I'm working on one project trying to reconstruct a sequence of multivariate signal data from another sequence of multivariate signal data. That is let $\{S_t\}_{t=1}^n$ be the first sequence of signals, where each $S_t$ is a m-by-1 vector and $\{r_t\}_{t=1}^n$ be the second sequence of signals, where each $r_t$ is a n-by-1 vector, we are trying to estimate $\{S_t\}_{t=1}^n$ using $\{r_t\}_{t=1}^n$.
For now, I'm modeling this problem as $\begin{cases}S_t=\phi_1S_{t-1}+\cdots+\phi_pS_{t-p}+\epsilon_t\\ r_t=\psi_0S_t+\psi_1S_{t-1}+\cdots+\psi_{p-1}S_{t-p+1}+\eta_t \end{cases}$, where $\epsilon_t\sim N(0,\Sigma_\epsilon)$, and $\eta_t\sim N(0,\Sigma_\eta)$ and they are independent of each other. And all the $\phi_i$'s and $\psi_i$'s are matrices.
By letting $\xi_t=\begin{bmatrix}S_t\\ S_{t-1}\\\vdots\\S_{t-p+1} \end{bmatrix}=\begin{bmatrix}\phi_1&\cdots&\phi_{p-1}&\phi_p\\ I &\cdots&0&0\\ \vdots&\cdots &\vdots&\vdots\\ 0&\cdots &I&0\end{bmatrix}\begin{bmatrix}S_{t-1}\\S_{t-2}\\ \vdots\\ S_{t-p} \end{bmatrix}+\begin{bmatrix}\epsilon_t\\\epsilon_{t-1}\\\vdots\\\epsilon_{t-p+1} \end{bmatrix}$, I converted the first equation into $\xi_t=F\xi_{t-1}+\mathscr{E_t}$. And the second equation can be written as $r_t=G\xi_t+\eta_t$, where $G=\begin{bmatrix}\psi_0&\cdots&\psi_{p-1} \end{bmatrix}$. Combining these two equations, I got $\begin{cases}\xi_t=F\xi_{t-1}+\mathscr{E_t}\\ r_t=G\xi_t+\eta_t \end{cases}$. This resembles a Kalman filter where I treat $S_t$ as latent variables. However, I'm having trouble estimating these coefficients including the variance-covariance matrices of $\mathscr{E_t}$ and $\eta_t$.
My data set consists of two parts, where the first part (the training data set) I can actually observe both $r_t$'s and $S_t$'s while in the second part (the testing data set) I can only observe $r_t$'s. However, I think in the standard setting of a Kalman filter, we do not need to observe the actual values of the latent variables to estimate the coefficients. So how can I make use of this extra piece of information to estimate the coefficients better in the training data set and estimate $S_t$ using these estimated coefficients and new observed $r_t$'s to predict $S_t$'s in the testing data set.
Thank you!


Your Answer

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

Browse other questions tagged or ask your own question.