Macroeconomics I University of Tokyo
Lecture 1 Time Series LS, Chapter 2
Julen Esteban-Pretel National Graduate Institute for Policy Studies Outline of Lecture 1
§ Definitions. • Stochastic process. • Stationary processes. § Markov Chains. § First-order stochastic linear difference equations. • Examples: - AR(2). - ARMA(1,1). - VAR. • First and second moments. • Impulse response functions.
Core Macro I - Spring 2013 Lecture 1: Time Series 2 Stochastic Processes
§ We will only focus on discrete-time processes.
§ Let xt be a vector of random variables. § Def: A stochastic process is a time ordered sequence of random variables.
We write it as {xt} and the starting date is t = 0 (some authors start at t=1). § Def: A stochastic process is stationary if the joint distribution does not change over time.
Core Macro I - Spring 2013 Lecture 1: Time Series 3 Covariance-Stationary Processes
§ Def: A stochastic process {xt} is covariance (or weakly) stationary if:
i. E(xt) does not depend of time, and
ii. Cov(xt+j, x t) exists, is finite, and depends only on j but not on t, for all j, t.
§ Def: The j-th order autocovariance, Cx( j ), is
Cx( j ) ≡ Cov(xt+j, x t) for t = 0,1,2,... (1.1)
§ Cx( j ) does not depend on t because of covariance stationarity.
§ Also due to covariance stationarity, Cx( j ) satisfies
Cx( j ) = Cx( - j )’. (1.2) § The 0-th order autocovariance is the variance
Cx( 0 ) = Var(xt). (1.3)
Core Macro I - Spring 2013 Lecture 1: Time Series 4 White Noise, MDS and i.i.d.
§ Def: A process {wt} is white noise if it is covariance stationary with mean zero and with no serial correlation (i.e. Cw(j) = 0 for j ≠ 0).
§ Def: A process {wt} is a martingale difference sequence (mds) if E(wt+1 | wt, w t-1, ...)=0.
§ Def: A process {wt} is said to be i.i.d. if it is independent and identically distributed.
§ “{wt} is i.i.d. with mean zero and finite second moments” ⇒ “{wt} is a
stationary mds with finite second moments” ⇒ “{wt} is white noise”
Core Macro I - Spring 2013 Lecture 1: Time Series 5 Markov Processes
§ Def: A stochastic process {xt} is said to have the Markov property (or it is a Markov process) if for k ≥ 1 and all t,
Prob(xt+1| xt, x t-1,..., xt-k) = Prob(xt+1| xt) (1.1.4)
§ Def: A state space is the space in which the possible values of each xt lie.
Core Macro I - Spring 2013 Lecture 1: Time Series 6 Markov Chains
§ Def: A Markov chain is a Markov process whose state space is a discrete set. We will focus on Markov chains defined over finite sets. § Def: A time-invariant Markov chain is described by a triple of objects:
i. An n-dimensional state space, ei, i = 1, 2,..., n, where ei is an n x 1 unit vector whose i-th entry is 1 and all other entries are zero. ii. An n x n transition matrix, P, which records the probability of moving from one state to another in one period. The elements of P are
Pij = Prob(xt+1 = ej | xt = ei). (1.5)
iii. An n x 1 vector, π0, of initial conditions, which specifies the probability of being in each state at date 0:
π0i = Prob(x0 = ei). (1.6) § We will assume that, n (1.7) - For i = 1, 2,...,n, the matrix P satisfies: j=1 Pij = 1 A matrix P satisfying this property is called a stochastic matrix. n - The vector π0 satisfies: i=1 0i = 1 (1.8)
A vector π0 satisfying this property is called a probability vector.
Core Macro I - Spring 2013 Lecture 1: Time Series 7 Markov Chains (Cont.)
§ A stochastic matrix, P, defines the probability of moving from each value of the state to any other in one period. § The probability of moving from one value of the state to any other in k periods is given by Pk.
P(k) = Prob(x = e x = e ) (1.9) ij t+k j| t i (k) k where P i j is the i, j element of P .
§ The unconditional probability t, i = P r o b ( x t = e i ) is the i-th element of πt where t is given by (n 1) t (1.10) t⇥ = 0⇥ P . (1 n) (1 n)(n n) § To derive (1.10) we use the following formula from probability theory:
p(x) = p(x, y), p(x, y) = p(x y)p(y), (1.11) y | where p ( x ) P r o b ( X = x), p(x, y) Prob(X = x, Y = y). Core Macro I - Spring 2013 Lecture 1: Time Series 8 Stationary Distributions
§ According to (1.10) the unconditional distributions evolve as (1.12) t +1 = t P.
§ Def: An unconditional distribution is stationary or invariant if t+1 = t.
§ Hence, if t is a stationary distribution, it satisfies (n 1) (1.13) = P
or = P
or (In P⇥) = 0 . (1.14) (n 1) which implies that π is an eigenvector associated with a unit eigenvalue of P’.
§ If π0 is a stationary distribution, then the Markov chain is a stationary Markov process.
Core Macro I - Spring 2013 Lecture 1: Time Series 9 Asymptotic Stationarity
§ Def: For a given arbitrary π0, a Markov chain is asymptotically stationary if
lim t = t ⇥ ⇥ where satisfies (1.14). § Def: A Markov chain is asymptotically stationary with a unique stationary distribution if the limit is independent of the initial distribution π0. § We call a stationary distribution or an invariant distribution of P. § Theorem 1: Let P be a stochastic matrix with Pij > 0 ∀(i, j). Then P has a unique stationary distribution and the process is asymptotically stationary.
n § Theorem 2: Let P be a stochastic matrix for which P i j > 0 ( i, j) for some value of n ≥ 1. Then P has a unique stationary distribution, and the process is asymptotically stationary.
Core Macro I - Spring 2013 Lecture 1: Time Series 10 Stochastic Linear Difference Equations
§ A first-order stochastic linear difference equation is an example of a continuous-state Markov process.
xt+1 = Ao xt + C wt+1 , t = 0, 1, 2, . . . . (1.15) (n 1) (n n)(n 1) (n m)(m 1)
§ {wt} is either white noise, stationary mds, or i.i.d. with zero mean, with
E(wtwt ) =Im.
§ We can append an observation equation yt=Gxt and use the augmented system
xt+1 = Aoxt + Cwt+1 (1.16a)
yt = Gxt. (1.16b)
where yt is the vector of variables observed at time t. § The system (1.16) is called a linear state-space system.
Core Macro I - Spring 2013 Lecture 1: Time Series 11 Stochastic Linear Difference Equations (cont.)
§ Re-stating the first-order stochastic linear difference equation
xt+1 = Aoxt + Cwt+1 (1.15)
§ Iterating (1.15) forward from t = 0, we obtain
t t 1 j (1.17) xt = Aox0 + j 0 AoCwt j. =
§ Hence, if x0 is uncorrelated with wt (t = 1,2,...), then xt is uncorrelated with wt+j ( j = 1,2,...).
Core Macro I - Spring 2013 Lecture 1: Time Series 12 Example 1: AR(2)
xt+1 = Aoxt + Cwt+1 (1.16a)
yt = Gxt. (1.16b) Scalar second-order autoregression
§ Assume that zt and wt are scalar processes and that zt+1 = + ⇥1zt + ⇥2zt 1 + wt+1. (1.18) § We can represent this relationship as a first-order system
zt+1 ⇥1 ⇥2 zt 1 zt = 1 0 0 zt 1 + 0 wt+1, (1.19a) 1 ⇥ 0 0 1⇥ 1 ⇥ 0⇥
⇤ ⌅ ⇤ ⌅ ⇤zt ⌅ ⇤ ⌅ zt = [1 0 0 ] zt 1 . (1.19b) 1 ⇥ § Hence an AR(2) can be written as in ⇤first-order⌅ stoch. linear diff. eq. where
zt ⇥1 ⇥2 1 xt = zt 1 , Ao = 1 0 0 , C = 0 , yt = zt, G = [1 0 0 ]. 1 ⇥ 0 0 1⇥ 0⇥
Core Macro I - Spring⇤ 2013 ⌅ ⇤ ⌅ ⇤ ⌅ Lecture 1: Time Series 13 Example 2: ARMA (1,1)
xt+1 = Aoxt + Cwt+1 (1.16a)
yt = Gxt. (1.16b) First-order scalar mixed moving average and autoregression
§ Assume that zt and wt are scalar processes and that zt+1 = ⇥zt + wt+1 + wt. (1.20) § We can represent this relationship as a first-order system z ⇥ z 1 t+1 = t + w , w 0 0 w 1 t+1 (1.21a) t+1⇥ ⇥ t⇥ ⇥ z z = [1 0 ] t . t w (1.21b) t⇥ § Hence an ARMA(1,1) can be written as in first-order stoch. linear diff. eq.: z ⇥ 1 x = t , A = , C = , y = z , G = [1 0 ] w o 0 0 1 t t t⇥ ⇥ ⇥ Core Macro I - Spring 2013 Lecture 1: Time Series 14 Example 3: VAR
xt+1 = Aoxt + Cwt+1 (1.16a)
yt = Gxt. (1.16b) n-dimensional 4-th order vector autoregression
§ Let zt be an n x 1 vector of random variables, wt+1 a mds with x 0⇥ = [ z 0 z 1 z 2 z 3 ] , and Aj an n x n matrix for each j. 4 zt+1 = Aj zt+1 j + Cy wt+1 . (1.22) (n 1) j 1 (n n) (n 1) (n n)(n 1) ⇥ = ⇥ ⇥ ⇥ ⇥ § We can represent this relationship as a first-order system
zt+1 A1 A2 A3 A4 zt Cy zt In 0 0 0 zt 1 0 ⇥ = ⇥ ⇥ + ⇥ wt+1. (1.23) zt 1 0 In 0 0 zt 2 0 ⇧zt 2⌃ ⇧ 0 0 In 0 ⌃ ⇧zt 3⌃ ⇧ 0 ⌃ ⇧ ⌃ ⇧ ⌃ ⇧ ⌃ ⇧ ⌃ ⇤ ⌅ ⇤ ⌅ ⇤ ⌅ ⇤ ⌅
Core Macro I - Spring 2013 Lecture 1: Time Series 15 First and Second Order Moments
xt+1 = Ao xt + C wt+1 , t = 0, 1, 2, . . . . (1.15) (n 1) (n n)(n 1) (n m)(m 1)
§ Let µt E(xt), t var(xt). (n 1) (n n) § Taking the unconditional expectation on both sides of (1.15)
µt+1 = Aoµt. (1.24) § Taking the variance of both sides of (1.15) and noting, as was shown before, that Cov(xt, w t+1) = 0,