Time series, Part 2 Linear stochastic processes Linear time series (stochastic process) X t i Z t i | Zt ~ WN(0, ) i X t is stationary E[X t ] 0 2 Z X ( ) i i i i 2 We assume μ=0 If i 0 for i<0 1 2 B i X t 1 X t 1 2 X t 2 | i 0 i Zt Xt Zt i Z t i autoregressive process AR(∞) i 1 X t is invertible 1 Zt ( B) moving average process MA(∞) i 1 Z t i X t i Z t | ( B) X t Z t X t ( B) i i A linear time series is expressed as … X t Zt 1Zt 1 2 Zt 2 Linear filter ( B) | ? Considering the lag operator : X t ( B) Z t i ( B) 1 ( B) the stochastic component Z t can be expressed in terms of the current and past observations X t Autoregressive processes 2 X t 1 X t 1 2 X t 2 Z t i X t i Z t autoregression AR(∞) i 1 We restrict the autoregression to the first p most recent terms X t 1 X t 1 2 X t 2 Zt ~ WN(0, Z2 ) Zt Autoregression process of order p, AR(p) X t 1 X t 1 2 X t 2 (1 1B 2 B2 p X t p Zt p B p ) X t Zt ( B) 1 1B 2 B 2 p B ( B) X t Z t p p ( B ) 1 i B i characteristic polynomial i 1 Condition of stationarity The roots of ( ) 0 must be outside the unit circle or p p 1 the roots of 1 p 1 p 0 must be inside the unit circle Autoregressive process of order one, AR(1) Zt ~ WN(0, Z2 ) X t X t 1 Zt Stationarity condition: | | 1 X t t t 1 t 2 2 Successive backward substitutions: 2 Var[X t ] (1 ) 1 2 i 0 2 2 4 2 2i i t i i 0 2 X Autocorrelation? (we assume stationarity) X t 1 X t X t 1 X t 1 Zt E[ X t 1 X t ] E[ X t 1 X t 1 ] E[ X t 1Zt ] X (1) X2 X (1) X t X t X t X t 1 Zt E[ X t X t ] E[ X t X t 1 ] E[ X t Zt ] X ( ) X ( 1) X ( ) () 1 0.5 0.5 () () () 1 0 -0.5 -1 0 0 -0.5 2 4 6 0.8 8 10 -1 0 2 4 6 0.8 8 10 Autoregressive process of order two, AR(2) X t 1 X t 1 2 X t 2 Zt Zt ~ WN(0, Z2 ) Stationarity condition The roots of ( B) 1 1B 2 B 2 must be outside the unit circle The roots of 2 1 2 0 must be inside the unit circle Ρίζες: 1,2 1 42 22 2 1 1,2 2 1 1 1 2 1 1 1 1 2 ? Stationarity condition for AR(2) 3 two real roots: 12 42 0 real distinct roots complex roots real single root 2 2 1 one double real root: 12 42 0 0 -1 complex conjugate roots: 12 42 0 -2 -3 -3 -2 -1 0 1 1 2 3 Autocorrelation (α) λ1=0.8+0.5i λ2=0.8-0.5i ( ) () 0 -0.5 0.5 -0.5 0 5 10 15 -1 20 0 5 10 15 2=-0.89 -0.5 15 20 5 10 15 2=0.76 0.5 20 1=-1.75 2=-0.76 0.5 0 -1 15 () 1=-0.15 20 10 (η) λ1=-0.8 λ2=-0.95 0 -0.5 -0.5 0 5 1 2=-0.64 0 -1 0 1=-1.6 -0.5 -1 20 () () 0 15 1 0.5 () 0.5 10 (στ) λ1=0.8 λ2=-0.95 () 1=-1.6 10 5 1 5 0 (δ) λ1=-0.8 λ2=-0.8 () 0 -0.5 1 0 0 -1 20 2=0.76 0.5 -0.5 (β) λ1=-0.8+0.5i λ2=-0.8-0.5i () 2=-0.76 0 -1 1=1.75 2=-0.64 () -1 1=0.15 1=1.6 () 0.5 () () 0.5 1 1 1=1.6 2=-0.89 () () 1 () 1 (ζ) λ1=-0.8 λ2=0.95 (ε) λ1=0.8 λ2=0.95 (γ) λ1=0.8 λ2=0.8 0 5 10 15 20 -1 0 5 10 15 20 Autoregressive process of order two, AR(2) Autocorrelation ? (we assume stationarity) X t 1 X t X t 1 1 X t 1 2 X t 2 Zt 1 1 2 1 E[ X t 1 X t ] 1E[ X t 1 X t 1 ] 2 E[ X t 1 X t 2 ] E[ X t 1Zt ] X (1) 1 X2 2 X (1) X (1) 1 2 X (1) 2 11 2 X t 2 X t X t 2 1 X t 1 2 X t 2 Zt E[ X t 2 X t ] 1E[ X t 2 X t 1 ] 2 E[ X t 2 X t 1 ] E[ X t 2 Zt ] X (2) 1 X (1) 2 X2 X (2) 1 X (1) 2 Για υστέρηση τ: 1 X ( ) 1 X ( 1) 2 X ( 2) 1 1 2 2 can be computed recursively variance 1 1 2 1 12 2 2 1 2 (1 1B 2 B2 ) 0 1 (1 2 ) 1 12 2 12 2 1 12 real roots: exponential decay characteristic polynomial complex roots: decaying harmonic function X t X t X t 1 X t 1 2 X t 2 Zt X2 1 X (1) 2 X (2) Z2 Z2 1 1 1 2 2 2 X Autoregressive process of order p, AR(p) X t 1 X t 1 2 X t 2 (1 1B 2 B2 p X t p Zt Zt ~ WN(0, Z2 ) p B p ) X t Zt Stationarity condition 2 Roots of ( B) 1 1B 2 B p B p must be outside the unit circle Autocorrelation ? (we assume stationarity) For lag τ: X t X t X t 1 X t 1 2 X t 2 p X t p Zt E[ X t X t ] 1E[ X t X t 1 ] 2 E[ X t X t 2 ] X ( ) 1 X ( 1) 2 X ( 2) X ( ) 1 X ( 1) 2 X ( 2) p E[ X t p X t 2 ] E[ X t p Zt ] p X ( p) p X ( p) ( B) 0 real roots : exponential decay complex roots : decaying harmonic function Autoregressive process of order p, AR(p) 1 1 2 2 1 2 1 1 2 1 1 p p 1 p 1 2 p 2 1 2 p 2 1 2 1 2 p p p p p p 1 p p 2 p 1 1 1 1 p p 1 p 2 Yule-Walker equations 2 1 p 1 p 2 p 3 1 p p p1p Variance X t X t X t 1 X t 1 2 X t 2 1 X (1) 2 X (2) 2 X p X t p Zt Z2 p X ( p ) 1 11 2 2 2 Z 2 X p p Partial autocorrelation 1 1 1 Yule-Walker 1 equations k 1 k 2 11 1 k 1 1 k 2 22 1 1 1 1 k 3 1 33 2 1 1 2 2 1 k 1 1 k 2 2 k 3 1 1 2 2 12 1 1 12 1 1 1 1 2 1 3 1 2 1 1 1 1 1 2 k k For each k we compute the coefficient k kk 1 1 kk k 1 1 1 1 1 k 1 k 2 1 1 k 2 2 1 k 3 2 1 k 2 k 3 1 1 k k 3 k 2 1 1 partial autocorrelation for lag (order) k Recursive algorithm of Durbin-Levinson the coefficients of AR(p) p1 , p 2 , 2 k 2 k 1 k 3 , pp , are computed recursively, and for each order k the coefficients are computed from the coefficients of order k-1 Partial autocorrelation (α) λ1=0.8+0.5i λ2=0.8-0.5i ( ) () 1=1.6 2=-0.89 2=-0.64 0.5 -0.5 0 -0.5 5 10 15 -1 20 0 5 10 15 -1 -0.5 10 15 20 15 15 20 20 () 1 1=-1.75 2=0.76 2=-0.76 0.5 0 -1 10 10 (η) λ1=-0.8 λ2=-0.95 -0.5 5 5 0.5 (,) 0 0 0 1=-0.15 2=-0.64 0.5 -1 -1 20 () -0.5 5 15 1=-1.6 (,) 0 10 1 1=-1.6 0 5 (στ) λ1=0.8 λ2=-0.95 () 0.5 0 1 2=-0.89 0 -0.5 (δ) λ1=-0.8 λ2=-0.8 () (,) 20 2=0.76 0.5 0 1 -1 0.5 -0.5 (β) λ1=-0.8+0.5i λ2=-0.8-0.5i 1=0.15 2=-0.76 (,) 0 1 1=1.75 (,) 0 () 1 1=1.6 (,) (,) 0.5 -1 () 1 (,) 1 (ζ) λ1=-0.8 λ2=0.95 (ε) λ1=0.8 λ2=0.95 (γ) λ1=0.8 λ2=0.8 0 -0.5 0 5 10 15 20 -1 0 5 10 15 20 Moving average processes 1 X t Zt 1Zt 1 2 Zt 2 Zt i Z t i i 1 moving average MA(∞) We constrain the white noise terms to the first q most recent terms X t Zt 1Zt 1 2 Zt 2 Zt ~ WN(0, Z2 ) Moving average process of order q, ΜΑ(q) X t Zt 1Zt 1 2 Zt 2 X t (1 1B 2 B2 ΜΑ(q) is stationary q Zt q q B q ) Zt i i X t ( B) Z t ? ΜΑ(q) is invertible if Zt 1 ( B) X t invertibility condition The roots of ( ) 0 must be outside the unit circle ( B) 1 1B 2 B2 q Bq characteristic polynomial Moving average process of order one, MA(1) X t Zt Zt 1 Invertibility condition: | | 1 Zt ~ WN(0, Z2 ) X t X t Zt Zt 1 Zt Zt 1 ... X2 (1 2 ) Z2 ? X t 1 X t Zt 1 Zt 2 Zt Zt 1 ... X (1) Z2 1 X t 2 X t Zt 2 Zt 3 Zt Zt 1 ... X (2) 0 1 2 0 Example X t Zt 0.4Zt 1 X t Zt 2.5Zt 1 1 2 1 2 | 1 | 1/ 2 ? For one 1 there are two solutions for θ and only one satisfies the invertibility condition X t Zt Zt 1 1 1 2.9 2 0 and X t Zt 1/ Zt 1 they have the same autocorrelation If the root of 1 B 0 is outside the unit circle the root of 1 1/ B 0 is inside the unit circle Moving average process of order one, MA(1) Partial autocorrelation 0.8 11 1 1 2 13 3 3,3 1 2 12 1 2 4 6 0.5 0.5 0 -0.5 -1 0 -0.5 0 2 4 6 8 -1 10 0 2 4 - ρτ of ΜΑ(1) decays the same as ϕττ of AR(1) 10 6 8 10 () 1 1 0.5 0.5 0 -0.5 -1 0 -0.5 0 2 4 6 - … but for MA(1), ρτ and ϕττ are always ≤0.5 8 0.8 () (,) - ϕττ of ΜΑ(1) decays the same as ρτ of AR(1) 6 1 () , (1 2 ) , 1 2( 1) () 1 (,) 12 2 1 12 1 2 4 () 2,2 ( ) 1 8 10 -1 0 2 4 Moving average process of order two, MA(2) X t ( B)Zt Zt 1Zt 1 2 Zt 2 , ( B) 1 1B 2 B2 Zt ~ WN(0, Z2 ) characteristic polynomial MA(2) is always stationary MA(2) is invertible if the roots of θ(Β) are outside the unit circle Variance X2 (1 12 22 ) Z2 Autocorrelation 1 (1 2 ) 1 2 2 1 1 2 2 2 2 2 1 1 2 0 2 Partial autocorrelation 11 1 2 12 2,2 1 12 13 12 (2 2 ) 3,3 1 22 2 12 (1 2 ) , ... complicated expression λ1=0.8 λ2=0.95 λ1=-0.8+0.5i λ2=-0.8-0.5i λ1=0.8+0.5i λ2=0.8-0.5i λ1=0.8 λ2=-0.95 Autocorrelation ( ) () 1=1.6 0.6 1=-1.6 0 0.2 0 0 -0.2 -0.2 -0.2 -0.4 -0.4 -0.4 -0.4 -0.6 -0.6 -0.6 -0.6 -0.8 -0.8 -0.8 5 10 15 20 0 5 10 15 20 2=0.76 0.4 -0.2 0 1=-0.15 0.6 2=-0.76 0.2 () () 1=1.75 0.4 0.2 0 0.8 0.6 2=-0.89 0.4 0.2 () 0.8 0.6 2=-0.89 0.4 () () 0.8 () 0.8 0 5 10 15 -0.8 20 0 5 10 15 20 Partial autocorrelation () () 1=1.6 0.6 1=-1.6 0 0.2 0 0 -0.2 -0.2 -0.2 -0.4 -0.4 -0.4 -0.4 -0.6 -0.6 -0.6 -0.6 0 5 10 15 20 - ϕττ of ΜΑ(2) decays same as ρτ of AR(2) -0.8 0 5 10 15 20 -0.8 2=0.76 0.4 -0.2 -0.8 1=-0.15 0.6 2=-0.76 0.2 (,) (,) 1=1.75 0.4 0.2 0 0.8 0.6 2=-0.89 0.4 0.2 () 0.8 0.6 2=-0.89 0.4 (,) () 0.8 (,) 0.8 0 5 - ρτ of ΜΑ(2) decays same as ϕττ of AR(2) 10 15 20 -0.8 0 5 10 15 - … but for MA(2), ρτ and ϕττ is always ≤0.5 20 Moving average process of order q, MA(q) X t ( B)Zt Zt 1Zt 1 2 Zt 2 ( B) 1 1B 2 B2 Variance q Bq q Zt q Zt ~ WN(0, Z2 ) characteristic polynomial X2 (1 12 q2 ) Z2 Autocovariance Z2 ( 1 1 0 q q ) 1, 2, , q q Autocorrelation 1 1 q q 1 12 22 q2 0 1, 2, , q q The partial autocorrelation decays in a way that is determined from the roots of the characteristic polynomial The expressions of ϕττ in terms of the coefficients θ1, θ2, ..., θq are complicated Relation between AR and MA processes Autoregressive process order p, AR(p) X t 1 X t 1 2 X t 2 p X t p Zt (1 1B 2 B2 p B p ) X t Zt ( B) 1 1B 2 B2 ( B) X t Z t Zt ~ WN(0, Z2 ) p B p Moving average process of order q, ΜΑ(q) X t Zt 1Zt 1 2 Zt 2 q Zt q X t (1 1B 2 B2 q Bq )Zt ( B) 1 1B 2 B2 X t ( B) Z t AR(p) stationary MA(q) invertible X t ( B) Zt ( B) 1 1 B 2 B 2 such that ( B) ( B) 1 1 X t ( B)Zt ΜΑ(∞) AR(p) ↔ MA(∞) Wold's decomposition (1) every covariance-stationary time series can be written as an infinite moving average (MA(∞)) process of its innovation process. q Bq ( B) 1 X t Zt ( B) 1 1 B 2 B 2 such that ( B) ( B) 1 ( B) X t Z t AR(p) and MA(q) have dual relation AR(∞) MA(q) ↔ AR(∞) The autocorrelation and partial autocorrelation of AR(p) and MA(q) have also dual relation AR(p): ρτ decays exponentially to 0, ϕττ gets zero for τ>p MA(q): ϕττ decays exponentially to 0, ρτ gets zero for τ>q Autoregressive moving average process ARMA(p,q) Autoregressive process of order p, AR(p) X t 1 X t 1 2 X t 2 p X t p Zt Moving average process of order q, ΜΑ(q) Zt ~ WN(0, Z2 ) X t Zt 1Zt 1 2 Zt 2 q Zt q X t 1 X t 1 2 X t 2 p X t p Zt 1Zt 1 2 Zt 2 q Zt q X t 1 X t 1 2 X t 2 p X t p Zt 1Zt 1 2 Zt 2 q Zt q ( B) X t ( B) Z t Xt ( B) Zt ( B) ( B) X t Zt ( B) Stationarity is determined by the AR part Invertibility is determined by the MA part Autocorrelation: X t X t X t (1 X t 1 2 X t 2 X ( ) 1 X ( 1) p X t p Zt 1Zt 1 2 Zt 2 q Zt q ) p X ( p) E[ X t t ] 1E[ X t t 1 ] For q 1 1 Για q mixing of autocorrelation for AR(p) and MA(q) p p 1 1 p p q E[ X t t q ] such as for AR(p) Process ARMA(1,1) X t X t 1 Zt Zt 1 (1 B) Zt (1 B) Stationarity condition: | | 1 (1 B) X t (1 B)Zt Autocorrelation: Xt Invertibility condition: | | 1 X t X t X t ( X t 1 Zt Zt 1 ) 1 E[ X t t ] E[ X t t 1 ] 0 1 1 2 1 2 2 2 2 2 Z 0 X 0 1 Z ( ) Z 2 1 ( )(1 ) 2 1 0 Z2 1 Z 2 1 1 such as for AR(1) ? ( )(1 ) 1 2 1 2 1 2 Partial autocorrelation : decays with the lag such as for MA(1) An ARMA(p,q) process with small p,q, exhibits correlation pattern (ρτ and ϕττ) that can be attained by AR(p) only for large order p, or by MA(q) only for large order q Estimation of models AR, MA, ARMA (stationary) time series (stochastic process) X t t mean value μ (stationary) time series of n observations autocovariance X t X t X t 2 autocorrelation ( ) , xn 1 n x xt n t 1 sample mean value ( ) ( X t )( X t ) x1 , x2 , sample autocovariance 1 n c c( ) ( xt xt x 2 ) n t 1 0,1, , n 1 sample autocorrelation ( ) (0) r r ( ) c( ) c(0) stochastic process AR(p) Estimation of the process (model) X t 1 X t 1 2 X t 2 ● AR, MA or ARMA ? other model ? p X t p Zt Zt ~ WN(0, Z2 ) stochastic process MA(q) X t Zt 1Zt 1 2 Zt 2 q Zt q stochastic process ARMA(p,q) X t 1 X t 1 2 X t 2 p X t p Zt 1Zt 1 2 Zt 2 q Zt q ● order p or/and q ? ● estimation of model parameters ? AR( p) :1 , 2 , , p , 2 ΜΑ(q) :1 ,2 , ,q , 2 ARΜΑ( p, q) :1 , 2 , ? , p ,1 ,2 , ,q , 2 Estimation of model AR(p) We assume a stochastic process AR(p) generate the time series estimation of parameters 1 , 2 , Fit of a model AR(p) x1 , x2 , , xn , p , 2 Method of moments or method of Yule-Walker (YW) Estimation of the parameters from the sample autocorrelations r1 , r2 , , rp , sX2 ˆ , ˆ , , ˆ , s 2 1 1 2 1 1 1 Yule-Walker 1 equations p 1 p 2 p 3 Estimation of 1 , 2 , , p , X2 r1 1 r 1 1 rp 1 rp 2 ˆ Rp1rp r2 r1 rp 3 2 p 1 p p 2 2 Z 2 X 1 11 2 2 p p 1 p p r1 , r2 , , rp , sX2 and then substitution … p 1 1 p 2 2 rp 1 ˆ1 r1 rp 2 ˆ2 r2 1 ˆ rp p sZ2 sX2 (1 ˆ1r1 ˆ2 r2 Rpˆ rp sZ2 s 1 ˆ1r1 ˆ2 r2 2 X ˆp rp ) ˆp rp x1 , x2 , , xn with a mean μ X t 1 ( X t 1 ) 2 ( X t 2 ) General form of AR(p) p ( X t p ) Zt The estimation method of ordinary least squares (OLS) Fit of model AR(p) to the data Minimization of the sum of squares of the fitting errors min S ( , 1 , p ) min ˆ , ˆ , ˆ , 1 2 , ˆ p n t p 1 xt 1 ( xt 1 ) p ( xt p ) zˆt ( xt ˆ ) ˆ1 ( xt 1 ˆ ) 2 w.r.t. , 1 , 2 , , p ˆ p ( xt p ˆ ), t p 1, n 1 s zˆt2 n p t p 1 2 Z ˆ ,n 1 n x xt n t 1 X t 1 ( X t 1 ) Zt AR(1) n S ( , 1 ) xt 1 ( xt 1 ) 2 t 2 x(2) ˆ x(1) ˆ 1 ˆ ˆ t 2 n 1 n x(1) xt 1 n 1 t 2 ( xt ˆ )( xt 1 ˆ ) 2 ˆ ( x ) t t 2 n n t 2 x(2) 1 n xt n 1 t 2 ( xt x )( xt 1 x ) n 2 ( x x ) t t 2 ˆ x 1 n ( xt x )( xt 1 x ) n t 2 1 n ˆ c0 t 1 ( xt x )2 n c1 c1 r1 c0 Other methods for estimation of AR(p) ● backward – forward approach (FB) ● Burg’s algorithm (Burg) ● maximum likelihood (ML) - conditioned - unconditioned The ML estimation is optimal, the other methods approximate it The ML reduces to OLS when the time series is from a Gaussian process Asymptotically (for large n) all methods converge to the same (ML) estimates The YW has the slowest convergence rate to ML Determination of order p of an AR model the criterion of partial autocorrelation correlation of partial autocorrelation for lag τ: xt , xt 1 , accounting for the correlation with x x z t 1,1 t 1 , xt 1 , xt t xt 1,2 xt 1 2,2 xt 2 zt xt 1,3 xt 1 2,3 xt 2 3,3 xt 3 zt The order is p if ˆp , p 0 and ˆ , 0 estimation of for model AR(τ) for τ>p (fall from non-zero to zero partial autocorrelation) the criterion based on fitting errors ● Akaike information criterion (AIC) ● Bayesian information criterion (BIC) ● Final prediction error (FPE) 2p n p ln(n) BIC( p) ln( sz2 ) n AIC( p) ln( sz2 ) FPE( p) sz2 n p n p Growth rate of gross national product (GNP) of USA Παράδειγμα quarter-annual observations, 2nd quarter 1947 – 1st quarter 1991). The time series is corrected for seasonality GNP of USA: increments incr.GNO(USA): autocorrelation 0.04 0.5 0.03 0.4 stationary ? 0.3 0.02 0.2 x t () 0.01 0 correlation ? 0 -0.01 -0.1 -0.02 -0.03 0.1 -0.2 0 50 100 0 150 5 10 15 20 t incr.GNO(USA): AIC incr.GNO(USA): partial autocorrelation 0.5 -9.16 0.4 -9.18 order of AR model ? 0.3 -9.2 AIC(p) , 0.2 0.1 0 -9.22 AR(3) ? -9.24 -0.1 -9.26 -0.2 0 5 10 15 20 0 5 10 p 15 20 parameter estimation ˆ 0.0077 xt ˆ xt OLS ˆ1 0.35 ˆ2 0.18 ˆ3 0.14 ˆ0 ˆ 1 ˆ1 ˆ2 ˆ3 0.0047 t 4, estimation xˆt 0.0047 0.35xt 1 0.18xt 2 0.14 xt 3 ,176 sz2 ˆ z2 0.0000989 zˆt xt xˆt errors or residual of fit xt 0.0047 0.35xt 1 0.18xt 2 0.14 xt 3 zt fitted AR(3) incr.GNP(USA): AR(3) fit sz ˆ z 0.0098 incr.GNP(USA): AR(3) fit 0.03 0.03 0.02 0.02 0.01 0.01 x(t) 0.04 x(t) 0.04 0 0 -0.01 -0.01 -0.02 -0.02 -0.03 0 50 100 time t 150 200 -0.03 100 110 120 time t Diagnostic check for model adequacy Is the residual time series independent? test for independence on 130 140 zˆt t p1 n Fit of the model MA(q) stochastic process AR(p) Estimation of the process (model) X t 1 X t 1 2 X t 2 ● AR, MA or ARMA ? other model ? p X t p Zt Zt ~ WN(0, Z2 ) stochastic process MA(q) X t Zt 1Zt 1 2 Zt 2 q Zt q stochastic process ARMA(p,q) X t 1 X t 1 2 X t 2 p X t p Zt 1Zt 1 2 Zt 2 q Zt q ● order p or/and q ? ● estimation of model parameters ? AR( p) :1 , 2 , , p , 2 ΜΑ(q) :1 ,2 , ,q , 2 ARΜΑ( p, q) :1 , 2 , We assume a stochastic process MA(q) for the time series Fit of the process (model) MA(q) x1 , x2 , ? , p ,1 ,2 , , xn parameter estimation 1 ,2 , ,q , 2 ,q , 2 MA(q) X t Zt 1Zt 1 2 Zt 2 Method of moments Variance X2 (1 12 q2 ) Z2 Nonlinear equation system w.r.t. the parameters 1 ,2 , Autocorrelation 1 1 q q 1 12 22 q2 0 Estimation of 1 , 2 , 1, 2, , q ,q q , q , X2 r1 , r2 , , rq , s X2 Innovation algorithm Method of ordinary least squares Fit of model MA(q) to the data Minimization of sum of squares of fitting errors min S ( ,1 , q ) min n (x z t q 1 t 1 t 1 q zt q ) 2 Numerical optimization method ˆ1 ,ˆ2 , ,ˆq w.r.t. ,1 ,2 , ,q q Zt q X t Zt Zt 1 MA(1) Method of moments 1 2 0 1 2 (1 ) 2 X 2 q 2 Z r | r1 | 0.5 ˆ 1 | r1 | 2 1 1 4 r 1 2 | r1 | 0.5 r1ˆ ˆ r1 0 ˆ1,2 2r1 We choose the solution | ˆ | 1 that gives rise to invertibility s X2 s 1 ˆ 2 2 Z Method of ordinary least squares 2n-2 solutions, we select the solution | ˆ | 1 We assume z0 0 (and 0 ) zt xt zt 1 z1 x1 z2 x2 z1 x2 x1 computational algorithm: least squares with constraints for invertibility z3 x3 z2 x3 ( x2 x1 ) x3 x2 2 x1 zn xn zn1 xn xn1 2 xn2 n n2 x2 n1 x1 min zt2 min x12 ( x2 x1 )2 ( x3 x2 x1 2 )2 t 1 min a0 a1 a2 n2 2n2 ( xn xn1 x1 n1 )2 Growth rate of gross national product (GNP) of USA Παράδειγμα quarter-annual observations, 2nd quarter 1947 – 1st quarter 1991). The time series is corrected for seasonality GNP of USA: increments incr.GNP(USA): autocorrelation 0.04 0.5 0.03 0.4 0.3 0.02 order of the MA model ? 0.2 x t r() 0.01 0 0 -0.01 -0.1 -0.02 -0.2 0 50 100 0 150 5 10 15 20 t incr.GNP(USA): AIC of MA models -9.14 -9.16 AIC(q) -0.03 0.1 -9.18 ΜΑ(2) ? -9.2 -9.22 -9.24 0 2 4 6 q 8 10 parameter estimation x 0.0077 OLS ˆ1 0.312 ˆ2 0.272 sz2 0.000097 variance of errors (residuals) sz 0.00983 xt 0.0077 zt 0.312 zt 1 0.272 zt 2 fitted ΜΑ(2) incr.GNP(USA): MA(2) fit t 1, ,176 incr.GNP(USA): MA(2) fit 0.04 0.03 0.03 0.02 0.02 0.01 0.01 fit with ΜΑ(2) x(t) x(t) 0.04 0 0 -0.01 -0.01 -0.02 -0.02 -0.03 0 50 100 time t 150 -0.03 100 200 incr.GNP(USA): AR(3) fit 120 time t 130 140 incr.GNP(USA): AR(3) fit 0.04 0.03 0.03 0.02 0.02 0.01 0.01 fit with AR(3) x(t) x(t) 0.04 0 0 -0.01 -0.01 -0.02 -0.02 -0.03 110 0 50 100 time t 150 200 -0.03 100 110 120 time t 130 140 Diagnostic check for model adequacy Is the residual time series independent? test for independence on zˆt t p1 n Εκτίμηση μοντέλου ARMA(p,q) stochastic process AR(p) Estimation of the process (model) X t 1 X t 1 2 X t 2 ● AR, MA or ARMA ? other model? p X t p Zt Zt ~ WN(0, Z2 ) stochastic process MA(q) X t Zt 1Zt 1 2 Zt 2 q Zt q ● order p or/and q ? ● estimation of model parameters ? stochastic process ARMA(p,q) X t 1 X t 1 2 X t 2 p X t p Zt 1Zt 1 2 Zt 2 AR( p) :1 , 2 , , p , 2 ΜΑ(q) :1 ,2 , ,q , 2 ARΜΑ( p, q) :1 , 2 , q Zt q , p ,1 ,2 , We assume a stochastic process ARMA(p,q) for the time series x1 , x2 , Fit of the process (model) ARMA(p,q) estimation of parameters 1 , 2 , , p ,1 ,2 , ,q , 2 The methods of moments and least squares as for MA(q) ? , xn ,q , 2 ARMA(1,1) X t ( X t 1 ) Zt Zt 1 Method of moments ( )(1 ) 1 1 2 2 1 2 1 2 2 2 2 X Z 2 1 Estimation of 1 , 2 , , X2 r1 , r2 , Solution of equation system w.r.t. , 1 ˆ 2 2 s s X ˆˆ 1 ˆ 2 2 2 Z Method of ordinary least squares We assume z0 0 (and x0 0 ) z1 x1 z2 x2 x1 z1 x2 ( ) x1 n min zt2 t 1 z3 x3 x2 z2 x3 ( ) x2 ( ) x1 zn xn xn1 zn1 xn ( ) xn1 ( ) xn2 computational algorithm of least squares with constraints for invertibility and stationarity n2 ( ) x1 , rp , sX2 ? Growth rate of gross national product (GNP) of USA Παράδειγμα quarter-annual observations, 2nd quarter 1947 – 1st quarter 1991). The time series is corrected for seasonality GNP of USA: increments incr.GNP(USA): autocorrelation 0.04 0.5 0.03 0.4 0.3 0.02 0.2 x t r() 0.01 0 0 -0.01 -0.1 -0.02 -0.03 0.1 -0.2 0 50 100 0 150 5 10 20 t incr.GNP(USA): AIC of ARMA models incr.GNP(USA): partial autocorrelation -9.14 0.5 q=0 q=1 q=2 q=3 q=4 q=5 0.4 -9.16 0.3 AIC(p,q) 0.2 p,p 15 0.1 0 -0.1 -9.18 order of ARMA ? ARMA(2,2) ? -9.2 -9.22 -0.2 0 2 4 6 p 8 10 -9.24 -1 0 1 2 3 p 4 5 6 parameter estimation x 0.0077 OLS ˆ1 0.614 ˆ2 0.455 variance of errors (residuals) fitted ARΜΑ(2,2) ˆ1 0.301 ˆ2 0.600 sz2 0.000097 sz 0.00983 xˆt 0.0065 0.614 xt 1 0.455xt 2 zt 0.301zt 1 0.600 zt 2 incr.GNP(USA): ARMA(2,2) fit incr.GNP(USA): ARMA(2,2) fit 0.03 0.03 0.02 0.02 0.01 0.01 fit with ARΜΑ(2,2) x(t) 0.04 x(t) 0.04 0 0 -0.01 -0.01 -0.02 -0.02 -0.03 0 50 100 time t 150 -0.03 100 200 incr.GNP(USA): MA(2) fit 110 120 time t 130 140 incr.GNP(USA): MA(2) fit 0.03 0.03 0.02 0.02 0.01 0.01 fit with ΜΑ(2) x(t) 0.04 x(t) 0.04 0 0 -0.01 -0.01 -0.02 -0.02 -0.03 0 50 100 time t 150 -0.03 100 200 incr.GNP(USA): AR(3) fit 110 120 time t 130 140 incr.GNP(USA): AR(3) fit 0.03 0.03 0.02 0.02 0.01 0.01 fit with AR(3) x(t) 0.04 x(t) 0.04 0 0 -0.01 -0.01 -0.02 -0.02 -0.03 0 50 100 time t 150 t 1, 200 -0.03 100 110 120 time t 130 140 ,176 Model for time series with trends (ARIMA) Yt t 1 random walk Yt Yt 1 X t X1 X 2 process AR(1) for Xt (non-stationary process) X t t iid E X t 0 1 First differences: X t (1 B)Yt Yt Yt 1 Yt t 1 E X t2 2 iid process non-stationary process that exhibits trends first differences: X t Yt Yt 1 ? stationary process YES NO second order differences: X t X t X t 1 Yt 2Yt 1 Yt 2 X t d Yt X t t 1 stationary after d order differences: (1 B)d Yt Usually d 1 stationary process YES NO AR(p), MA(q), ARMA(p,q) Yt t 1 non-stationary process ARIMA(p,d,q) X t 1 X t 1 2 X t 2 d The polynomial ( B)(1 B) has a unit root and all other roots are outside the unit circle ? p X t p Zt 1Zt 1 2 Zt 2 ( B) X t ( B) Z t ( B)d Yt ( B)Zt ( B)(1 B)d Yt ( B)Zt ? q Zt q Fit of model ARIMA (Box-Jenkins approach) time series y1 , y2 , , yn time series history diagram autocorrelation (strong and slowly decaying) other ? if autocorrelation indication that there is trend decays to zero d d-order differences xt (1 B) yt other ? the time series is stationary stationary time series x , x , , x ? 1 2 n model order estimation of model parameters if the autocorrelation is statistically not significant fit of model AR(p), MA(q), ARMA(p,q) ? diagnostic test model adequacy model ARMA(p,q) for x1 , x2 , is it iid ? YES , xn test for independence STOP d then using the inverse transform xt (1 B) yt we get the model ARΙMA(p, d,q) for y1 , y2 , , yn ? NO nonlinear model ? prediction ? Annual index of global temperature (anomaly of surface temperature of the north hemisphere at grid 5ο x 5ο), time period 1850-2011 Example y1 , y2 , , yn real observations Source: http://www.cru.uea.ac.uk/cru/data/temperature annual global temperature: autocorrelation 0.8 1 0.6 0.5 0.4 stationary time series? r() global temperature annual land air temperature anomalies 1.5 0 0.2 -0.5 0 -1 1840 1860 x1 , x2 , 1880 1900 1920 1940 year 1960 1980 2000 -0.2 2020 NO 0 5 10 15 , xn first differences first differences of annual land air temperature anomalies first difference of annual global temperature: autocorrelation 0.3 1 0.2 0.1 stationary time series? 0 r() d(temp) 0.5 0 -0.1 -0.2 -0.5 YES -0.3 -0.4 -1 1840 1860 1880 1900 1920 1940 year 1960 1980 2000 2020 -0.5 0 5 10 15 Model for time series x1 , x2 , , xn ? partial autocorrelation autocorrelation first difference of annual global temperature: autocorrelation 0.3 0.2 diff of temp: partial autocorrelation -2.85 0.2 0.1 -2.9 0.1 diff of temp: AIC of ARMA models -0.2 -0.1 -0.2 -3.05 -3.15 -0.4 -0.4 0 5 10 -0.5 15 -3.2 0 5 10 15 0 1 Model for time series y1 , y2 , 2 3 4 5 6 p , yn ARIΜΑ(0,1,4) (1 B)Yt 4 ( B) Zt fit of ΜΑ(4) ( x 0.008) xt 0.008 zt 0.758zt 1 0.022 zt 2 0.219 zt 3 0.275zt 4 diff of global temperature: ARMA(0,4) fit sz2 0.0414 sz 0.2035 diff of global temperature: ARMA(0,4) fit 1 0.5 0.5 x(t) 1 0 -0.5 -1 1840 -3.25 -1 The most appropriate model ? x(t) -3 -3.1 -0.3 -0.3 -0.5 AIC(p,q) () -0.1 q=0 q=1 q=2 q=3 q=4 q=5 -2.95 0 0 r() AIC criterion 0 -0.5 1860 1880 1900 1920 1940 time t 1960 1980 2000 2020 -1 1930 1935 1940 1945 time t 1950 1955 1960 Model of time series with seasonality (ARMAs) Given the time series y1 , y2 , , yn without trend and with seasonality (periodicity) Removal of seasonality of period s, k n / s : k 1 si yi js xt yt st Estimation of the periodic components si i=1,…,s k j 1 1 Symmetric moving xt (0.5 yt s /2 yt s /21 yt s /21 0.5 yt s /2 ) s even s average of order s 1 ( s 1)/2 xt yt i s odd s i ( s 1)/2 s – differences (difference of lag s) X t sYt (1 B s )Yt Yt Yt s Given the time series x1 , x2 , , xn without trend and with seasonality s Hypothesis: there are correlations but only between the same components of each period (the dependence occurs at time steps s) : x1 , x2 , x3 , , xs , xs 1 , xs 2 , , x2 s , x2 s 1 , x2 s 2 , , x3s , x3s 1, x3s 2 , xn k cycles of period s model ARMA(P,Q)s for i 1, 2, X i st 1 X i s (t 1) ,s t Ps 1, Ps 2, xi , xis , xi2s , ,n X t 1 X t s , xi ks the same for i 1, 2, P X i s ( P1) Zi st 1Zi s (t 1) P X t Ps Zt 1Zt s ( B s ) X t ( B s )Zt Q Zt Qs model ARMA(P,Q)s for ,s Q Zi s (Q 1) x1 , x2 , , xn Model of time series with seasonality (ARIMAs) This is an extension of ARMA(P,Q)s when the time series has “seasonal trend”, meaning trend at the time points t, t+s, t+s, … Given the time series y1 , y2 , , yn with seasonal trend and given that the correlations are between components with the same periodic order s – differences (difference of lag s) y1 , y2 , , yn xs1 , xs2 , , xn xt s yt (1 B ) yt yt yt s s ARMA(P,Q)s ARIMA(P,1,Q)s ( B s )(1 B s )Yt ( B s )Zt ( B s ) X t ( B s )Zt In general, ARIMA(P,D,Q)s ( B s )(1 B s ) D Yt ( B s )Zt Example Mean monthly temperature at Thessaloniki station, period 1930-2000 part of the record x1 , x2 , , xn n 71*12 852 Estimation of seasonal component Temp Thess: subtract average period strong seasonality 6 YR JAN FEB MAR APR MAY JUN ΙJUL AUG SEP OCT NOV DEC (periodicity) 4 1930 6,7 6,7 11,3 15,7 19 22,6 26,2 26 22,8 17,5 12,1 8,9 2 Temperature Thessaloniki: autocorrelation Temp 4,1 9,2 5,5 10 9,2 7,2 8 8,8 7,7 4,2 4,8 7,3 1 -4 0 -6 30 -0.5 -1 35 40 45 50 55 60 65 70 year 75 80 85 90 95 00 05 90 95 00 05 95 00 05 moving average 0 20 40 60 80 100 Temp Thess: moving average with order 12 18 Temperature Thessaloniki, period 1/1930-12/2000 17.5 30 20 17 Temp removal of seasonality 25 Temp 0 -2 0.5 r() 1931 7,9 8,8 10 12,7 19,7 24,9 27,4 26,9 21,5 16 10,8 1932 5 2,9 7,4 14,1 19,4 24,1 27,2 26,1 24,1 21,5 11,7 1933 5,2 7,6 9,1 13,5 17,6 22,8 25,5 25,3 20,5 17,7 14,2 1934 5,3 5,7 12,6 15,9 20,5 23,9 26,9 26,3 23,1 17,6 13,4 1935 4,5 6,8 8,2 14,7 18,7 24,9 26,1 26 22,8 19,8 12,1 1936 10,5 8,3 12,2 16 18,4 23 27,1 25,9 21,6 16,3 12,3 1937 4,8 8,2 12,9 14,5 19,7 24,4 26,4 26 23,6 16,9 12,8 1938 4,8 6,6 11 12,9 18,4 24,1 27,5 27,1 22,2 17,9 12,8 1939 7,9 7,9 8,5 15,5 19,8 23,1 27,2 26,6 21,9 18,5 12,3 1940 3,1 6,8 8,1 13,9 17,5 22,9 26,6 24,3 21,4 18,5 13 1941 6,9 10,2 11 15,6 19,2 23,7 26 26 18,9 15,7 10,1 1942 0,9 5,6 8,6 13,9 20,7 24,6 25,4 26,1 23,8 17,5 10,7 16.5 16 15 15.5 10 15 5 0 30 14.5 30 35 40 45 50 55 60 65 70 year 75 80 85 90 95 00 35 40 45 05 50 55 60 65 70 year 75 80 85 12-order difference Temperature Thessaloniki, period 1/1930-12/2000 - month Temp Thess: 12-differencing 30 10 25 Temp 15 10 5 Temp same model ARMA(P,Q)s for each month 20 0 -5 ? 5 0 1930 1940 1950 1960 1970 year 1980 1990 2000 -10 30 35 40 45 50 55 60 65 70 year 75 80 85 90 model ARMA(P,Q)s Temp Thess: ARMA(1,1)12 fit 30 Temp Thess: AIC of ARMA 12 models 5 25 4.5 20 AIC(p,q) 4 3.5 3 x(t) q=0 q=1 q=2 q=3 q=4 q=5 15 10 5 2.5 2 -1 0 30 0 1 2 3 4 5 35 40 45 50 55 60 65 70 time t 75 80 6 p Temp TempThess: Thess: ARMA(1,1) ARMA(1,1)12 fitfit 12 30 30 x 15.928 xt 0.0075 0.9995xt 12 zt 0.5242 zt 12 25 25 sz2 3.733 sz 1.932 standard deviation of residual time series x(t) x(t) Fit of ARMA(1,1)12 20 20 15 15 10 10 sz 1.603 55 00 60 60 time timett Fit with the estimation of the periodic component sz 1.427 85 90 95 00 05 Model of time series with trend and seasonality (SARIMA) Given that the time series y1 , y2 , , yn has trend and removal of trend y1 , y2 , , yn dependence between successive observations (time step 1) xt 2 , xt 1 , xt , xt 1 , xt 2 seasonality s removal of seasonality x1 , x2 , , xn dependence between seasonal components of the same seasonal order (time step s) xt 2 s , xt s , xt , xt s , xt 2s ARIMA(P,1,Q)s ARIMA(p,1,q) ( B)(1 B)d Yt ( B)Zt ( B s )(1 B s ) D Yt ( B s )Zt ( B)( B s )(1 B)d (1 B s ) D Yt ( B)( B s )Zt SARIMA(p,d,q)×(P,D,Q)s Seasonal multiplicative model d 0 and D0 SARMA(p,q)×(P, Q)s most often d 1 D0 Monthly index of global temperature (anomaly of surface temperature of the north hemisphere at grid 5ο x 5ο), time period 1850-2011 Example y1 , y2 , , yn real observations Source: http://www.cru.uea.ac.uk/cru/data/temperature monthly global temperature: autocorrelation 1.2 2 1 0.8 1 0.6 0 r() global temperature land air temperature anomalies, period 1/1850-12/2011 3 0.4 -1 0.2 -2 0 -3 50 60 70 80 90 00 10 20 30 40 50 60 year 70 80 90 00 10 20 -0.2 0 20 40 60 80 100 land air temperature anomalies, period 1/1850-12/2011 3 removal of trend ? global temperature 2 removal of seasonality / periodicity 1 0 dependences between successive ? observations (time step 1) -1 Jan May Sep -2 -3 1840 1860 1880 1900 1920 1940 year 1960 1980 2000 ? 2020 dependence between seasonal components of the same seasonal order (time step s) ? first differences differences of lag 12 first differences of month global temperature first differences of monthly global temperature 1.5 1 1 0.5 d(temp) d(temp) 0.5 0 -0.5 0 -0.5 -1 -1 -1.5 Jan50 Jan52 Jan54 Jan56 year Jan58 Jan60 Jan62 -1.5 Jan50 Jan52 Jan54 first difference of monthly global temperature: autocorrelation 0.3 Jan58 Jan60 Jan62 12-difference of monthly global temperature: autocorrelation 0.4 significant autocorrelations 0.2 Jan56 year 0.2 0.1 r() -0.1 for τ=12,24,… -0.2 0 r() for τ=1,2,… 0 -0.2 -0.4 -0.3 -0.4 0 20 40 60 80 100 -0.6 0 20 40 60 80 100 diff of temp: partial autocorrelation diff of monthly temp: AIC of SARMA(p,q)x(1,0) diff of monthly temp: AIC of SARMA(p,q)x(0,0) 0.1 -1.25 -1.35 -0.2 -0.3 -0.4 0 20 40 60 80 100 -1.4 -1.45 -1.55 -1.55 -1.6 -1.6 -1 -1.4 2 3 4 5 p diff of monthly temp: AIC of SARMA(p,q)x(1,1) -1.4 -1 6 -1.45 -1.55 -1.55 -1.6 -1.6 -1.6 -1.35 -1.4 0 1 2 3 4 5 p diff of monthly temp: AIC of SARMA(p,q)x(1,2) -1.25 q=0 q=1 q=2 q=3 q=4 -1.3 -1 -1.35 -1.45 -1.4 -1.45 -1.55 -1.55 -1.6 -1.6 -1.6 3 4 5 6 -1 0 1 2 p min(AIC)=-1.622 for SARMA(3,3) (1,2)12 3 p 4 5 3 4 5 p diff of monthly temp: AIC of SARMA(p,q)x(2,3) 6 6 q=0 q=1 q=2 q=3 q=4 -1.45 -1.55 2 2 -1.4 -1.5 1 1 -1.35 -1.5 0 0 -1.3 -1.5 -1 q=0 q=1 q=2 q=3 q=4 -1.25 q=0 q=1 q=2 q=3 q=4 -1.3 -1 6 AIC(p,q) -1.25 6 6 -1.45 -1.55 3 4 5 p diff of monthly temp: AIC of SARMA(p,q)x(0,2) 3 4 5 p diff of monthly temp: AIC of SARMA(p,q)x(2,0) -1.4 -1.5 2 2 -1.35 -1.5 1 1 -1.3 -1.5 0 0 -1.25 q=0 q=1 q=2 q=3 q=4 -1.35 -1.45 -1 AIC(p,q) 1 -1.3 AIC(p,q) AIC(p,q) -1.35 0 -1.25 AIC(p,q) -1.3 -1.45 -1.5 diff of monthly temp: AIC of SARMA(p,q)x(0,1) q=0 q=1 q=2 q=3 q=4 -1.4 -1.5 -1.25 -1.35 AIC(p,q) () AIC(p,q) -0.1 q=0 q=1 q=2 q=3 q=4 -1.3 AIC(p,q) -1.3 0 -1.25 q=0 q=1 q=2 q=3 q=4 -1 0 1 2 3 4 p SARMA(1,2) (1,1)12 AIC=-1.618 5 6 SARMA(3,3) (1,2)12 xt 1.12 xt 1 0.70 xt 2 0.22 xt 3 0.95 xt 12 1.11xt 13 0.70 xt 14 0.18 xt 15 zt 0.42 zt 1 0.22 zt 2 0.95 zt 3 1.01zt 12 0.48 zt 13 0.23zt 14 0.93zt 15 x 0.0013 sz 0.445 0.13zt 24 0.08 zt 25 0.05 zt 26 0.08 zt 27 diff global temp: ARMA(3,3)x(1,2)12 fit 4 2 2 x(t) x(t) diff global temp: ARMA(3,3)x(1,2)12 fit 4 0 -2 -4 50 60 70 0 -2 80 90 00 10 20 30 40 50 60 time t 70 80 90 00 10 -4 60 20 time t SARMA(1,2) (1,1)12 xt 0.35 xt 1 0.98 xt 12 0.34 xt 13 sz 0.446 zt 1.04 zt 1 0.1zt 2 0.93zt 12 0.99 zt 13 0.12 zt 14 diff global temp: ARMA(1,2)x(1,1)12 fit 4 4 2 2 x(t) x(t) diff global temp: ARMA(1,2)x(1,1)12 fit 0 -2 -2 -4 50 60 70 0 80 90 00 10 20 30 40 50 60 time t 70 80 90 00 10 20 -4 60 time t Prediction of time series Models for time series (AR, MA, ARMA, ARIMA, SARIMA) prediction Many applications Index and volume of the Athens Stock Exchange (ASE) Can we predict the index or volume the first day(s) of May 2002 given the observations until the end of April 2002? General index of consumer prices (GICP) General Index of Comsumer Prices General Index of Comsumer Prices, period Jan 2001 - Aug 2005 125 120 115 110 105 100 01 02 03 04 05 06 years At what level GICP is to be moved in the next months? Sunspots Annual sunspots, period 1960 - 2001 200 Annual sunspots, period 1700 - 2001 200 150 number of sunspots number of sunspots 150 100 50 0 1700 1750 1800 1850 1900 1950 years Annual sunspots, period 1900 - 2001 100 50 2000 0 1960 200 1970 1980 years 1990 2000 180 number of sunspots 160 Given the number of sunspots up to the current date, how many sunspots will be next year(s)? 140 120 100 80 60 40 20 1900 1920 1940 1960 years 1980 2000 Heart rate What is the next heart rate(s) ? The problem of time series prediction • We are given the time series up to time n • We want to estimate xn+k Prediction xn(k) Prediction error: en (k ) xnk xn (k ) x1 , x2 ,, xn Stochastic process { X n } prediction Xn(k) is the estimation of the observation Xn+k of { X n } Best prediction : X n (k ) E[ X nk | X n , X n1 , ] Properties of a good prediction: • unbiasedness : E[ X n (k )] X nk • efficiency, meaning small prediction error Var[ n (k )] Var[ X nk X n (k )] Optimizing both unbiasedness and efficiency minimization of the mean square prediction error 2 E X n k X n ( k ) Evaluation of a prediction model : Given x1 , x2 ,, xn learning / given also training set {xn1 , xn 2 , test / validation set , xnl } prediction model prediction errors k time step ahead xn (k ), xn1 (k ), , xnl k (k ) en (k ), en1 (k ), , enl k (k ) e j (k ) x j k x j (k ) j n, n 1, ,n l k Statistical measures of error Estimation of mean square error (mse) 1 n l k 1 n l k 2 mse(k ) x j k x j (k ) e j (k ) l k 1 l k 1 j n j n 2 root mean square error (rmse) 2 1 n l k 1 n l k 2 rmse(k ) e ( k ) x x ( k ) j j k j l k 1 j n l k 1 j n normalized root mean square error (nrmse) nrmse(k ) 1 l k 1 n l k x j n jk x j (k ) 2 1 n l k x jk x l k 1 j n 2 nrmse 0 very good prediction nrmse ≈ 1 prediction at the level of mean value prediction Predictions: Prediction many steps ahead for a given current time Given x1 , x2 ,, xn , we predict xn (1), xn (2),, xn (k ) Prediction at a given time step ahead for different current times Given x1 , x2 ,, xn , xn1 ,, xnl , we want to evaluate the predictability of a prediction model 1. We estimate the model parameters based on the time series x1 , x2 ,, xn 2. We pursue predictions for some time step ahead k xn (k ), xn1 (k ),, xnl k (k ) 3. We compute a statistic of prediction errors 2 1 n l k 1 n l k 2 rmse(k ) x j k x j (k ) e j (k ) l k 1 l k 1 j n j n prediction limits xn (k ) c1 / 2 Var en (k ) zt ~ N(0, z2 ) c1 / 2 z1 / 2 Simple prediction techniques Deterministic trend (revisited) xt t zt white noise xn k n k znk Prediction: zt ~ WN(0, z2 ) trend, a small varying function of time xn (k ) E nk znk | xn , xn1, , x1 n k Solution: Extrapolation of function μt for times > n Prediction error: μt = ? en (k ) z nk known simple substitution global (fit to x1 , x2 ,, xn ) unknown estimation e.g. polynomial local (fit only to the m last observations) pm (t ) c0 c1t cmt m xnm1, xnm2 ,, xn Index and volume of ASE, prediction with trend extrapolation (polynomial fit of trend) Deterministic seasonal term xt st zt deterministic seasonal term and deterministic trend xt t st zt Same approach: estimation of the deterministic term {xt }56 t 1 xt xt t GICP, January 2001 – August 2005 General Index of Comsumer Prices, linear trend is subtracted 3 t 103.9 + 0.31t 2 120 xt t st zt 115 110 1 detrended GICP General Index of Comsumer Prices General Index of Comsumer Prices, period Jan 2001 - Aug 2005 125 0 -1 -2 105 100 01 -3 02 03 04 05 -4 01 06 02 03 st t 1 xn (k ) n k snk 3 year cycle of GICP 2 Prediction of Sept 2005 57 103.9 + 0.31*57 121.70 s9 0.16 1 0 -1 -2 x56 (1) 121.86 -3 03 04 years 06 General Index of Comsumer Prices, trend and period comp. subtracted 4 detrended and deseasoned GICP General Index of Comsumer Prices, year cycle 4 02 05 zt xt st xt t st n -4 01 04 years years 05 06 3 2 1 0 -1 -2 -3 -4 01 02 03 04 years 05 06 Exponential smoothing Estimation of xn+k as a weighted sum of former observations n 1 xn (k ) c0 xn c1 xn1 cn1 x1 c j xn j j 0 Desired condition on the weights: c0 c1 cn1 Determination of the weights with a single parameter : c j (1 ) j , j 0,1,, n 1, 0 1 recursive relation : xn (k ) xn (1 ) xn1 (k ) n 1 j 0 cj 1 Prediction with exponential smoothing: Examples Index and volume of ASE Prediction at one time step ahead for all days in May 2002 Comparison of the prediction performance of exponential smoothing for different Large (weighting most the most recent observations) gives the best prediction Sunspots Heart rate Prediction of stationary time series with linear models Predictions with AR, MA and ARMA Prediction with autoregressive models (AR) Given the time series AR(1) model t n 1 x1 , x2 ,, xn xt xt 1 zt white noise xn1 xn z n1 Prediction error: en (1) z n1 Optimal prediction at time step 1: xn (1) xn t n2 Var en (1) z2 Prediction error : xn2 xn1 z n2 Optimal prediction at time step 2: xn (2) xn (1) xn 2 t nk k Optimal prediction at time step k: xn (k ) xn zt ~ WN(0, z2 ) en (2) zn1 zn 2 Var en (2) ( 2 1) z2 Prediction error : en (k ) k 1 zn1 znk 1 zn k 1 2k Var en (k ) 1 2 2 z Given the time series AR(p) model t n 1 x1 , x2 ,, xn xt 1 xt 1 p xt p zt xn1 1 xn p xn p1 z n1 Optimal prediction at time step 1: Var en (1) z2 xn (1) 1 xn p xn p1 t n2 xn2 1 xn1 p xn p 2 zn 2 Optimal prediction at time step 2: xn (2) 1 xn (1) 2 xn1 p xn p 2 t nk xnk 1 xn1 k p xn p k znk Optimal prediction at time step k: xn (k ) 1 xn (k 1) p xn (k p) prediction xn ( j ) where xn ( j ) observation xn j j 0 j0 Prediction error : en (1) z n1 Prediction error : en (2) 1en (1) zn2 1 zn1 zn2 Var en (2) ( 2 1) z2 Prediction error : en (k ) 1en (k 1) k en (1) znk k 1 en ( k ) b j z n k j j 0 Var en (k ) k 1 2 z b j 0 2 j Index ASE, multi-step prediction for May 2002 AR(1) AR(6) AR(11) 0.9995 1.1535 1.1523 -0.2126 -0.2131 0.0944 0.0961 -0.0655 -0.0663 0.0103 0.0135 0.0194 -0.0031 -0.0058 0.0190 -0.0175 0.0458 -0.0213 xn (k ), n 30.04.2002, k 1,,20 Index ASE, one step ahead prediction in May 2002 xn (1), n 2.05.2002 31.05.2002 Volume ASE, multi-step prediction for May 2002 AR(1) 0.9097 AR(6) 0.3412 0.2092 0.1557 0.1369 0.0773 0.0528 AR(11) 0.3251 0.1955 0.1380 0.1138 0.0455 0.0009 0.0350 0.0068 0.0249 0.0420 0.0527 xn (k ), n 30.04.2002, k 1,,20 Volume ASE, one step ahead prediction in May 2002 xn (1), n 2.05.2002 31.05.2002 Sunspots, multi step prediction from 1991 to 2001 AR(1) 0.8205 AR(6) 1.3231 -0.5297 -0.1655 0.1895 -0.2576 0.1702 AR(11) 1.1848 -0.4385 -0.1718 0.1933 -0.1324 0.0311 0.0157 -0.0203 0.1993 -0.0186 0.0352 xn (k ), n 1990, k 1,,21 Sunspots, prediction one year ahead in period 1991-2001 xn (1), n 1991 2001 Heart rate, prediction of the next 21 heart rates AR(1) 0.8065 AR(6) 0.7850 -0.1205 0.1983 0.1438 -0.1407 -0.0465 AR(11) 0.7803 -0.0736 0.1759 0.0858 -0.1239 -0.1899 0.1413 0.0761 0.0073 -0.0463 0.0347 xn (k ), n 1060, k 1,,21 Heart rate, prediction of the next heart rates xn (1), n 1061 1081 Growth rate of GNP of USA The observations are at annual-quarters, from the second quarter of 1947 till the first quarter of 1991 (n=176) Rate growth of GNP of USA Autocorrelation of rate growth 0.04 0.5 0.03 0.4 0.3 0.02 0.2 xt r() 0.01 0.1 0 0 -0.01 -0.1 -0.02 -0.2 -0.03 0 50 100 0 150 5 10 t 0.4 -9.16 0.3 -9.17 0.2 -9.18 p,p AIC(p) -9.15 0.1 -9.19 0 -9.2 -0.1 -9.21 -0.2 -9.22 2 4 6 p 8 20 AIC for Rate Growth Partial Autocorrelation for Rate Growth 0.5 0 15 10 -9.23 0 2 4 6 p 8 10 Growth rate of GNP of USA xn (k ), n 170, k 1, xt 0 1 xt 1 2 xt 2 3 xt 3 zt AR(3) ˆ 0.0077 ˆ1 0.35 ˆ2 0.18 ˆ3 0.14 ˆ0 ˆ 1 ˆ1 ˆ2 ˆ3 0.0047 ,6 Prediction of rate growth with AR(1) 0.04 0.03 0.02 0.01 xt 0.0047 0.35xt 1 0.18xt 2 0.14 xt 3 zt sz ˆ z 0.0098 0 -0.01 -0.02 -0.03 164 166 168 170 172 174 176 Prediction of rate growth with AR(3) 0.04 AR(1) 0.03 xt 0.0047 0.38xt 1 zt 0.02 0.01 sz ˆ z 0.0099 0 -0.01 -0.02 -0.03 164 166 168 170 172 174 176 Growth rate of GNP of USA predictability for k step ahead AR(p), p=1,…,10 xn (1), n 126 176 xn (1), n 146 176 nrmse(k) on the last 50 data nrmse(k) on the last 30 data 1.1 1.1 1 1 nrmse(p) nrmse(p) k=1 k=2 0.9 0.8 0.7 0 2 4 6 p 8 0.9 k=1 k=2 0.8 10 0.7 0 2 4 6 p 8 10 zt ~ WN(0, ) 2 z MA(1) model t n 1 Ez n j xt zt zt 1 xn1 z n1 z n Optimal prediction at time step 1: xn (1) zn t n2 αν j 0 αν j 0 0 | xn , xn1 , z n j xn 2 zn 2 zn1 Optimal prediction at time step 2: xn (2) 0 Prediction error : en (1) z n1 Var en (1) z2 Prediction error : en (2) xn 2 Var en (2) Var xn 2 x2 For time step k: z n xn ( k ) 0 για k 1 για k 1 z en (k ) n1 xn k για k 1 για k 1 MA(q) model t n 1 xt zt 1 zt 1 q zt q xn1 z n1 1 z n q z nq1 Prediction error : en (1) z n1 Optimal prediction at time step 1: xn (1) 1 z n q z nq1 Var en (1) z2 t n 2 xn 2 zn 2 1 zn1 2 zn q znq 2 Prediction error : Optimal prediction at time step 2: en (2) zn 2 1 zn1 Var en (2) ( 2 1) z2 xn (2) 2 zn q znq 2 t nk xnk zn k 1 zn k 1 q zn k q Optimal prediction at time step k: z z xn (k ) k n k 1 n 1 0 q zn q k Prediction error : en (k ) znk 1 znk 1 k 1 zn1 if k q if k q k 1 en (k ) j zn k j j 0 Var en (k ) k 1 2 z j 0 2 j Growth rate of GNP of USA The observations are at annual-quarters, from the second quarter of 1947 till the first quarter of 1991 (n=176) ΜΑ(2) xt 0.0077 zt 0.41zt 1 0.40zt 2 sz ˆ z 0.0109 xn (k ), n 170, k 1, xn (1), n 146 176 ,6 Prediction of rate growth with MA(2) nrmse(k) with MA(q) on the last 30 data 0.04 k=1 k=2 0.03 1.1 0.02 nrmse(q) 0.01 0 1 0.9 -0.01 0.8 -0.02 -0.03 164 166 168 170 172 174 176 0.7 0 2 4 6 q 8 10 ARMA(p,q) model xt 1 xt 1 p xt p zt 1 zt 1 q zt q t n 1 xn1 1 xn p xn p1 zn1 1 zn q znq1 Prediction error : en (1) z n1 Optimal prediction at time step 1: xn (1) 1 xn p xn p1 1 zn q znq1 Var en (1) z2 Optimal prediction at time step k: 1 xn (k 1) xn (k ) 1 xn (k 1) p xn (k p) k zn p xn (k p) q zn q k if k q if k q Prediction with ARMA: merging of the prediction with AR and MA Growth rate of GNP of USA ARMA(3,2) xt 0.0034 0.15xt 1 0.29 xt 2 0.12 xt 3 zt 0.33zt 1 0.13zt 2 sz ˆ z 0.0105 xn (k ), n 170, k 1, xn (1), n 146 176 ,6 Prediction of rate growth with ARMA(3,2) nrmse(k) with ARMA(p,1) on the last 30 data 0.04 k=1 k=2 0.03 1.1 nrmse(p) 0.02 0.01 0 1 0.9 -0.01 0.8 -0.02 -0.03 164 166 168 170 172 174 176 0.7 0 2 4 6 p 8 10 Prediction of non-stationary time series Given a non-stationary time series y1 , y2 , Stages of prediction: y1 , y2 , 1. transformation to stationary 2. prediction of xn+k with some model 3. inverse transform on the prediction standard decomposition model for yt : t n k prediction of yn+k : , yn yn (k ) 1 x1 , x2 , 3 , xn 2 xn (k ) yt t st xt ynk nk snk xnk Estimation of μt and st as functions of time t 1. xt yt t st 2. xn (k ) : prediction (of type ARMA) of xn+k 3. yn (k ) nk snk xn (k ) , yn Removal of μt and st (using differences) prediction with models ARIMA or SARIMA ARIMA(p,1,q) Stages of prediction of yn (1) : 1. transformation : y1 , y2 , , yn x2 , x3 , xn stationary xt yt yt 1 2. prediction of xn+1 with ARMA(p,q) xn (1) 3. inverse transform : yn (1) xn (1) prediction error : e (1) e~ (1) n yn (1) yn xn (1) n prediction error of xn (1) For prediction at k steps ahead : yn (k ) yn (k 1) xn (k ) known from the prediction of yn+k-1 ARMA(p,q) prediction of xn+k Similar procedure for the prediction with models ARIMA(p,d,q) or SARIMA(p,d,q)(P,D,Q)s Period from January 2002 to September 2005 ASE index Autocorrelation of ASE General Index 1 3000 0.8 2500 0.6 r() close index ASE General Index, Jan 2002 - Sep 2005 3500 2000 0.4 1500 0.2 1000 02 Returns 03 04 years 05 0 0 06 yt yt 1 xt yt 1 10 20 30 40 50 Autocorrelation of returns of ASE General Index 0.2 Returns of ASE General Index 0.05 0.15 0.04 0.1 0.05 0.02 r() close index returns 0.03 0.01 0 0 -0.05 -0.01 -0.1 -0.02 -0.15 -0.03 -0.04 02 -0.2 0 03 04 years 05 06 5 10 15 20 Period from January 2002 to September 2005 ASE index Order of AR model Partial autocorrelation of returns of general index 0.2 xt yt yt 1 yt 1 AIC of returns of general index -9.105 0.15 -9.11 0.1 -9.115 AIC(p) p,p 0.05 0 -0.05 -9.12 -9.125 -0.1 -9.13 -0.15 -0.2 0 5 10 p 15 -9.135 0 20 5 10 p 15 20 Prediction of many steps, all for current time on 20/9/2005 ASE index returns of ASE xn(k) of general index, n=20.9.2005 yn(k) of index return, n=20.9.2005 yt yt 1 (1 xt ) 0.015 return of index y (T), AR(7) n 0.01 3400 yn1 yn (1 xn1 ) 0.005 Prediction 0 yn (1) yn (1 xn (1)) -0.005 close index returns of index 3450 3350 3300 3250 yn (k ) yn (k 1)(1 xn (k )) -0.01 -0.015 18 25 02 days 09 16 general index xn(T), AR(7) 3200 18 25 02 days 09 16 ASE index Period from January 2002 to September 2005 One step ahead prediction for period 20/9/2005 – 12/10/2005 Estimation of prediction error with ΑR(p) models for the period 20/9/2005 – 12/10/2005 xn(1) of general index n=20.9.2005 to 12.10.2005 nrmse of AR for general index, 20.9.2005-12.10.2005 3450 1.5 k=1 k=2 k=5 3350 nrmse(p) close index 3400 general index AR(1) AR(7) 3300 1 3250 3200 18 25 02 days 09 16 0.5 0 5 10 p 15 20