EXPONENTIAL

SMOOTHING
S T Q S 3 11 3 S TAT I S T I C A L M O D E L L I N G

DR. RAZIK RIDZUAN

INTRODUCTION
• Exponential smoothing is a forecasting method that effective when the components
(trend, seasonal factors) may be changing over time.
• This method will weigh the observed time series values unequally.
• Recent observations are weighted more heavily than remote observations.
• The unequal weight is accomplished by using one or more smoothing constants.
• The smoothing constants will determine how much weight is given to each observation

2
ADDITIVE HOLT-WINTERS METHOD (AHW)

• Additive Holt-Winters method is used for the time series with constant seasonal variations and
having linear trend (at least locally).
• It is generally considered to be best suited to forecasting a time series that can be described by
the equation:

• Fixed seasonal pattern, and fixed growth rate .

• However, if the time series exhibit constant seasonal variation with the level, growth rate,
and the seasonal pattern may be changing (slowly) over time, the additive Holt-Winters
method will be useful.

3
• At time , the estimate of the level of the time series is

• The estimate of the growth rate (trend) is

• The estimate of the seasoanal factor is

where are the smoothing constants and is the number of seasons ( for monthly data or
for quarterly data)
• You can also rewrite the three equations above in the form of error correction.

4
FORECASTING
• A point forecast made in time for is

where is the “latest” estimate of the seasonal factor for the season corresponding to
time .
• A 95% prediction interval for is

where
– when , ,
– when ,

5
 – when ,

where

• The standard error is

6
EXAMPLE MOUNTAIN BIKE SALES DATA

Year
Q
1 2 3 4
1 10 11 14 19
2 31 33 36 41
3 43 45 50 55
4 16 17 21 25

7
• Linear upward trend over the 4-year period.
• Magnitude of seasonal variation is almost constant as the level of the time series increases.
• Additive Holt-Winters method can be applied to forecast future sales.

• Step 1: Obtain initial values for the level , the growth rate , and the seasonal factors , ,
and , by fitting a least squares trend line to at least four or five years of the historical
data.

• We have

8
• Step 2: Find the initial seasonal factors (3 procedures)
1. Compute for each time period that is used in finding the least squares regression
equation. In this example, .

2. Detrend the data by computing for each observation used in the squares regression
equation.

9
3. Compute the average seasonal values for each of the seasons. For example, for
quarter 1,

For other quarters,

20.8500
• Compute 0.9809
the average -14.2163factors.6.5528
of the L seasonal The average18.5719
should be 0.-10.9090

10
• Step 3: Calculate a point forecast of from time 0 using the initial values.

The forecast error is

• Step 4 – Update the estimates , and by using predetermined values of smoothing

constants
Let

Let , the

11
Step 5 – Find the optimum values of and that gives the smallest SSE.

0.1 0.1 0.1 25.7291

0.2 0.4 0.05 34.6746
0.5 0.4 0.2 25.6421
0.5607 0 0 18.7983

Therefore, are the optimum values of smoothing constants.

12
FORECASTING

• In AHW, the point estimate forecast for steps ahead at time is

• In this example,

13
14
MULTIPLICATIVE HOLTS-WINTER METHOD
(MHW)

• If the time series has a linear trend with fixed growth rate, fixed seasonal pattern, but
increasing seasonal variation, thus it can be described by the equation:

where : seasonal pattern and : irregular component

• MHW is appropriate when a time series has a linear trend with a multiplicative
seasonal pattern for which the level (), growth rate (), and the seasonal pattern () may
be slowly changing over time (with increasing seasonal variation).

15
• At time , the estimate of the level of the time series is

• The estimate of the growth rate (trend) is

• The estimate of the seasonal factor is

where are the smoothing constants and is the number of seasons ( for monthly data or
for quarterly data)

16
FORECASTING
• A point forecast made in time for is

where is the “latest” estimate of the seasonal factor for the season corresponding to
time .
• A 95% prediction interval for is

where
– when ,

17
 – when ,

• The relative standard error is

18
 Quarterly sales of sports drink
Year
Q
1 2 3 4 5 6 7 8
1 72 77 81 87 94 102 106 115
2 116 123 131 140 147 162 170 177
3 136 146 158 167 177 191 200 218
4 96 101 109 120 128 134 142 149

19
• Observations:
– Linear upward trend over the 8-year
period
– Magnitude of the seasonal span increases
as the level of the time series increases
• Multiplicative Holt-Winters method can be
applied to forecast future sales

20
• Step 1 – Obtain the initial values of the level , the growth rate , and the seasonal
factors and by fitting a least squares trend line to at least 4-5 years of historical
data.
In this example, we use 4 years data

• Step 2 – Find the initial seasonal factors

1. Compute for

21
2. Detrend the data by computing for

22
3. Compute the average seasonal values for each seasons
For quarter 1,

For other quarters,

23
4. Multiply the average seasonal values by the normalizing constant

such that the average of the seasonal factor is 1. The initial seasonal factors are

For this example,

The seasonal factors for initial period of time

24
• Step 3 – Calculate a point forecast of at time 0

The forecast error is

• Step 4 – Update the estimates , and by using some predetermined smoothing constant
values.
Let

25
The forecast error is

26
Step 5 – Find the optimum values of and that gives the smallest SSE.

0.1 0.1 0.1 216.93

0.3 0.7 0.4 251.89
0.07 0.06 0.3 231.54
0.3355 0.0456 0.1340 168.45

Therefore, are the optimum values of smoothing constants.

27
FORECASTING

• The forecasts -step ahead made at time

For this data, we forecast 1-year ahead (4-step).

• Try obtaining the prediction intervals for these 4 forecasts.

28
29