Signal Extraction Problems[1]#
This section extends to dynamic, serially correlated settings the static signal-extraction problem of Chapter X, where an agent estimates an unobserved variable from a noisy observation by linear least squares projection. Here the signal and the noise are themselves stochastic processes, and the projection runs over their entire histories.
Let \(y_t\) be a covariance stationary stochastic process with \(m\)th order moving average representation
where
where \(u_t\) is a white noise with variance \(\sigma_u^2\) that is not necessarily fundamental for \(y_t\). Suppose that \(x_t\) is the sum of \(y_t\) and an orthogonal serially uncorrelated white noise \(\eta_t\) with variance \(\sigma_\eta^2\), where
where
Suppose that an agent observes \(\{x_t, x_{t-1}, \ldots\}\) at \(t\), and wishes to construct linear least squares forecasts of \(x\)’s on the basis of this information set. To construct the linear least squares forecast for \(x_{t+k}\) given \(\{x_t, x_{t-1}, \ldots\}\), one uses the Wiener-Kolmogorov formula (248), which requires that a Wold moving average representation \(x_t = d(L)\epsilon_t\) be obtained for \(x_t\).
To obtain the Wold representation for \(x_t\), we simply use the method of Deriving the Moving Average Representation. In particular, the covariance generating function of \(x_t\) is
We find the zeros of \(g_x(z)\), which come in reciprocal pairs, and prepare the factorization
where the zeros of \(d(z) = (1 - \lambda_1 z)\cdots(1-\lambda_m z)\) do not lie inside the unit circle, where \(\sigma_\epsilon^2\) solves
The Wiener-Kolmogorov formula (248) can then be used to calculate \(P[x_{t+k}|x_t, x_{t-1},\ldots]\).
Moving into a richer class of examples, we now let \(y_t\) be a process with mixed moving average, autoregressive representation
where \(u_t\) is a white noise with variance \(\sigma_u^2\), and
where the \(\alpha_j\)’s can be either side of the unit circle. Suppose that \(x_t\) is the sum of \(y_t\) and a serially uncorrelated white noise \(\eta_t\) with variance \(\sigma_\eta^2\),
where \(E \eta_t u_{t-s} = 0\) for all \(s\). Again we desire to find \(P[x_{t+k}|x_t, x_{t-1}, \ldots]\), so we need to find a Wold representation for \(x_t\). We use the method of Finding a Wold Representation.
The covariance generating function of \(x\) is
Taking the right-hand side to a common denominator gives
The numerator polynomial is of order \(p = \max(n,m)\), and can be factored to be of the form
where
and where \(\sigma_\epsilon^2\) solves
The Wold moving average representation for \(x_t\) is then
The Wiener-Kolmogorov formula can be applied to (271).
A famous application of the preceding analysis is due to Muth (1960). Muth assumed that income \(x_t\) is the sum of a first order Markov process \([1/(1-\rho L)]u_t\), \(|\rho| < 1\), and an uncorrelated white noise \(\eta_t\). The agent’s problem was to predict his future income. Setting \(a(L)=1\), \(b(L) = (1 - \rho L)\), we find that equation (270) becomes
The expression on the left can be written
Applying the quadratic formula, and setting \(\lambda_1\) equal to the root that is smaller in absolute value, we have
The limiting value of \(\lambda_1\) as \(\rho\) approaches 1 from below is
which is the expression obtained by Muth (1960). Thus we have that \(x_t\) has the first-order moving average, first-order autoregressive representation
where \(\epsilon_t\) is a fundamental white noise for \(x_t\) with variance \(\sigma_\epsilon^2\) that solves
The result (249) of Some Examples now applies with \(\beta \equiv \rho\) and \(\lambda_1 \equiv -a\). Thus we have
so that projections of future \(x\)’s are a geometric average of past \(x\)’s.