Model specification (identification) We already know about the sample autocorrelation function (SAC): n rk Y Y Y Y Y t k 1 t k 2 t n t 1 Y t Properties: • Not unbiased (since a ratio between two random variables) • Bias decreases with n • Variance complicated, common to use general large-sample results Large-sample results (asymptotics): For large n the random vector H n r1 1 , n r2 2 ,, n rm m Note : n rj j rj j 1 n A kind of st udent izing ; has an approximate multivariate normal distribution with zero mean vector and covariance matrix ( cij ) where cij r 2 2 2 2 r i r j r i r j i r r j j r r i i j r This gives that c c Var (rk) 0 as n rk is asymptotic ally N k , kk (i.e. rk kk ) n n ckj Covrk , rj does not diminish as n ckk c jj Hence, the distribution of rk will depend on the correlation structure of Yt and accordingly on the model behind (i.e. if it is and AR(1), an ARMA(2,1) etc.) For an AR(1), i.e. Yt = Yt – 1 + et 1 1 2 1 2k Var rk ckk n 1 2 1 2 n 2 k 2k 2 1 1 n 1 2 k 0 k large i.e. not dependent on k for large lags For an MA(q) q 1 2 Varrk ckk 1 2 j for k q n j 1 i.e. not dependent on k after the qth lag For white noise 1 1 Varrk since j n 0 j 0 ckk j0 r Only some terms 0 2 1 n i.e. whitenoise - based 95 % bounds k 1 2 1 2 rj2 n j 1 i.e. bounds for rk based on an MAk - process with j estimated by rj Partial autocorrelation function kk k Corr Yt , Yt k Yt 1 , Yt 2 ,, Yt k 1 Describes the “specific” part of the correlation between Yt and Yt – k that is not due to successive serial correlations between the variables Yt – 1 , Yt – 2 , …, Yt – k . Partial correlations are used for other types of data as well (for instance in linear models of cross-sectional data. Patterns • For an AR(p)-process, k cuts off after lag p (i.e. the same type of behaviour like k has for an MA(q)-process • For an MA(q)-process k shows approximately the same pattern as does k for an AR(p)-process Estimation from data , Sample Partial Autocorrelation function (SPAC): No explicit formula, estimation has to be made recursively Properties of SPAC: More involved, but for an AR(p)-process SPAC-values for lags greater than p are approximately normally distributed with zero mean and variance 1/n Extended Autocorrelation function (EACF) One (of several) tool to improve the choice of orders of ARMA(p, q)-processes. Very clear as a theoretical function, but noisy when estimated on series not too long. AR, MA or ARMA? No pattern at all? EACF table for Y AR/MA 0 1 2 3 4 5 6 7 8 9 10 11 12 13 0ooooooooooo o o o 1ooooooooooo o o o 2xoooooooooo o o o 3ooxoooooooo o o o 4xoooooooooo o o o 5xxooooooooo o o o 6oxoooxooooo o o o 7oooooxooooo o o o True process: Yt = 1.3 + 0.2Yt – 1 + et – 0.1et – 1 ARMA(0,0) or ARMA(1,0)? Model selection from more analytical tools Dickey-Fuller Unit-Root test H0: The process Yt is difference non-stationary (Yt is stationary) Ha : The process Yt is stationary Augmented Dickey-Fuller test statistic (ADF): , X t an AR p - process Yt Yt 1 X t Yt Yt Yt 1 1Yt 1 X t 1Yt 1 1 X t 1 p X t p et If =1 (difference non-stationary) X t Yt Yt 1 Yt Yt 1 1 1 Yt 1 1 Yt 1 Yt 2 p Yt p Yt p 1 et 0 Fit the model Yt Yt 1 1 Yt 1 1 Yt 1 Yt 2 p Yt p Yt p 1 et a Yt 1 1 Yt 1 Yt 2 p Yt p Yt p 1 et and test H0: a = 0 (difference non-stationary) vs. Ha : a < 0 (stationary) using the test statistic ADF aˆ s.e.aˆ However, not t-distributed under H0. Another sampling distributions has been derived and tables (programmed in R) Akaike’s criteria For an ARMA(p,q)-process, let k = p + q + 1 and find the values of the parameters p, q, 1 , …, p , 1 , …, q of the model 1 B 1 p q B Y 1 B B et p t 1 q that minimizes • –2log{max L(p, q, 1 , …, p , 1 , …, q )} + 2k AIC [Akaike’s Information Criterion]. Works well when the true process have (at least one) infinite order • –2log{max L(p, q, 1 , …, p , 1 , …, q )} + klog(n) BIC [(Schwarz) Bayesian Information Criterion]. Works well when we “know” that the true process is a finite-order ARMA(p,q)