Optimal Filtering Formula#
It is convenient to have a formula for the projection of a random variable \(y_t\) against current and past values of a covariance stationary, indeterministic random process \(x_t\). We assume that \(y_t\) and \(x_t\) have means of zero and are jointly covariance stationary, indeterministic processes. That is, we seek the \(h_j\) that characterize the one-sided projection
where \(E x_{t-j} u_t = 0\) for all \(j \geq 0\). First, suppose that \(x_t\) has the moving average representation
where \(\{\epsilon_t\}\) is a serially uncorrelated process of innovations in \(x\), i.e., \(\epsilon_t\) is fundamental for \(x\). As an intermediate step,[1] think of projecting \(y_t\) on current and past \(\epsilon\)’s:
where \(E u_t \epsilon_{t-j} = 0\) for all \(j \geq 0\). We assume that \(x_t\) has both a moving average and an autoregressive representation, so that it is easy to see that \(\{ x_t, x_{t-1},\ldots\}\) and \(\{\epsilon_t, \epsilon_{t-1},\ldots\}\) span the same space. For this reason, \(u_t\) in (303) equals \(u_t\) in (304). Since the \(\epsilon\)’s form an orthogonal process, we have that the \(\phi_j\) are the simple least squares coefficients:
where \(\sigma^2 = E \epsilon_t^2\). Thus we can write
where \([\,]_{+}\) again means “ignore negative powers of \(L\)” and \(g_{y \epsilon}(L)\) is the cross-covariance generating function
We can relate \(g_{y \epsilon}(L)\) to the cross-covariance generating function \(g_{y x}(L)\) as follows:
Thus we have \(g_{y \epsilon}(z) = g_{y x}(z)/d(z^{-1})\). Substituting this into (305), we obtain
So we have
so that in (303) we have
A classic application of this formula is due to Muth (1960). Suppose that income evolves according to \(x_t = y_t + \epsilon_t\), where \(y_t = \rho y_{t-1} + u_t\), \(|\rho| < 1\), and where \(u_t\) and \(\epsilon_t\) are mutually orthogonal at all lags and serially uncorrelated. Here \(x_t\) is measured income, while \(y_t\) is “systematic” or permanent income. The consumer only “sees” \(x_t, x_{t-1}, \ldots\) and desires to estimate systematic income \(y_t\) by a linear function of \(x_t, x_{t-1}, \ldots\). The consumer is assumed to know all the relevant moments. This problem can be solved quickly using formula (307), and the reader is invited to do so.