Exponential Smoothing Methods

Exponential
Smoothing
Methods
Chapter Topics
Introduction to exponential smoothing
Simple Exponential Smoothing
Holts Trend Corrected Exponential
Smoothing
Holt-Winters Methods
Multiplicative Holt-Winters method
Additive Holt-Winters method
Slide 2
Motivation of Exponential
Smoothing
Simple moving average method assigns
equal weights (1/k) to all k data points.
Arguably, recent observations provide
more relevant information than do
observations in the past.
So we want a weighting scheme that
assigns decreasing weights to the more
distant observations.
Slide 3
Exponential Smoothing
Exponential smoothing methods give
larger weights to more recent
observations, and the weights decrease
exponentially as the observations become
more distant.
These methods are most effective when
the parameters describing the time series
are changing SLOWLY over time.
Slide 4
Data vs Methods
No trend or
seasonal
pattern?
Y
Single
Exponential
Smoothing
Method
Linear trend
and no seasonal
pattern?
Y
Holts Trend
Corrected
Exponential
Smoothing
Method
Both trend
and seasonal
pattern?
Holt-Winters
Methods
Use Other
Methods
Slide 5

The Simple Exponential Smoothing
method is used for forecasting a time
series when there is no trend or seasonal
pattern, but the mean (or level) of the
time series yt is slowly changing over
time.
NOyTREND
model
t o t
Slide 6
Procedures of Simple Exponential

Smoothing Method
Step 1: Compute the initial estimate of the
mean (or level)
of the series at time period t
n
=0
y
0 y
t 1
Step 2: Compute the updated estimate by

using the smoothing equation
l T yT (1 )l T 1
where is a smoothing constant between 0 and

1.
Slide 7
Procedures of Simple Exponential

Smoothing Method
Note that
l T yT (1 )l T 1
yT (1 )[ yT 1 (1 )l T 2 ]
yT (1 ) yT 1 (1 ) 2 l T 2
yT (1 ) yT 1 (1 ) 2 yT 2 ... (1 )T 1 y1 (1 )T l 0
The coefficients measuring the contributions of

the observations decrease exponentially over
time.
Slide 8

Point forecast made at time T for yT+p
yT p (T ) l T
( p 1, 2,3,...)
SSE, MSE, and the standard errors at time T

T
SSE [ y t y t (t 1)] 2
t 1
MSE
SSE
,
T 1
MSE
Note: There is no theoretical justification for dividing SSE

by (T number of smoothing constants). However, we use
this divisor because it agrees to the computation of s in
Box-Jenkins models introduced later.
Slide 9
Example: Cod Catch

The Bay City Seafood Company recorded the monthly
cod catch for the previous two years, as given below.
Cod Catch (In Tons)
Month
Year 1
Year 2
January
362
276
February
381
334
March
317
394
April
297
334
May
399
384
June
402
314
July
375
344
August
349
337
September
386
345
October
328
362
November
389
314
December
343
365
Slide 10
Example: Cod Catch

The plot of these data suggests that there is no trend
or seasonal pattern. Therefore, a NO TREND model is
suggested: yt o t
It is also possible that the mean (or level) is slowly
changing over time.
Slide 11
Example: Cod Catch

Step 1: Compute 0 by averaging the first twelve
time series values.
12
l0
y
t 1
12
362 381 ... 343

360.6667
12
Though there is no theoretical justification, it is a

common practice to calculate initial estimates of
exponential smoothing procedures by using
HALF of the historical data.
Slide 12
Example: Cod Catch

Step 2: Begin with the initial estimate 0 =
360.6667 and update it by applying the
smoothing equation to the 24 observed cod
catches.
Set = 0.1 arbitrarily and judge the
appropriateness of this choice of by the
models
fit. 0.9(360.6667) 360.8000
l y (1in-sample
)l 0.1(362)
1
l 2 y2 (1 )l 1 0.1(381) 0.9(360.8000) 362.8200
Slide 13
One-period-ahead Forecasting
Slide 14
Example: Cod Catch

Results associated with different values of
Smoothing
Constant
Sum of Squared
Errors
0.1
28735.11
0.2
30771.73
0.3
33155.54
0.4
35687.69
0.5
38364.24
0.6
41224.69
0.7
44324.09
0.8
47734.09
Slide 15
Example: Cod Catch

Step 3: Find a good value of that
provides the minimum value for MSE (or
SSE).
SSEas an illustration
Use Solver in Excel
alpha
Slide 16
Example: Cod Catch
Slide 17

Smoothing
If a time series is increasing or decreasing
approximately at a fixed rate, then it may be
described by the LINEAR TREND model
yt 0 1t t
If the values of the parameters 0 and 1 are slowly
changing over time, Holts trend corrected exponential
smoothing method can be applied to the time series
observations.
Note: When neither 0 nor 1 is changing over time,
regression can be used to forecast future values of yt.
Level (or mean) at time T: 0 + 1T
Growth rate (or trend): 1
Slide 18

Smoothing
A smoothing approach for forecasting such a
time series that employs two smoothing
constants, denoted by and .
There are two estimates T-1 and bT-1.
T-1 is the estimate of the level of the time
series constructed in time period T1 (This is
usually called the permanent component).
bT-1 is the estimate of the growth rate of the
time series constructed in time period T1
(This is usually called the trend component).
Slide 19

Smoothing
Level estimate
l T yT (1 )(l T 1 bT 1 )
Trend estimate
bT (l T l T 1 ) (1 )bT 1
where = smoothing constant for the level (0 1)
= smoothing constant for the trend (0 1)
Slide 20

Smoothing
yT p (T ) l T pbT
( p 1, 2,3,...)
MSE and the standard error s at time T

T
SSE [ y t y t (t 1)] 2
t 1
MSE
SSE
,
T 2
s MSE
Slide 21
Yt
Yt 0 1t t
T 1 bT 1 YT 2 (T 1)
( T 1 T ) (1 )bT
bT 1
YT+1
T+1
Y (T ) T bT
T 1
YT (T 1) T 1 bT 1
( T T 1 ) (1 )bT 1
bT
T
YT
T 1
T+1
T+2
Slide 22
Procedures of Holts Trend

Corrected Exponential Smoothing
Use the example of Thermostat Sales as an
illustration
Slide 23

Findings:
Overall an upward trend
The growth rate has been changing over
the 52-week period
There is no seasonal pattern
Holts trend corrected exponential
smoothing method can be applied
Slide 24

Step 1: Obtain initial estimates 0 and b0 by
fitting a least squares trend line to HALF of
the historical data.
y-intercept = 0; slope = b0
Slide 25

Example
Fit a least squares trend line to the first 26
observations
Trend line
y t 202.6246 0.3682t
0 = 202.6246; b0 = 0.3682
Slide 26

Step 2: Calculate a point forecast of y1 from
time 0
yT p (T ) l T pbT
T 0, p 1
Example
y1 (0) l 0 b0 202.6246 0.3682 202.2564
Slide 27

Step 3: Update the estimates T and bT by using
some predetermined values of smoothing
constants.
Example:
)(=
l y let
(1
l 0.2
b ) and = 0.1
1
0.2(206) 0.8(202.6246 0.3682) 203.0051

b1 (l 1 l 0 ) (1 )b0
0.1(203.0051 202.6246) 0.9( 0.3682) 0.2933
y 2 (1) l 1 b1 203.0051 0.2933 202.7118
Slide 28
Slide 29

Step 4: Find the best combination of
and that minimizes SSE (or MSE)
Example: Use Solver in Excel as an
SSE
illustration
alpha
gamma
Slide 30
Slide 31

Smoothing
p-step-ahead forecast made at time T
yT p (T ) l T pbT
( p 1, 2,3,...)
Example
- In period 52, the one-period-ahead sales
forecast for period 53 is
y53 (52) l 52 b52 315.9460 4.5040 320.45
In period 52, the three-period-ahead sales

forecast for period 55 is
y55 (52) l 52 3b52 315.9460 3(4.5040) 329.458
Slide 32

Smoothing
Example
If we observe y53 = 330, we can either find a
new set of (optimal) and that minimize the
SSE for 53 periods, or
we can simply revise the estimate for the
level and growth rate and recalculate the
l 53 y53 (1as
follows:
)(l 52 b52 )
forecasts
0.247(330) 0.753(315.946 4.5040) 322.8089
b53 (l 53 l 52 ) (1 )b52
0.095(322.8089 315.9460) 0.905(4.5040) 4.7281
y54 (53) l 53 b53 322.8089 4.7281 327.537
y55 (53) l 53 2b53 322.8089 2(4.7281) 332.2651
Slide 33
Two Holt-Winters methods are designed for time
series that exhibit linear trend
- Additive Holt-Winters method: used for time
series
with constant (additive) seasonal variations
Multiplicative Holt-Winters method: used for time
series with increasing (multiplicative) seasonal
variations
Holt-Winters method is an exponential smoothing
approach for handling SEASONAL data.
The multiplicative Holt-Winters method is the better
known of the two methods.
Slide 34
Multiplicative Holt-Winters Method

It is generally considered to be best suited to
forecasting time series that can be described by
theyt equation:
( 0 1t ) SN t IRt
SNt: seasonal pattern
IRt: irregular component
This method is appropriate when a time series
has a linear trend with a multiplicative seasonal
pattern for which the level (0+ 1t), growth rate
(1), and the seasonal pattern (SNt) may be slowly
changing over time.
Slide 35

Estimate of the level
l T ( yT / snT L ) (1 )(l T 1 bT 1 )
Estimate of the growth rate (or trend)
bT (l T l T 1 ) (1 )bT 1
Estimate of the seasonal factor

snT ( yT / l T ) (1 ) snT L
where , , and are smoothing constants between 0
and 1,
L = number of seasons in a year (L = 12 for
monthly data,
and L = 4 for quarterly
data)
Slide 36

yT p (T ) (l T pbT ) snT p L
( p 1, 2,3,...)
T
MSE
2
errors at time T
SSEand
[the
y t ystandard
(
t
1
)]
t
t 1
MSE
SSE
,
T 3
MSE
Slide 37
Procedures of Multiplicative HoltWinters Method

Use the Sports Drink example as an illustration
Slide 38

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
Slide 39

Step 1: Obtain initial values for the level 0,
the growth rate b0, and the seasonal factors
sn-3, sn-2, sn-1, and sn0, by fitting a least
squares trend line to at least four or five
years of the historical data.
Slide 40

Example
Fit a least squares trend line to the first 16
observations
Trend line
yt 95.2500 2.4706t
0 = 95.2500; b0 = 2.4706
Slide 41

Step 2: Find the initial seasonal factors
1. Compute yt for the in-sample observations
used for fitting the regression. In this
example, t = 1, 2, , 16.
y1 95.2500 2.4706(1) 97.7206
y 2 95.2500 2.4706(2) 100.1912
......
y16 95.2500 2.4706(16) 134.7794
Slide 42

St yt / yt
2. Detrend the data by computing
for
each time period that is used in finding the
least squares regression equation. In this
example,
2, , 0.7368
16.
S y / ty =
721,
/ 97.7206
1
S 2 y2 / y 2 116 /100.1912 1.1578

......
S16 y16 / y16 120 /134.7794 0.8903
Slide 43

3. Compute the average seasonal values for
each of the L seasons. The L averages are
found by computing the average of the
detrended values for the corresponding
season. For example, for quarter 1,
S1 S5 S9 S13
S[1]
4
0.7368 0.7156 0.6894 0.6831
0.7062
4
Slide 44

4. Multiply the average seasonal values by the
normalizing constant
CF
L
L
S
i 1
[i ]
such that the average of the seasonal factors

is 1. The initial seasonal factors are
sni L S[ i ] (CF )
(i 1, 2,..., L )
Slide 45

4. Multiply the average seasonal values by the
normalizing constant such that the average of
the seasonal factors is 1.
Example
CF = 4/3.9999 = 1.0000
sn3 sn14 S[1] (CF ) 0.7062(1) 0.7062
sn2 sn2 4 S[2] (CF ) 1.1114(1) 1.1114
sn1 sn3 4 S[3] (CF ) 1.2937(1) 1.2937
sn0 sn4 4 S[1] (CF ) 0.8886(1) 0.8886
Slide 46

time 0 using the initial values
(T 0, p 1)
y1 (0) (l 0 b0 ) sn1 4 (l 0 b0 ) sn3

(95.2500 2.4706)(0.7062)
69.0103
Slide 47

Step 4: Update the estimates T, bT, and snT by
using some predetermined values of smoothing
constants.
Example:
l ( y /let
sn
) =
(1 0.2,
)(l b=
) 0.1, and = 0.1
1
1 4
0.2(72 / 0.7062) 0.8(95.2500 2.4706) 98.5673
b1 (l 1 l 0 ) (1 )b0
0.1(98.5673 95.2500) 0.9(2.4706) 2.5553
sn1 ( y1 / l 1 ) (1 ) sn1 4
0.1(72 / 98.5673) 0.9(0.7062) 0.7086
y 2 (1) (l 1 b1 ) sn2 4
(98.5673 2.5553)(1.1114) 112.3876
Slide 48
2 y2 sn24 1 1 b1
0.2116 1.1114 0.8 98.5673 2.5553

101.7727
b2 2 1 1 b1
0.1101.7727 98.5673 0.9 2.5553

2.62031
sn2 y2 2 1 sn24
0.1116 101.7727 0.91.1114

1.114239
y 3 2 2 b2 sn34
101.7727 2.620311.2937
135.053
Slide 49
4 y4 sn44 1 3 b3
0.2 96 0.8886 0.8104.5393 2.6349

107.3464
b4 4 3 1 b3
0.1107.3464 104.5393 0.9 2.6349

2.65212
sn4 y4 4 1 sn4 4
0.1 96 107.3464 0.9 0.8886

0.889170
y 5 4 4 b4 sn54
107.3464 2.65212 0.7086

77.945
Slide 50
Slide 51

Step 5: Find the most suitable
combination of , , and that minimizes
SSE (or MSE)
Example: Use Solver inSSE
Excel as an
illustration
alpha
gamma
delta
Slide 52
Slide 53

( p 1, 2,3,...)
Example
y33 (32) (l 32 b32 ) sn33 4 (168.1213 2.3028)(0.7044) 120.0467
y34 (32) (l 32 2b32 ) sn34 4 [168.1213 2(2.3028)](1.1038) 190.6560
y35 (32) (l 32 3b32 ) sn35 4 [(168.1213 3(2.3028)](1.2934) 226.3834
y36 (32) (l 32 4b32 ) sn36 4 [(168.1213 4(2.3028)](0.8908) 157.9678
Slide 54

Example
Slide 55
Additive Holt-Winters Method

It is generally considered to be best suited to
forecasting a time series that can be described
by ythe
equation:
t ( 0 1t ) SN t IRt
SNt: seasonal pattern
IRt: irregular component
This method is appropriate when a time series
has a linear trend with a constant (additive)
seasonal pattern such that the level (0+ 1t),
growth rate (1), and the seasonal pattern (SNt)
may be slowly changing over time.
Slide 56

Estimate of the level
T ( yT snT L ) (1 )(T 1 bT 1 )
Estimate of the growth rate (or trend)

bT (T T 1 ) (1 )bT 1
Estimate of the seasonal factor

snT ( yT T ) (1 ) snT L
where , , and are smoothing constants between 0
and 1,
L = number of seasons in a year (L = 12 for
monthly data,
and L = 4 for quarterly
data)
Slide 57

y T p (T ) T pbT snT p L
( p 1, 2, 3,...)
MSE andT the standard error s at time T

2
SSE [ y t y t (t 1)]
t 1
MSE
SSE
,
T 3
MSE
Slide 58
Procedures of Additive HoltWinters Method

Consider the Mountain Bike example,
Slide 59

Observations:
Linear upward trend over the 4-year
period
Magnitude of seasonal span is almost
constant as the level of the time series
increases
Additive Holt-Winters method can be
applied to forecast future sales
Slide 60

Step 1: Obtain initial values for the level 0,
the growth rate b0, and the seasonal factors
sn-3, sn-2, sn-1, and sn0, by fitting a least
squares trend line to at least four or five
years of the historical data.
Slide 61

Example
Fit a least squares trend line to all 16
observations
Trend
y t 20line
.85 0.980882t
0 = 20.85; b0 = 0.9809
Slide 62

1. Compute yt for each time period that is
used in finding the least squares regression
equation. In this example, t = 1, 2, , 16.
y1 20.85 0.980882(1) 21.8309
y 2 20.85 0.980882(2) 22.8118
......
y16 20.85 0.980882(16) 36.5441
Slide 63

S t y t y t
2. Detrend the data by computing
for
each observation used in the least squares
fit. In this example, t = 1, 2, , 16.
S1 y1 y1 10 21.8309 11.8309
S 2 y 2 y 2 31 22.8112 8.1882
......
S16 y16 y16 25 36.5441 11.5441
Slide 64

3. Compute the average seasonal values for
each of the L seasons. The L averages are
found by computing the average of the
detrended values for the corresponding
season. For example, for quarter 1,
S1 S 5 S 9 S13
4
(11.8309) (14.7544) (15.6779) (14.6015)
14.2162
4
S [1]
Slide 65

4. Compute the average of the L seasonal
factors. The average should be 0.
Slide 66

time 0 using the initial values
(T 0, p 1)
y1 (0) 0 b0 sn1 4 0 b0 sn 3
20.85 0.9809 (-14.2162) 7.6147
Slide 67

Step 4: Update the estimates T, bT, and snT by
using some predetermined values of smoothing
constants.
Example:
= 0.2, = 0.1, and = 0.1
1 ( y1let
sn1
4 ) (1 )(0 b0 )
0.2(10 (14.2162)) 0.8(20.85 0.9808) 22.3079
b1 (1 0 ) (1 )b0
0.1(22.3079 20.85) 0.9(0.9809) 1.0286
sn1 ( y1 1 ) (1 ) sn1 4
0.1(10 22.3079) 0.9(14.2162) 14.0254
y 2 (1) 1 b1 sn 2 4 1 b1 sn 2
22.3079 1.0286 6.5529 29.8895
Slide 68
Slide 69

Step 5: Find the most suitable
combination of , , and that minimizes
SSE (or MSE)
Example: Use Solver inSSE
Excel as an
illustration
alpha
gamma
delta
Slide 70
Slide 71

( p 1, 2, 3,...)
Example
y17 (16) 16 b16 sn17 4 36.3426 0.9809 14.2162 23.1073
y18 (16) 16 2b16 sn18 4 36.3426 2(0.9809) 6.5529 44.8573
y19 (16) 16 3b16 sn19 4 36.3426 3(0.9809) 18.5721 57.8573
y 20 (16) 16 4b16 sn20 4 36.3426 4(0.9809) 10.9088 29.3573
Slide 72

Example
Slide 73
Chapter Summary
No trend, no seasonal pattern
Smoothing
Trend, no seasonal pattern
Both trend and seasonal pattern
Multiplicative Holt-Winters method
Slide 74

Exponential Smoothing Methods

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Exponential Smoothing Methods

Uploaded by

Copyright:

Available Formats

Exponential

Simple Exponential Smoothing

Procedures of Simple Exponential

Step 2: Compute the updated estimate by

where is a smoothing constant between 0 and

Procedures of Simple Exponential

The coefficients measuring the contributions of

Simple Exponential Smoothing

SSE, MSE, and the standard errors at time T

Note: There is no theoretical justification for dividing SSE

Example: Cod Catch

Example: Cod Catch

Example: Cod Catch

362 381 ... 343

Though there is no theoretical justification, it is a

Example: Cod Catch

l 2 y2 (1 )l 1 0.1(381) 0.9(360.8000) 362.8200

Example: Cod Catch

Example: Cod Catch

Example: Cod Catch

Holts Trend Corrected Exponential

Holts Trend Corrected Exponential

Holts Trend Corrected Exponential

Holts Trend Corrected Exponential

MSE and the standard error s at time T

Procedures of Holts Trend

Procedures of Holts Trend

Procedures of Holts Trend

Procedures of Holts Trend

Procedures of Holts Trend

Procedures of Holts Trend

0.2(206) 0.8(202.6246 0.3682) 203.0051

Procedures of Holts Trend

Holts Trend Corrected Exponential

In period 52, the three-period-ahead sales

Holts Trend Corrected Exponential

Multiplicative Holt-Winters Method

Multiplicative Holt-Winters Method

Estimate of the seasonal factor

Multiplicative Holt-Winters Method

Procedures of Multiplicative HoltWinters Method

Procedures of Multiplicative HoltWinters Method

Procedures of Multiplicative HoltWinters Method

Procedures of Multiplicative HoltWinters Method

Procedures of Multiplicative HoltWinters Method

Procedures of Multiplicative HoltWinters Method

S 2 y2 / y 2 116 /100.1912 1.1578

Procedures of Multiplicative HoltWinters Method

Procedures of Multiplicative HoltWinters Method

such that the average of the seasonal factors

Procedures of Multiplicative HoltWinters Method

Procedures of Multiplicative HoltWinters Method

y1 (0) (l 0 b0 ) sn1 4 (l 0 b0 ) sn3

Procedures of Multiplicative HoltWinters Method

0.2(72 / 0.7062) 0.8(95.2500 2.4706) 98.5673

0.2116 1.1114 0.8 98.5673 2.5553

0.1101.7727 98.5673 0.9 2.5553

0.1116 101.7727 0.91.1114

0.2 96 0.8886 0.8104.5393 2.6349

0.1107.3464 104.5393 0.9 2.6349

0.1 96 107.3464 0.9 0.8886

107.3464 2.65212 0.7086

Procedures of Multiplicative HoltWinters Method