Stat 882 (Winter 2012) – Peter F. Craigmile Stationary autoregressive moving average processes Part 1 Reading: Brockwell and Davis [1991, Chapter 3] • LTI filtering of random variables • LTI filtering of stationary processes • Defining the autoregressive moving average (ARMA) process – The moving average (MA) and autoregressive (AR) processes. • Simulating ARMA Processes in R • Example: The AR(1) process • An aside: a quick review of complex numbers • ARMA processes and the roots of their polynomials • Checking stationarity • Causal ARMA processes • Converting between ARMA and MA representations • Invertible ARMA processes 1 Motivation • The class of autoregressive moving average processes (ARMA) are the most famous class of time series models. • Popularly used since the 1960s. • This very general class of processes includes the moving average (MA) and autoregressive (AR) processes. • Why do we use these models? – Much is known about the statistical properties of these models. – Under suitable assumptions, we can exactly represent the ACVF at a finite collection of lags using an ARMA model of a large enough model order. • Before defining ARMA models, we discuss what it means to LTI filter a collection of random variables. 2 LTI filtering of random variables Brockwell and Davis [1991, Prop 3.1.1] • For LTI filter coefficients {ψj } that are absolutely summable; P i.e., j∈Z |ψj | < ∞, let ψ(B) = X ψj B j j∈Z (where B is the usual backshift operator). • Now let {Xt} be any collection of random variables such that supt E|Xt| < ∞. Then the filtered sequence Yt = ψ(B)Xt = X ψj Xt−j j∈Z converges absolutely with probability one. (Proof via the Monotone Convergence Theorem). • Now let {Xt} be any collection of random variables such that supt E|Xt| < ∞ and supt E|Xt|2 < ∞. Then Yt = ψ(B)Xt converges in mean square to the same limit. 3 LTI filtering of stationary processes (filtering preserves stationarity) Brockwell and Davis [1991, Prop 3.1.2] • Suppose {Xt} is a stationary process with mean µX and ACVF γX (·). • Let {ψj } be a set of absolutely summable coefficients. • Then Yt = ψ(B)Xt convergences absolutely with probability one and in mean square to the same limit (follows from the previous page). • Show that {Yt} is a stationary process: 4 LTI filtering of stationary processes, continued 5 Remarks • Commonly the process that we are filtering is a white noise or IID process. – Such a filtering defines what is called a linear process. • Thinking of ψ(·) as a polynomial in the backshift operator it helps to think about power series when filtering {Xt} using ψ(·) to yield {Yt}. (We will demonstrate with examples in these notes.) • Filter cascades will also be important. 6 Defining the autoregressive moving average (ARMA) process • The process {Xt} is an ARMA(p, q) process if 1. {Xt} is stationary, and 2. for every t we can write φ(B)Xt = θ(B)Zt, for some WN(0, σ 2) process {Zt}, where φ(B) = 1 − p X φj B j j=1 is the autoregressive (AR) polynomial of order p, and θ(B) = 1 + q X θj B j j=1 is the moving average (MA) polynomial of order q. 7 ARMA processes: remarks and subclasses • We can show later that {Xt} is a mean zero process. We obtain a mean µ, ARMA(p, q) process {Yt} by taking an ARMA process {Xt} and letting Yt = µ + Xt, for each t. • With φ(B) = θ(B) = 1, Xt = Zt for all t. Thus an ARMA(0, 0) process is a white noise process. 8 Subclasses of ARMA model • When φ(B) = 1, we obtain the moving average, MA(q), process of order q: X t = Zt + q X θj Zt−j . j=1 • When θ(B) = 1, we obtain the autoregressive, AR(p), process of order p. Xt − p X φj Xt−j = Zt j=1 The AR process is attributed to George Udny Yule (1871– 1951) (See the biography by O’Connor and Robertson and look at Yule [1921] and Yule [1927]. The AR(1) process has also been called the Markov process). • If θ(B) contains absolutely summable coefficients {θj }, then it makes sense to define a MA(∞) process: ∞ X Xt = θj Zt−j . j=0 (Similarly we can also define an AR(∞) process). 9 Simulating ARMA Processes in R • To simulate a series, called x, of length n from a time series model we use: x <- arima.sim(n=, model) • Here, model is a list of two possible items: 1. ar: a vector of AR coeffcients {φj }. 2. ma: a vector of MA coeffcients {θj }. • Some examples: ## simulate an AR(1) process with phi_1=0.7 x <- arima.sim(n=100, model=list(ar=0.7)) ## simulate an ARMA(1,1) process with phi_1=0.4 and theta_1=0.5 y <- arima.sim(n=100, model=list(ar=0.4, ma=0.5)) • To change the stdev. of {Zt} use the sd argument: ## simulate an AR(1) process with phi_1=0.7 and sigma^2=2 x <- arima.sim(n=100, model=list(ar=0.7), sd=sqrt(2)) 10 Example: The AR(1) process • Consider the AR(1) process {Xt} defined by Xt = φXt−1 + Zt, where {Zt} is a WN(0, σ 2) process. • When φ = 1 or φ = −1 we obtain a form of random walk. This process is not stationary. • What about other values of φ? Is the process stationary in these other cases? • With φ(B) = (1 − φB) we have φ(B)Xt = Zt. 11 An aside: a quick review of complex numbers • Let i be the square root of negative one, √ i = −1. • Then any complex number ζ can be written as ζ = a + b i, for real numbers a and b. • The polar representation of ζ is ζ = |ζ|eiArg(ζ). • Here, the modulus of a complex number ζ, |ζ| is |ζ| = √ a2 + b2 , and the argument of ζ is Arg(ζ) = “ arctan(b/a)00. 12 When is the AR(1) process stationary? • The key idea is to see whether there exists a LTI filter ψ(B) such that ψ(B)φ(B)Xt = ψ(B)Zt, with ψ(B)φ(B) = 1. • Then we have written Xt = ψ(B)Zt, and by the LTI filtering preserves stationarity result, {Xt} is a mean zero stationary process. • Forget about the backward shift operator for the moment. • For the polynomial φ(z) = 1 − φz, with z ∈ C (a complexvalued number), it follows that when |φ| < 1 ∞ X 1 1 = = φj z j . φ(z) 1 − φz j=0 13 The |φ| < 1 case 14 The |φ| < 1 case, continued 15 The |φ| > 1 case • When |φ| > 1 we have Xt = φXt−1 + Zt. Dividing by φ we get φ−1Xt = Xt−1 + φ−1Zt, ı.e., Xt−1 = φ−1Xt − φ−1Zt. • Can write this as a linear combination of future Zt’s. As this is also a filtering of a stationary process we have a stationary solution. – BUT, Xt depends on future values of the {Zt} – not very practical! • If we assume that Xs and Zt are uncorrelated for each t > s, |φ| < 1 is the only stationary solution to the AR equation. 16 ARMA processes and the roots of their polynomials • In the next few slides we discuss a number of properties of ARMA processes: – Stationarity – Casuality – Invertibility Each property involve rewriting the ARMA equations in a different form, and can be tested by examining the roots of the AR or the MA polynomial. e.g., examining the values ζ ∈ C such that φ(ζ) = 0. • From now on we will assume that φ(·) and θ(·) do not have common roots. – If there are common roots, either the ARMA process can be simplified, or there may be more than one solution solution to the ARMA equations [Brockwell and Davis, 1991, p.87, Remark 1]. 17 Decomposing roots • The Fundamental Theorem of Algebra tells us that any polynomial of degree p, f (·) can always be rewritten as f (z) = K n Y (z − ζj ), j=1 where ζ1, . . . , ζn are the roots of the polynomial f (·). (The roots are real- or complex-valued). • When we think about f (·) being, for example, the AR polynomial φ(·), it is more convenient to decompose the polynomial as n Y φ(z) = K1 (1 − ζj z). j=1 18 Checking stationarity in general • For any ARMA(p,q) process, a stationary and unique solution exists if and only if φ(z) = 1 − φ1z − . . . − φpz p 6= 0, for all |z| = 1. • General strategy: 1. Find the roots of φ(·), i.e., the values of ζ such that φ(ζ) = 0. 2. Show that the roots do not have modulus 1; i.e., show |ζ| = 6 1 for each root ζ. • We can use the polyroot function in R to find the roots of a polynomial, and Mod to calculate the modulus. 19 Causal ARMA processes • An ARMA process is causal if there exists constants {ψj } P with ∞ j=0 |ψj | < ∞ and Xt = ∞ X ψj Zt−j ; j=0 that is, we can write {Xt} as an MA(∞) process depending only on the current and past values of {Zt}. • Equivalently, an ARMA process is causal if and only if φ(z) = 1 − φ1z − . . . − φpz p 6= 0, for all |z| ≤ 1. 20 Converting between ARMA and MA representations • For an causal AR polynomial φ(·) and an MA polynomial θ(·) suppose we observe the ARMA process {Xt} φ(B)Xt = θ(B)Zt. • The MA(∞) representation of this process is written as Xt = ψ(B)Zt. • We want to identify the polynomial ψ(·), which satisfies φ(B)ψ(B) = θ(B). • Next we change the B to some z ∈ C: (1 − φ1z − . . . − φpz p)(ψ0 + ψ1z + ψ2z 2 . . .) = 1 + θ1 z + . . . + θq z q 21 Converting, continued • We match the constant terms, z terms, z 2 terms, etc. of both sides of the equation, and obtain 1 = ψ0 θ1 = ψ1 − ψ0φ1 θ2 = ψ2 − ψ1φ1 − ψ0φ2 ... • More generally for j = 1, 2, . . . min{j,p} θj = ψj − X φk ψj−k . k=1 • Now we solve the linear system of equations for {ψj }. • We can solve these equations recursively. • In R we can convert from an ARMA model to the MA representation using the function ARMAtoMA(ar, ma, lag.max). 22 Invertible ARMA processes • An ARMA process is invertible if there exists constants P {πj } with ∞ j=0 |πj | < ∞ and Zt = ∞ X πj Xt−j ; j=0 i.e., we can write {Zt} as an MA process depending only on the current and past values of {Xt}. • The process is invertible if and only if θ(z) = 1 + θ1z + . . . + θpz q 6= 0, for all |z| ≤ 1. 23 References P. J. Brockwell and R. A. Davis. Time Series. Theory and Methods (Second Edition). Springer-Verlag, New York, NY, 1991. J. J. O’Connor and E. F. Robertson. A biography of George Udny Yule. The MacTutor History of Mathematics archive. URL http://www-gap.dcs.st-and.ac.uk/ ~history/Biographies/Yule.html. G. U. Yule. On the time-correlation problem, with special reference to the variatedifference correlation method. Journal of Royal Statistical Society, 84:497–537, 1921. URL http://www.jstor.org/stable/2341101. G. U. Yule. On a method of investigating periodicities in disturbed series, with special reference to Wolfer’s sunspot numbers. Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character, 226: 267–298, 1927. URL http://www.jstor.org/stable/91170. 24