1
$\begingroup$

Consider a State Space Model, where the observed process is $Y_t$ $$ Y_t = B F_t + \epsilon_t \\ F_t = \Phi F_{t-1} + \nu_t $$ where the error terms are white noise. Later on, I want to compute the state vector $F_t$ with the Kalman Filter.

I often find the assumption that $(F_t)_t$ is stationary in the literature but it's not clear to me why this assumption is meaningful. So, what are the properties or advantages of the model if I chose $(F_t)_t$ to be a stationary process?

$\endgroup$
  • 1
    $\begingroup$ I don't know about "often". The basic structural time series models have state variables that are explicitly nonstationary. For example, the local level model has $y_t = \mu_t + \epsilon_t$ with $\mu_t = \mu_{t-1} + \eta_t$. Here the state variable $\mu_t$ is a random walk. $\endgroup$ – The Laconic Apr 1 '18 at 16:42
  • 2
    $\begingroup$ When the model is stationary, the Kalman Filter (KF) converges, which means that the prediction and filtering covariances P(t|t−1) and P(t|t) converge for t→∞. So if you find the stationary distribution, the KF update can be avoided in some cases. $\endgroup$ – Yves Apr 2 '18 at 15:06
  • $\begingroup$ Thanks you @Yves, do you have a good reference for this finding? $\endgroup$ – ReinerZufall Apr 2 '18 at 17:51
  • 1
    $\begingroup$ @Larry This relates to the so-called Riccati equation. See e.g. the famous book Harvey, A. Forecasting Structural Time Series Models and the Kalman Filter. I would encourage you to consider the stationary $\text{AR}(p)$ where the convergence is reached (exactly) in $p$ steps. $\endgroup$ – Yves Apr 2 '18 at 18:22
1
$\begingroup$

Kalman filter is a tool that can be still optimal even with non-stationary state or noise. The assumption of being stationary often aims to compute some specific equation. For example, in the state equation, $F_t = \Phi F_{t-1} + v_t$, if we want to compute the autocorrelation matrix $\mathbf{E}[F_tF_t^T]$, it can come from $\mathbf{E}[F_tF_t^T] = \Phi \mathbf{E}[F_{t-1}F_{t-1}^T] \Phi^T+ Q_t$ by setting $\mathbf{E}[F_tF_t^T] = \mathbf{E}[F_{t-1}F_{t-1}^T]$ (can only be adopted if $F_t$ is stationary).

But note that Kalman filter does not require the state to be stationary. If it does, most likely it needs to obtain a solution to some specific equations.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.