I.FirstOrderAutoregressiveProcesses.docVIP

I. First Order Autoregressive Processes First order autoregressive processes, AR(1), have a structure similar to random walks: AR(t) = b*AR(t-1) + e(t) . If the coefficient, b, equals one we have a random walk. If b is greater than one the process is also evolutionary. If b is positive and less than one there is a positive dependence on the past but the process is stationary. Many economic series are characterized with a positive autoregressive structure. If b equals zero, we have a process of order zero, i.e. white noise. If b is between minus one and zero the process depends negatively on the past but is covariance stationary. If b is algebraically less than one, the process is evolutionary. The following plot shows three evolutionary autoregressive processes for values of b of 1.5, 1, and -1.5, respectively. Using the lag operator, an autoregressive process can be expressed as: [1-bZ] AR(t) = e(t), so that {1-bZ] is the filter that converts a first order autoregressive process to white noise. Multiplying both sides by the inverse of this filter: [1-bZ]-1[1-bZ] AR(t) = AR(t) = [1-bZ]-1e(t) , or: AR(t) = [1 + bZ + b2Z2 + b3Z3 + ...] e(t) = e(t) + bZe(t) + b2Z2e(t) ..., AR(t) = e(t) + be(t-1) + b2e(t-2) + b3e(t-3) ... . The mean function of this autoregressive process is: m(t) = E[AR(t)] = Ee(t) + bEe(t-1) + b2Ee(t-2) + b3Ee(t-3) ... = 0, and the autocovariance function at lag zero is: ?AR,AR???E{[e(t) + be(t-1) + b2e(t-2) ...][e(t) + be(t-1) + b2e(t-2)...]} = ??????b2????+ b4???+ ... = ???/(1-b2? If we lag the autoregressive process by one period, AR(t-1) = e(t-1) + be(t-2) + b2e(t-3) + b3e(t-4) ... , we see that it depends upon past values of the white noise series and if the parameter b is less than one in absolute value, the influence of the past will fade. The autocorrelation at lag one can be calculated from: ?AR,AR(1)???E[AR(t) AR(t-1)] =E{AR(t-1)][ b*AR(t-1) + e(t)]} = b*E[AR(t-1) AR(t-1)] = b*??AR.AR (0), so that the autocorrel