Seasonally-adjusted data

Apr-00 Apr-00

Jul-00 Jul-00

Oct-00 Oct-00

Jan-01 Jan-01

Forecasting
Simple smoothing
Apr-01 Apr-01
Simple smoothing

Jul-01 Jul-01

Oct-01 Oct-01

Jan-02 Jan-02

Apr-02 Apr-02

Jul-02 Jul-02

Oct-02 Oct-02

33
4.3 Trend smoothing, nonseasonal data (TREND)
Exponential smoothing with a trend works much like simple smoothing except that two
components must be updated each period: level and trend. The level is a smoothed estimate of
the value of the data at the end of each period. The trend is a smoothed estimate of average
growth at the end of each period.

To explain this type of forecasting, let's review an application at Alief Precision Arms, a
company that manufactures high-quality replicas of the Colt Single-Action Army revolver and
other revolvers from the nineteenth century. Alief was founded in 1987 and, as shown in Figure
4-8, experienced rapid growth through about 1994. Since 1994, growth has slowed and this is
reflected in the forecasts.

Figure 4-8

Trend smoothing
50

Actual
45 Forecast

40

35

30

25

20
1987

1990

1993

1996

1999

2002

2005

2008

34 Forecasting
TREND in Figure 4-9 computes several different types of trend forecasts. This worksheet can
produce a linear or straight-line trend, a damped trend in which the amount of growth declines
each period in the future (as in Figure 4-8), or an exponential trend in which the amount of
growth increases each period in the future.

To get started, initial values for level and trend are computed in cells G22 and H22. The model
sets the initial trend equal to the average of the first four differences among the data. These
differences are (23.1 - 20.8), (27.2 - 23.1), (32.3 - 27.2), and (34.4 - 32.3). The average
difference or initial trend is 3.4. This value is our estimate of the average growth per period at
the beginning of the data. The initial level is the first data observation minus the initial trend or
20.8 – 3.4 = 17.4.

The forecasting system works as follows:

Forecast error = Actual data - current forecast
Current level = Current forecast + (level weight x error)
Current trend = (Trend modifier x previous trend) + (trend weight x error)
Next forecast = Current level + (trend modifier x current trend)

Level and trend are independent components of the forecasting model and require separate
smoothing weights. Experience shows that the level weight is usually much larger than the trend
weight. Typical level weights range anywhere from 0.10 to 0.90, while trend weights are usually
small, in the range of 0.05 to 0.20. The trend modifier is usually in the range 0.70 to 1.00. If the
trend modifier is less than 1.00, the effect is to reduce the amount of growth extrapolated into the
future. If the modifier equals 1.00, we have a linear trend with a constant amount of growth each
period in the future. If the modifier exceeds 1.00, growth accelerates, a dangerous assumption
more than a few time periods into the future.

Let’s work through the computations at the end of 1987. The forecast error in 1987 is data
minus forecast or 20.80 – 20.26 = 0.54. The current level is the forecast for 1987 plus the level
weight times the error, or 20.26 + 0.46 x 0.54 = 20.51. The current trend is the trend modifier
times the previous trend plus the trend weight times the error, or 0.84 x 3.40 + 0.10 x 0.54 =
2.92. The forecast for 1988 is the current level plus the trend modifier times the current trend or
20.51 + 0.84 x 2.92 = 22.96.

Now look at the forecasts for more than one period ahead. Let n be the number of periods ahead.
To forecast more than one period into the future, the formula is:

Forecast for n > 1 = (previous forecast) + [(trend modifier)n] x (final computed trend estimate)

Let's forecast the years 2000 - 2003, or 2 - 4 years into the future. The previous forecast needed
to get started is the 1999 forecast of 45.20. The final computed trend estimate was 0.83 at the
end of 1998 (it is a coincidence that the final trend value and trend modifier are so close). The
forecasts are:

Forecasting 35
Figure 4-9

A B C D E F G H I J K L M
1 TREND INTERMEDIATE CALCULATIONS:
2 Exponential smoothing with damped trend RMSE (Square root of MSE) 2.45
3 3 X RMSE 7.34
4 Avg. of first 4 differences 3.4
5 Warm-up SSE 35.92
6 Warm-up MSE 5.986613
7 Forecasting SSE 38.64
8 Fcst SSE - Warm-up SSE 2.72
9 Nbr. of forecast periods 6
10 INPUT: OUTPUT: Forecasting MSE 0.45
11 Level weight 0.46 Number of data 12 Warm-up Sum abs. err. 12.11
12 Trend weight 0.10 Number of outliers 0 Warm-up MAD 2.01853
13 Trend modifier 0.84 Warm-up MSE 5.99 Forecasting Sum abs. err. 14.63
14 Number of warm-up data 6 Forecasting MSE 0.45 Fcst Sum abs. - Warm-up Sum a 2.52
15 Last period to forecast 24 Warm-up MAD 2.02 Nbr. of forecast periods 6
16 Forecasting MAD 0.42 Forecasting MAD 0.42
17 Last forecast at period 24
18
19
20 Month Month Period Outlier Sum Period
21 & year nbr. nbr. Data Fcst Error Level Trend Indicator nbr out nbr.
22 17.40 3.40
23 1987 1 1 20.8 20.26 0.54 20.51 2.92 0 0 1
24 1988 2 2 23.1 22.96 0.14 23.03 2.47 0 0 2
25 1989 3 3 27.2 25.10 2.10 26.07 2.29 0 0 3
26 1990 4 4 32.3 28.00 4.30 29.99 2.37 0 0 4
27 1991 5 5 34.4 31.98 2.42 33.10 2.24 0 0 5
28 1992 6 6 37.6 34.98 2.62 36.19 2.15 0 0 6
29 1993 7 7 38.0 38.00 0.00 38.00 1.81 0 0 7
30 1994 8 8 41.0 39.52 1.48 40.21 1.67 0 0 8
31 1995 9 9 41.6 41.61 -0.01 41.61 1.41 0 0 9
32 1996 10 10 42.2 42.79 -0.59 42.52 1.12 0 0 10
33 1997 11 11 43.9 43.47 0.43 43.67 0.99 0 0 11
34 1998 12 12 44.5 44.50 0.00 44.50 0.83 0 0 12
35 1999 13 #N/A 45.20 #N/A 44.50 0.83 #N/A #N/A 13
36 2000 14 #N/A 45.79 #N/A 44.50 0.83 #N/A #N/A 14
37 2001 15 #N/A 46.29 #N/A 44.50 0.83 #N/A #N/A 15
38 2002 16 #N/A 46.71 #N/A 44.50 0.83 #N/A #N/A 16
39 2003 17 #N/A 47.06 #N/A 44.50 0.83 #N/A #N/A 17
40 2004 18 #N/A 47.36 #N/A 44.50 0.83 #N/A #N/A 18

36 Forecasting
 Forecast for 2 years ahead (2000) = 45.20 + .842 x 0.83 = 45.79
Forecast for 3 years ahead (2001) = 45.79 + .843 x 0.83 = 46.29
Forecast for 4 years ahead (2002) = 46.29 + .844 x 0.83 = 46.71
Forecast for 5 years ahead (2003) = 46.71 + .845 x 0.83 = 47.06

The trend modifier is a fractional number. Raising a fractional number to a power produces
smaller numbers as we move farther into the future. The result is called a damped trend because
the amount of trend added to each new forecast declines. This makes sense because growth
cannot continue indefinitely. The damped trend was selected by Alief management because it
reflects slowing growth, probably the best that can be expected given political and economic
conditions in the firearms market at the end of 1998.

By changing the trend modifier, you can produce different kinds of trend. A modifier equal to
1.0 yields a linear trend, where the amount of growth in the forecasts is constant beyond the end
of the data. In the example, change the trend modifier to 1.0 to produce a graph showing growth
that runs well above the last few data observations. Optimists can also use a modifier greater
than 1.0 to produce an exponential trend in which the amount of growth gets larger each time
period.

TREND requires that you choose the best combination of three parameters: level weight, trend
weight, and trend modifier. There are numerous ways to do this in Excel. The quickest is to use
the Excel Solver. Suppose that we want to choose parameters that minimize the forecasting
MAD in the Alief data. Under the Tools menu, select Solver. In the menu box, set the target cell
to $H$15 (the forecasting MAD). In the “Equal To” row, select Min. In the entry area for “By
Changing Cells”, enter $D$11..$D$13 (the parameters). Finally, click on Solve. The result is as
shown in Figure 4-6. A word of caution is in order here. Sometimes the Solver returns extreme
parameters, such as a level or trend weight of zero. We recommend that limits be imposed on
the parameters: 0.10-0.90 for the level weight, 0.05-0.20 for the trend weight, and 0.70-1.00 for
the trend modifier.

Another approach is to set the trend modifier equal to 1.0. and use the data table starting at the
top of column Q in TREND to find the best combination of level and trend parameters, the
combination that minimizes the MSE or MAD. Search over the range 0.10-0.90 for the level
weight in increments of 0.10. Search over the range 0.05-0.20 for the trend weight in increments
of 0.05. Then fix the level and trend parameters and try alternative values of the trend modifier
in the range 0.70 to 1.00 in increments of 0.05. Once you find the best trend modifier, run the
data table again, then do another search for the trend modifier. Keep going until the forecasting
MSE or MAD stabilizes. Great precision is not necessary. TREND is a robust model, relatively
insensitive to smoothing parameters provided that they are approximately correct.

Forecasting 37
 The mathematical formulation of the model in TREND, originally developed by Gardner and
McKenzie,1 is:
^
et = Xt - Xt-1(1) (4-4)
^
St = Xt-1(1) + h1et (4-5)

Tt = φTt-1 + h2et (4-6)

^ m
Xt(m) = St + φi Tt (4-7)
i=1

Compared to the simple model, there is one additional smoothing parameter, h2, for the trend,
denoted Tt. The trend modification parameter is φ. In the trend equation (4-6), notice that the
old trend is multiplied by φ, while in the forecast equation (4-7), the new trend is multiplied by
φ i.

4.4 Trend smoothing with seasonality (TRENDADDSEAS and

TRENDMULSEAS)
If the data are seasonal and trending, use TRENDADDSEAS for additive seasonality or
TRENDMULSEAS for multiplicative seasonality. TRENDADDSEAS is a combination of the
ADDITMON and TREND models. Data entry is done in the first tab, named Additive (identical
to ADDITMON and not shown here). Seasonally-adjusted data from the first tab is
automatically used in the second tab, named Trend, shown in Figure 4-10. The data are the
Houston-London yields discussed in Chapter 2. Column D repeats seasonally-adjusted data from
the Additive tab, while column E repeats the seasonal indices. In columns F-I, trend smoothing
is applied to the seasonally-adjusted data. Errors in column G are equal to seasonally-adjusted
data minus seasonally-adjusted forecasts. In column J, seasonally-adjusted forecasts are
converted to final forecasts by adding back the appropriate seasonal index—that is, seasonal
adjustment in the forecasts is reversed.

Two forecast graphs are included in the TRENDADDSEAS workbook (Figures 4-11 and 4-12).
The first graph shows seasonally-adjusted data and seasonally-adjusted forecasts, while the
second shows original data and final forecasts.

TRENDMULSEAS is a combination of the MULTIMON and TREND models and is not shown
here. All calculations and worksheet operations are similar to TRENDADDSEAS except that
seasonality is multiplicative.

1
Everette S. Gardner, Jr. and E. McKenzie, "Forecasting Trends in Time Series," Management Science, Vol. 31, No. 10
(October, 1985), pp. 1237-1246.

38 Forecasting
Figure 4-10

A B C D E F G H I J K L M
1 TRENDADDSEAS INTERMEDIATE CALCULATIONS:
2 Trend smoothing with damped trend and additive seasonality. RMSE (Square root of MSE) 0.00235997
3 Note: data entry is done in the additive tab. 3 X RMSE 0.00707992
4 Avg. of first 4 differences -0.00029177
5 Warm-up SSE 0.00016708
6 Warm-up MSE 0.00000557
7 Forecasting SSE 0.00034162
8 Fcst SSE - Warm-up SSE 0.00017454
9 Nbr. of forecast periods 30
10 INPUT: OUTPUT: Forecasting MSE 0.00000582
11 Level weight 0.30 Number of data 60 Warm-up Sum abs. err. 0.05553602
12 Trend weight 0.05 Number of outliers 0 Warm-up MAD 0.00185120
13 Trend modifier 0.90 Warm-up MSE 0.00000557 Forecasting Sum abs. err. 0.11603014
14 Number of warm-up data 30 Forecasting MSE 0.00000582 Fcst Sum abs. - Warm-up Sum ab 0.06049411
15 Last period to forecast 72 Warm-up MAD 0.00185120 Nbr. of forecast periods 30
16 Forecasting MAD 0.00201647 Forecasting MAD 0.00201647
17 Last forecast at period 72
18
19
20 Month Month Period S. Adjuste Seas S. Adjusted Final Original Outlier Sum
21 & year nbr. nbr. Data Index Fcst Error Level Trend Fcst Data Indicator nbr out
22 0.05365 -0.00029
23 Jan-97 1 1 0.0534 -0.00357 0.05339 -0.00003 0.05338 -0.00026 0.04982 0.04979 0 0
24 Feb-97 2 2 0.0540 -0.00845 0.05314 0.00084 0.05339 -0.00020 0.04468 0.04552 0 0
25 Mar-97 3 3 0.0539 -0.00829 0.05321 0.00071 0.05343 -0.00014 0.04492 0.04563 0 0
26 Apr-97 4 4 0.0523 -0.00484 0.05330 -0.00097 0.05301 -0.00018 0.04845 0.04748 0 0
27 May-97 5 5 0.0522 -0.00171 0.05285 -0.00066 0.05265 -0.00019 0.05114 0.05048 0 0
28 Jun-97 6 6 0.0545 0.00010 0.05248 0.00202 0.05309 -0.00007 0.05258 0.05460 0 0
29 Jul-97 7 7 0.0549 0.01273 0.05302 0.00190 0.05359 0.00003 0.06575 0.06765 0 0
30 Aug-97 8 8 0.0542 0.01370 0.05362 0.00059 0.05380 0.00006 0.06732 0.06791 0 0
31 Sep-97 9 9 0.0514 0.00436 0.05385 -0.00245 0.05312 -0.00007 0.05821 0.05576 0 0
32 Oct-97 10 10 0.0522 0.00102 0.05305 -0.00082 0.05280 -0.00010 0.05407 0.05325 0 0
33 Nov-97 11 11 0.0503 -0.00334 0.05271 -0.00237 0.05200 -0.00021 0.04937 0.04700 0 0
34 Dec-97 12 12 0.0508 -0.00170 0.05181 -0.00103 0.05150 -0.00024 0.05011 0.04908 0 0
35 Jan-98 13 13 0.0508 -0.00357 0.05128 -0.00046 0.05114 -0.00024 0.04771 0.04725 0 0
36 Feb-98 14 14 0.0484 -0.00845 0.05092 -0.00253 0.05017 -0.00034 0.04247 0.03994 0 0
37 Mar-98 15 15 0.0487 -0.00829 0.04986 -0.00119 0.04950 -0.00037 0.04157 0.04038 0 0
38 Apr-98 16 16 0.0518 -0.00484 0.04917 0.00263 0.04996 -0.00020 0.04432 0.04695 0 0
39 May-98 17 17 0.0479 -0.00171 0.04977 -0.00183 0.04923 -0.00027 0.04807 0.04624 0 0
40 Jun-98 18 18 0.0485 0.00010 0.04898 -0.00052 0.04883 -0.00027 0.04908 0.04856 0 0

Forecasting 39
 40
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.02
0.03
0.04
0.05
0.06
0.07
0.08
Jan-97 Jan-97

Apr-97 Apr-97
Figure 4-11

Figure 4-12
Jul-97 Jul-97

Oct-97 Oct-97

Jan-98 Jan-98

Forecast
Forecast

Apr-98 Apr-98

Original data
Jul-98 Jul-98

Oct-98 Oct-98
Seasonally-adjusted data

Jan-99 Jan-99

Apr-99 Apr-99

Jul-99 Jul-99

Oct-99 Oct-99

Jan-00 Jan-00

Apr-00 Apr-00

Forecasting
Trend smoothing
Trend smoothing

Jul-00 Jul-00

Oct-00 Oct-00

Jan-01 Jan-01

Apr-01 Apr-01

Jul-01 Jul-01

Oct-01 Oct-01

Jan-02 Jan-02

Apr-02 Apr-02

Jul-02 Jul-02

Oct-02 Oct-02