# complex exponential smoothing - 18-03-2015 ¢ introduction...

Post on 21-Aug-2020

0 views

Embed Size (px)

TRANSCRIPT

Complex exponential Smoothing

Ivan Svetunkov Nikolaos Kourentzes Robert Fildes

18 March 2015

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.

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.

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.

Introduction • But the trend is unobservable and arbitrary!

DGP: ETS(A,N,N)

Reminder • yt is the real number, actual value,

• yt+ipt is the complex number 12 i

y t0 y1=15

y t0 y1=15

p t

p1=5

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

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

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

ŷ t+i p̂t=(α0+iα1 ) ( y t−1+i ςt−1 )+(1−α0+i−iα1 ) ( ŷ t−1+i p̂ t−1 )

Theoretical framework

• Complex variables -> system of real variables:

• Final forecast of CES consists of two parts: – smoothed level, – information potential part.

ŷ t+i p̂t=(α0+iα1 ) ( y t−1+i ςt−1 )+(1−α0+i−iα1 ) ( ŷ t−1+i p̂ t−1 )

{ ŷ t=(α0 y t−1+ (1−α0 ) ŷ t−1 )−(α1ς t−1+(1−α1) p̂ t−1 )p̂ t=(α0ς t−1+(1−α0) p̂ t−1 )+(α1 y t−1+(1−α1 ) ŷ t−1)

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.

State-space form • Any time series model consists of:

– transition equation: x t=F x t−1+gε t εt∼N (0,σ 2)

State-space form • Any time series model consists of:

– measurement equation: y t=w' x t−1+ε t

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

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

State-space form • State-space form:

• Likelihood function:

• Maximizing it is equivalent to minimization of SSE:

f (α0+iα1,σ 2∣y)=( 1σ√2 π )

T

exp(−12∑t=1 T

(ε tσ ) 2

)

SSE=∑ t=1

T

ε t 2

y t=l t−1+εt

( ltc t)=( 1−(1−α1) 1 1−α0 )( l t−1c t−1)+(−α1α0) ςt+(α0α1)ε t

Time series simulation

• Series N2692 from M3 Se

rie s

N 26

92

1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993

62 00

66 00

70 00

74 00

78 00

Example. Trended series

• 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

• CES• CES

Example. Trended series

α0+ iα1=2.00056+1.00364 i

• CES

Example. Trended series

• Series N1661 in M3 Se

rie s

N 16

61

1990.0 1990.5 1991.0 1991.5 1992.0 1992.5 1993.0 1993.5 1994.0

0 10

00 30

00 50

00 70

00

Example. Stationary series

• ETS(M,N,N) Forecasts from ETS(M,N,N)

1990.0 1991.0 1992.0 1993.0 1994.0 1995.0

0 10

00 30

00 50

00 70

00

Example. Stationary series

• CES

Example. Stationary series

α0+ iα1=0.99999+1.00034 i

• CES

Example. Stationary series

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.

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 Overall Forecasting 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

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

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

Empirical results. Nemenyi test

Empirical results

Forecasting horizon

M A

S E

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 0.000

0.500

1.000

1.500

2.000

2.500

3.000

CES SES Theta Holt Damped MMN MMdN Auto ETS ARIMA

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.

Thank you!

Ivan Svetunkov, email: i.svetunkov@lancaster.ac.uk

mailto:i.svetunkov@lancaster.ac.uk

Slide 1 Slide 2 Slide 3 Slide 4 Slide 5 Slide 6 Slide 7 Slide 8 Slide 9 Slide 10 Slide 11 Slide 12 Slide 13 Slide 14 Slide 15 Slide 16 Theoretical framework CES. Examples. Trended series Slide 19 Slide 20 Slide 21 CES. Examples. Stationary series Slide 23 Slide 24 Slide 25 Slide 26 Slide 27 Slide 28 Slide 29 Empirical results Slide 31 Slide 32 Slide 33