the caterpillar ssa approach to time series analysis and
play

The Caterpillar-SSA approach to time series analysis and its - PowerPoint PPT Presentation

The Caterpillar-SSA approach to time series analysis and its automatization Th.Alexandrov, N.Golyandina theo@pdmi.ras.ru, nina@ng1174.spb.edu St.Petersburg State University Caterpillar-SSA and its automatization p. 1/15


  1. The “Caterpillar”-SSA approach to time series analysis and its automatization Th.Alexandrov, N.Golyandina theo@pdmi.ras.ru, nina@ng1174.spb.edu St.Petersburg State University “Caterpillar”-SSA and its automatization – p. 1/15

  2. Outline 1. History 2. Possibilities and advantages 3. “Caterpillar”-SSA: basic algorithm 4. Decomposition feasibility 5. Identification of SVD components 6. Example: trend and seasonality extraction 7. Forecast 8. Example: signal forecast 9. AutoSSA: 1) extraction of trend 2) extraction of cyclic components 3) model example 4) conclusions 10. Future plans “Caterpillar”-SSA and its automatization – p. 2/15

  3. History Origins: � Singular system approach to the method of delays . Dynamic Systems – analysis of attractors [middle of 80’s] ( Broomhead ) � Singular Spectrum Analysis . Geophysics/meteorology – signal/noise enhancing, distinguishing of a time series from the red noise realization (Monte Carlo SSA) [90’s] ( Vautard, Ghil, Fraedrich ) � “Caterpillar” . Principal Component Analysis evolution [end of 90’s] ( Danilov, Zhigljavskij, Solntsev, Nekrutkin, Goljandina ) Books: � Elsner, Tsonis. Singular Spectrum Analysis. A New Tool in Time Series Analysis , 1996. � Golyandina, Nekrutkin, and Zhigljavsky. Analysis of Time Series Structure: SSA and Related Techniques , 2001. Internet links and software: � http://www.atmos.ucla.edu/tcd/ssa/ � http://www.gistatgroup.com/cat/ � http://www.pdmi.ras.ru/˜theo/autossa/ “Caterpillar”-SSA and its automatization – p. 3/15

  4. Possibilities and advantages Basic possibilities of the “Caterpillar”-SSA technique: � Finding trend of different resolution � Smoothing � Extraction of seasonality components Simultaneous extraction of cycles with small and large periods • Extraction periodicities with varying amplitudes • Simultaneous extraction of complex trends and periodicities • � Forecast � Change-point detection Advantages: � Doesn’t require the knowledge of parametric model of time series � Works with wide spectrum of real-life time series � Matches up for non-stationary time series � Allows to find structure in short time series “Caterpillar”-SSA and its automatization – p. 4/15

  5. “Caterpillar”-SSA: basic algorithm F N = F (1) N + . . . + F ( m ) � Decomposes time series into sum of additive components: N � Provides the information about each component   Algorithm: f 0 f 1 . . . f N − L   1. Trajectory matrix construction:   f 1 f 2 . . . f N − L +1     F N = ( f 0 , . . . , f N − 1 ) , F N → X ∈ R L × K X = . .   ... ... . .   . .   ( L – window length, parameter) f L − 1 f L . . . f N − 1 � λ j U j V T X j = 2. Singular Value Decomposition (SVD): j X = � X j λ j – eigenvalue, U j – e.vector of XX T , � V j – e.vector of X T X , V j = X T U j / λ j X ( k ) = � 3. Grouping of SVD components: { 1 , . . . , d } = � I k , j ∈ I k X j 4. Reconstruction by diagonal averaging: F (1) F ( m ) F N = � + . . . + � X ( k ) → � F ( k ) N N N “Caterpillar”-SSA and its automatization – p. 5/15

  6. Decomposition feasibility X = X (1) + X (2) F N = F (1) + F (2) N , N Separability: we can form the group I 1 of SVD components so that I 1 ↔ X (1) � Separability is the necessary condition for “Caterpillar”-SSA application � Exact separability impose strong constraints at the spectrum of time series which could be processed Real-life: asymptotic separability I 1 ↔ X (1) ↔ � F N = F (1) + F (2) F (1) –approximation of the F (1) The case of finite N : N , N N N Examples of asymptotic separability: � A determinate signal is asympt. separable from a white noise � A periodicity is asympt. separable from a trend Fulfilment of (asymptotic) separability conditions places limitations on the value of window length L Only time series which generate finite amount of SVD components – hard constraint? Any linear combination of product of exponents , harmonics and polynomials generates finite amount of SVD components “Caterpillar”-SSA and its automatization – p. 6/15

  7. Identification of SVD components Identification – choosing of SVD components on the stage of grouping. Important examples f n = Ae αn Exponential trend: � it generates one SVD component, � eigenvector: U = ( u 1 , . . . , u L ) T : u k = Ce αk ( “exponential” form with the same α ) f n = Ae αn cos(2 πωn ) Exponentially-modulated harmonic: � it generates two SVD components, � eigenvectors: L ) T : = C 1 e αk cos(2 πωk ) U 1 = ( u (1) 1 , . . . , u (1) u (1) k L ) T : = C 2 e αk sin(2 πωk ) U 2 = ( u (2) 1 , . . . , u (2) u (2) k ( “exponentially-modulated” form with the same α and ω ) “Caterpillar”-SSA and its automatization – p. 7/15

  8. Identification of SVD components Identification – choosing of SVD components on the stage of grouping. Important examples f n = Ae αn Exponential trend: � it generates one SVD component, � eigenvector: U = ( u 1 , . . . , u L ) T : u k = Ce αk ( “exponential” form with the same α ) f n = Ae αn cos(2 πωn ) Exponentially-modulated harmonic: � it generates two SVD components, � eigenvectors: L ) T : = C 1 e αk cos(2 πωk ) U 1 = ( u (1) 1 , . . . , u (1) u (1) k L ) T : = C 2 e αk sin(2 πωk ) U 2 = ( u (2) 1 , . . . , u (2) u (2) k ( “exponentially-modulated” form with the same α and ω ) “Caterpillar”-SSA and its automatization – p. 7/15

  9. Identification of SVD components Identification – choosing of SVD components on the stage of grouping. Important examples f n = Ae αn Exponential trend: � it generates one SVD component, � eigenvector: U = ( u 1 , . . . , u L ) T : u k = Ce αk ( “exponential” form with the same α ) f n = Ae αn cos(2 πωn ) Exponentially-modulated harmonic: � it generates two SVD components, � eigenvectors: L ) T : = C 1 e αk cos(2 πωk ) U 1 = ( u (1) 1 , . . . , u (1) u (1) k L ) T : = C 2 e αk sin(2 πωk ) U 2 = ( u (2) 1 , . . . , u (2) u (2) k ( “exponentially-modulated” form with the same α and ω ) “Caterpillar”-SSA and its automatization – p. 7/15

  10. Example: trend and seasonality extraction Traffic fatalities. Ontario, monthly, 1960-1974 Abraham, Redolter. Stat. Methods for Forecasting , 1983 N=180, L=60 “Caterpillar”-SSA and its automatization – p. 8/15

  11. Example: trend and seasonality extraction Traffic fatalities. Ontario, monthly, 1960-1974 Abraham, Redolter. Stat. Methods for Forecasting , 1983 N=180, L=60 “Caterpillar”-SSA and its automatization – p. 8/15

  12. Example: trend and seasonality extraction Traffic fatalities. Ontario, monthly, 1960-1974 Abraham, Redolter. Stat. Methods for Forecasting , 1983 SVD components: 1, 4, 5 N=180, L=60 “Caterpillar”-SSA and its automatization – p. 8/15

  13. Example: trend and seasonality extraction Traffic fatalities. Ontario, monthly, 1960-1974 Abraham, Redolter. Stat. Methods for Forecasting , 1983 SVD components: 1, 4, 5 N=180, L=60 “Caterpillar”-SSA and its automatization – p. 8/15

  14. Example: trend and seasonality extraction Traffic fatalities. Ontario, monthly, 1960-1974 Abraham, Redolter. Stat. Methods for Forecasting , 1983 SVD components: 1, 4, 5 N=180, L=60 SVD components with periods estimations: 2-3 (12), 4-5, 6-7(6), 8(2), 9-10(10), 11-12(4), 13-14(2.4) “Caterpillar”-SSA and its automatization – p. 8/15

  15. Forecast Recurrent forecast in the framework of the “Caterpillar”-SSA. Linear Recurrent Formula (LRF): f n = � d F N = ( f 0 , . . . , f N − 1 ) : k =1 a k f n − k Theory: � All time series with finite number of SVD components generates finite order LRFs F N = F (1) + F (2) N , if { U j } j ∈ I 1 ↔ F (1) then we can find a LRF governing the F (1) � N N N Algorithm of F (1) one-step forecast: N 1. Choose window length L Identify SVD components corresponding to the F (1) { U j } j ∈ I 1 ↔ F (1) 2. N : N F (1) { U j } j ∈ I 1 ⇒ � 3. Reconstruct via diagonal averaging: N F (1) Find a LRF governing � 4. using the theorem N F (1) Apply the LRF to last values of the � f N 5. N : “Caterpillar”-SSA and its automatization – p. 9/15

  16. Example: signal forecast First 119 points were given as the base for the signal reconstruction and forecast Remaining part of the time series is figured to estimate the forecast quality N=119, L=60, forecast of points 120-180 SVD components: 1 (trend); 2-3, 5-6, 9-10 (harmonics with periods 12, 4, 2.4); 4 (harmonic with period 2) “Caterpillar”-SSA and its automatization – p. 10/15

  17. AutoSSA: trend extraction Let us investigate every eigenvector U j and take U = ( u 1 , . . . , u L ) T . Low Frequencies method � � � + ( − 1) n c L/ 2 , u n = c 0 + c k cos(2 πnk/L ) + s k sin(2 πnk/L ) � 1 � k � L − 1 2  � Periodogram:  2 c 02 , k = 0 ,   c k 2 + s k 2 , U ( k/L ) = L Π L 1 � k � L − 1 , 2  2   2 c L/ 22 , L is even and k = L/ 2 . Π L U ( ω ) , ω ∈ { k/L } , reflects the contribution of a harmonic with the frequency ω into the Fourier decomposition of U . � Parameter: ω 0 – upper boundary for the “low frequencies” interval � 0 � k � L ω 0 Π L U ( k / L ) C ( U ) = � U ( k / L ) – contribution of LF frequencies. 0 � k � L / 2 Π L C ( U ) � C 0 ⇒ eigenvector U corresponds to a trend , where C 0 ∈ (0 , 1) – threshold “Usually” C 0 = 0 . 1 “Caterpillar”-SSA and its automatization – p. 11/15

Recommend


More recommend