Time Series Analysis - II
Arnab K Laha
Indian Institute of Management
Ahmedabad
Single Exponential Smoothing
This single exponential smoothing method is appropriate
for series that move randomly above and below a constant
mean with no trend nor seasonal patterns.
To correct past forecast mistakes
Next forecast = current forecast + “correction.”
“Correction” = a fraction of current forecast error.
Correction factor α is between 0 and 1.
α is called the smoothing constant.
Next forecast depends on value of correction factor α
and the current forecast error.
Single Exponential Smoothing
or
It continually revises an estimate in
the light of more recent experiences
Autopilot
Single Exponential Smoothing
α is the smoothing factor. The smaller is the α, the
smoother is the forecasted series. By repeated
substitution, we can rewrite the recursion as
t −1
Ŷt +1 = αYt + α(1 − α )Yt −1 + ... + α(1 − α ) Y1
t −1
s
= α ∑ (1 − α ) Yt −s
s =0
This shows why this method is called exponential smoothing –
the forecast is a weighted average of the past values of the series,
where the weights decline exponentially with time.
Single Exponential Smoothing:
Forecasting
The forecasts from single exponential smoothing
are constant for all future observations.
Let T = n + 1 where n is the number of observatio ns
ŶT +k = ŶT for k = 1,2,...
How to estimate α ???
•hunch
•MSE
INITIAL VALUE
How do we start?
the first estimate
of smoothed series = the first observation
just… average!
( use an average of first 5 or 6 observation)
Single Exponential Smoothing
Bowerman and O'Connell (1979) suggest that
values of around 0.01 to 0.30 work quite well.
You can also estimate α to minimize the sum of
squares of one-step forecast errors.
No. air passengers by week in Sweden
No.passengers at Swedish airports
No. passengers
13.6
13.4
13.2
(thousands)
13.0
12.8
12.6
12.4
12.2
1992 1996 2000 2004
Single exponential forecast of the number of air passengers
(α
α=0.2)
Ft +1 = Ft + α (Yt − Ft ) = αYt + (1− α )Ft
Smoothing Plot for No. passengers
Single Exponential Method
Variable
13.50
A ctual
Fits
Forecasts
95.0% PI
13.25
Smoothing Constant
No. passengers
A lpha 0.2
A ccuracy Measures
13.00
MAPE 0.688525
MAD 0.089820
MSD 0.014213
12.75
12.50
1 73 146 219 292 365 438 511 584 657 730
Index
Single exponential forecast of the number of air passengers
(α
α=0.05)
Ft +1 = Ft + α (Yt − Ft ) = αYt + (1 − α )Ft
Smoothing Plot for No. passengers
Single Exponential Method
Variable
13.50
Actual
Fits
Forecasts
95.0% PI
13.25
Smoothing C onstant
No. passengers
Alpha 0.05
Accuracy Measures
13.00
MAPE 0.773411
MAD 0.100988
MSD 0.017768
12.75
12.50
1 73 146 219 292 365 438 511 584 657 730
Index
HOLT’S LINEAR METHOD
Holt’s two-parameter method,
1957
Smoothes the level and the slope, using
constants
Constants provide estimates of level and slope,
adapt over time as new observations become
available
Great deal of flexibility in selecting of the
rates at which the level and trend are changed.
Holt’s Linear Method
This method is appropriate for series with a linear
time trend and no seasonal variation.
Ideas behind smoothing with trend:
``De-trend'' time-series by separating base
from trend effects
Smooth base in usual manner using α
Smooth trend forecasts in usual manner using
β
Also called Double Exponential Smoothing
Holt’s Method:
Smooth the base forecast Bt
B t = α Y t + (1 − α )( B t − 1 + T t − 1 )
Smooth the trend forecast Tt
Tt = β (B t − B t −1 ) + (1 − β )Tt −1
Forecast k periods into future Ft+k with base
and trend
Ft + k = B t + kTt
Holt’s Linear Method: How do we
start?
the first estimate of the smoothed
series = the first observation Ti = 0
Or,
use an average of first 5 or 6
observation
Choosing α and β
Grid of value of α and β
Select a combination: lowest MSE
Example: US Housing Starts
2200
US Housing Starts
2100
2000
1900
1800
Index 1 2 3 4 5 6 7 8 9
Single Exponential Smoothing (α=0.2)
Single Exponential Smoothing
2250 Actual
Predicted
Forecast
2150 Actual
US Housing S
Predicted
Forecast
2050
Smoothing Constant
1950 Alpha: 0.200
MAPE: 3.8
1850 MAD: 76.5
MSD: 13354.3
0 5 10
Time
Single Exponential Smoothing:
Optimal α
Single Exponential Smoothing
2240 Actual
Predicted
Forecast
2140 Actual
US Housing S
Predicted
Forecast
2040
Smoothing Constant
1940 Alpha: 0.017
MAPE: 3.7
MAD: 74.9
1840
MSD: 11845.7
0 5 10
Time
Double Exponential Smoothing
(α=0.2, β=0.2)
Double Exponential Smoothing for US Housing S
2250 Actual
2200 Predicted
Forecast
2150
Actual
US Housing S
2100 Predicted
Forecast
2050
2000
Smoothing Constants
1950 Alpha (level): 0.200
Gamma (trend): 0.200
1900
1850 MAPE: 4.0
MAD: 80.4
1800 MSD: 13972.4
0 5 10
Time
Double Exponential Smoothing:
Optimal α, β
Double Exponential Smoothing for US Housing S
Actual
2500 Predicted
Forecast
Actual
US Housing S
Predicted
Forecast
2000
Smoothing Constants
Alpha (level): 0.010
Gamma (trend): 99.990
MAPE: 11.5
1500 MAD: 230.7
MSD: 77792.9
0 5 10
Time
Holt-Winter’s Multiplicative
Method:
Ideas behind smoothing with trend and
seasonality:
“De-trend’: and “de-seasonalize” time-series by
separating base from trend and seasonality
effects
Smooth base in usual manner using α
Smooth trend forecasts in usual manner using β
Smooth seasonality forecasts using γ
Holt-Winter’s Multiplicative
Method:
Smooth the base forecast Bt
Yt
Bt = α + (1 − α )( B t − 1 + T t − 1 )
S t−m
(m = no. of seasons)
Smooth the trend forecast Tt
T t = β ( B t − B t −1 ) + (1 − β ) T t −1
Smooth the seasonality forecast St
Yt
St = γ + (1 − γ )S t −m
Bt
Holt-Winter’s Multiplicative
Method:
Forecast Ft+k (k=1,2,…) with trend and
seasonality
Ft+k = (Bt + kTt )St+k−m
Holt-Winter’s Multiplicative
Method: How do we start?
Bm = (Y1+…+Ym)/m
Tm =[{(Ys+1-Y1)/m} + {(Ys+2-Y2)/m} +…+
{(Ys+s-Ys)/m}]/m
S1=Y1/Bm, S2=Y2/Bm,…, Sm=Ym/Bm
Holt-Winter’s Additive Method:
Smooth the base forecast Bt
B t = α ( Y t − S t − m ) + (1 − α )( B t − 1 + T t − 1 )
(m = no. of seasons)
Smooth the trend forecast Tt
T t = β ( B t − B t − 1 ) + (1 − β ) T t − 1
Smooth the seasonality forecast St
S t = γ ( Yt − B t ) + (1 − γ )S t −m
Holt-Winter’s Additive Method:
Forecast Ft+k (k=1,2,…) with trend and
seasonality
Ft+k = Bt + kTt + St+k−m
Holt-Winter’s Additive Method:
How do we start?
Bm = (Y1+…+Ym)/m
Tm =[{(Ys+1-Y1)/m} + {(Ys+2-Y2)/m} +…+
{(Ys+s-Ys)/m}]/m
S1=Y1-Bm, S2=Y2-Bm,…, Sm=Ym-Bm
Holt-Winter’s Additive Method:
(α = β = γ =0.5)
Winters' Additive Model for Meals Served
Actual
15000
Predicted
14000 Forecast
Actual
Meals Served
13000 Predicted
Forecast
12000
11000 Smoothing Constants
Alpha (level): 0.500
10000 Gamma (trend): 0.500
Delta (season): 0.500
9000
MAPE: 5
MAD: 556
8000
MSD: 419183
0 5 10 15
Time
