Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential...
Transcript of Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential...
![Page 1: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/1.jpg)
Complex exponential Smoothing
Ivan Svetunkov
Nikolaos Kourentzes
Robert Fildes
18 March 2015
![Page 2: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/2.jpg)
Introduction• Exponential Smoothing methods performed very well in
many competitions:
– M-Competitions in 1982 and 2000,
– Competition on telecommunication data in 1998 and 2008,
– Tourism forecasting competition in 2011.
• In practice forecasters usually use:
– SES for the level time series,
– Holt’s method for trend time series,
– Holt-Winters method for a trend-seasonal data.
![Page 3: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/3.jpg)
Introduction• Holt’s method is not performing consistently. Examples:
– M-Competitions;
– Taylor, 2008;
– Gardner & Diaz-Saiz, 2008;
– Acar & Gardner, 2012.
• Holt’s method is still very popular in publications:
– Gelper et. al, 2010;
– Maia & de Carvalho, 2011.
![Page 4: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/4.jpg)
Introduction• Several modifications for different types of trends were
proposed over the years:
– Multiplicative trend (Pegels, 1969);
– Damped trend (Gardner & McKenzie, 1985);
– Damped multiplicative trend (Taylor, 2003);
– Prior data transformation using cross-validation (Bermudez et. al., 2009).
• Model selection procedure based on IC is usually used.
![Page 5: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/5.jpg)
Introduction• But the trend is unobservable and arbitrary!
DGP: ETS(A,N,N)
![Page 6: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/6.jpg)
Reminder
• yt is the real number, actual value,
• yt+ip
t is the complex number
12 i
y t0 y1=15
y t0 y1=15
p t
p1=5
![Page 7: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/7.jpg)
Remark• The fact that imaginary numbers has hitherto been
surrounded by mysterious obscurity, is to be attributed largely to an ill adapted notation.
• If “+1”, “-1”, and “√-1” had been called “direct”, “inverse” and “lateral” units, instead of “positive”, “negative” and “imaginary”, such an obscurity would have been out of the question.
Carl Friedrich Gauss
![Page 8: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/8.jpg)
New approach• We propose a different approach to time series modelling.
• where pt is information potential
• Instead of:
• forecasting model now has a form:
y t+i pt
y t+i pt= f ( y t−1+i p t−1 , y t−2+i p t− 2 , ... , x1 , x2, ...)+ε t+i ξt
y t= f ( y t−1 , y t−2 ,... , x1 , x2, ...)+εt
![Page 9: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/9.jpg)
Theoretical framework• Simple exponential smoothing:
• Principle of CES: smooth level and combine it with information potential estimate.
• Basic form of CES:
ttt yy ˆ ststt yy ˆ tt y ,..., ,2,1 ttt xxf
ttt yyy ˆ1ˆ 1
y t+i pt=(α0+iα1 ) ( y t−1+i ςt−1 )+(1−α0+i−iα1 ) ( y t−1+i p t−1 )
![Page 10: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/10.jpg)
Theoretical framework
• Complex variables -> system of real variables:
• Final forecast of CES consists of two parts:
– smoothed level,
– information potential part.
y t+i pt=(α0+iα1 ) ( y t−1+i ςt−1 )+(1−α0+i−iα1 ) ( y t−1+i p t−1 )
{ y t=(α0 y t−1+ (1−α0 ) y t−1 )−(α1ς t−1+(1−α1) p t−1 )p t=(α0ς t−1+(1−α0) p t−1 )+(α1 y t−1+(1−α1 ) y t−1)
![Page 11: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/11.jpg)
State-space form• Any exponential smoothing method has an underlying
statistical model.
• State-space model with Single Source of Error.
• Every time series consists of components:– level,
– trend,
– seasonality,
– error.
![Page 12: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/12.jpg)
State-space form• Any time series model consists of:
– transition equation: x t=F x t−1+gε t εt∼N (0,σ2)
![Page 13: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/13.jpg)
State-space form• Any time series model consists of:
– measurement equation: y t=w' x t−1+ε t
![Page 14: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/14.jpg)
State-space form• State-space model with Single Source of Error:
– measurement equation:
– transition equation:
y t=w' x t−1+ε t
x t=F x t−1+gε t
![Page 15: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/15.jpg)
State-space form• State-space form of CES:
– measurement equation:
– transition equation:
y t=l t−1+ε t
(l tc t)=(1−(1−α1)
1 1−α0)( l t−1ct−1)+(
−α1α0)ς t+(
α0α1)ε t
![Page 16: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/16.jpg)
State-space form• State-space form:
• Likelihood function:
• Maximizing it is equivalent to minimization of SSE:
f (α0+iα1,σ2∣y)=( 1
σ√2 π )T
exp(−12∑t=1T
(ε tσ )
2
)
SSE=∑t=1
T
ε t2
y t=l t−1+εt
( ltc t)=(1−(1−α1)
1 1−α0)( l t−1c t−1)+(
−α1α0) ςt+(
α0α1)ε t
![Page 17: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/17.jpg)
Time series simulation
![Page 18: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/18.jpg)
• Series N2692 from M3S
eri
es
N2
69
2
1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993
62
00
66
00
70
00
74
00
78
00
Example. Trended series
![Page 19: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/19.jpg)
• ETS(M,A,N)Forecasts from ETS(M,A,N)
Se
rie
s N
26
92
1983 1985 1987 1989 1991 1993
62
00
66
00
70
00
74
00
78
00
82
00
Example. Trended series
![Page 20: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/20.jpg)
• CES• CES
Example. Trended series
α0+ iα1=2.00056+1.00364 i
![Page 21: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/21.jpg)
• CES
Example. Trended series
![Page 22: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/22.jpg)
• Series N1661 in M3S
eri
es
N1
66
1
1990.0 1990.5 1991.0 1991.5 1992.0 1992.5 1993.0 1993.5 1994.0
01
00
03
00
05
00
07
00
0
Example. Stationary series
![Page 23: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/23.jpg)
• ETS(M,N,N)Forecasts from ETS(M,N,N)
1990.0 1991.0 1992.0 1993.0 1994.0 1995.0
01
00
03
00
05
00
07
00
0
Example. Stationary series
![Page 24: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/24.jpg)
• CES
Example. Stationary series
α0+ iα1=0.99999+1.00034 i
![Page 25: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/25.jpg)
• CES
Example. Stationary series
![Page 26: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/26.jpg)
Empirical results: setup• M3-Competition data. 3003 time series.• Rolling origin.• Automated ETS was used to split data into categories:
– level non-seasonal,
– level seasonal,
– trend non-seasonal,
– trend seasonal.
![Page 27: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/27.jpg)
Empirical results: setup• M3-Competition data. 3003 time series.• Rolling origin.• Automated ETS was used to split data into categories.
Series type
Number of series OverallForecasting horizon
Rolling origin horizon
Level series
Trend series
year 255 390 645 6 12
quart 306 450 756 8 16
month 686 742 1428 18 24
other 61 113 174 8 16
Overall 1308 1695 3003
![Page 28: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/28.jpg)
1. Naive (Naive),
2. Simple exponential smoothing (SES),
3. Holt’s additive trend (AAN),
4. Pegels’ multiplicative trend (MMN),
5. State-space ETS with AICc model selection (ZZN),
6. Gardner’s Damped trend (AAdN),
7. Taylor’s Damped multiplicative trend (MMdN),
8. Theta using Hyndman & Billah, 2003 (Theta),
9. Hyndman & Khandakar 2008 Auto ARIMA (ARIMA),
10.Complex exponential smoothing (CES).
Empirical results: competitors
![Page 29: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/29.jpg)
Empirical results• MASE was calculated for each of the horizons from each
of the origins,• Nemenyi test was conducted to compare methods for
each of the series type.• General results for CES:
– at least as good as SES on level series,
– outperforms MMN and AAN on level series,
– at least as good as MMN and AAN on trend series,
– outperforms all the methods on monthly trend series.
Ú
![Page 30: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/30.jpg)
Empirical results. Nemenyi test
![Page 31: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/31.jpg)
Empirical results
Forecasting horizon
MA
SE
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 180.000
0.500
1.000
1.500
2.000
2.500
3.000
CES
SES
Theta
Holt
Damped
MMN
MMdN
Auto ETS
ARIMA
![Page 32: Complex exponential Smoothing - svetunkov.ru · 2015-03-18 · Introduction •Exponential Smoothing methods performed very well in many competitions: – M-Competitions in 1982 and](https://reader033.fdocuments.net/reader033/viewer/2022050420/5f92c47063b5ad3f831ec124/html5/thumbnails/32.jpg)
Conclusions• CES
– is flexible,
– has an underlying statistical model,
– is able to identify trends and levels,
– does it better than Holt and Pegels,
– is at least as good as SES,
– outperforms all the other methods on monthly data,
– is more accurate on long-term horizons.