2
$\begingroup$

The classical autoregressive model is a linear model for the dynamic variable $x$, where the added noise $\epsilon$ is directly affecting the dynamics of the model $$x_{t} = \sum_i \alpha_i x_{t-i} + \epsilon_t$$ I am considering a related model, where the noise is only due to imperfect observation of the underlying variable, but does not directly affect the dynamics. Then, the dynamics of a hidden variable $x$ is given by

$$x_t = \sum_i \alpha_i x_{t-i}$$

and of the observable $y$

$$y_t = x_t + \nu_t$$

where $\nu_t$ is the observable noise.

Questions:

  • Under what name is the "observable noise AR model" known in the literature?
  • Under what name is the composition of the above two models known in the literature? That is, a model that possesses both dynamic and observational noise
$\endgroup$
  • $\begingroup$ In your first equation, $\epsilon_i$ is supposed to be $\epsilon_t$ $\endgroup$ – Firebug Oct 19 at 15:57
  • $\begingroup$ You seem to be looking for state space models. $\endgroup$ – Chris Haug Oct 19 at 17:23
  • $\begingroup$ @Firebug, it was a typo, thanks for spotting $\endgroup$ – Aleksejs Fomins Oct 20 at 7:07
2
$\begingroup$

Consider the model \begin{cases} x_t = \sum_i \alpha_i x_{t-i}\\ y_t = x_t + \nu_t \end{cases}

Substituting $$y_t = \sum_i \alpha_i x_{t-i} + \nu_t\\ y_t = \sum_i \alpha_i (y_{t-i}-\nu_{t-i}) + \nu_t\\$$

This is an ARMA model in $y$ with a very specific constraint.

\begin{cases} y_t = \sum_i \phi_i y_{t-i} + \sum_i \theta_i \nu_{t-i} + \nu_t\\ \theta_i=-\phi_i \end{cases}

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Nice observation. What about the composite model? I have tried the same math as you did for composite model, and it looks almost like ARMA, but there are two sources of noise, one window-averaged and another instantaneous $\endgroup$ – Aleksejs Fomins Oct 20 at 7:22
  • $\begingroup$ @AleksejsFomins how did you implement the other model given that $x$ is hidden? $\endgroup$ – Firebug Oct 20 at 11:16
  • $\begingroup$ The point is to infer the hidden variables $x$ from observations $y$. The variables form a closed system, given a prior for $x_0$, $\alpha$ and the noise model. In case the observable is obtained by means of a function of the hidden variable, not just the noise, the model is called Dynamical Causal Model, however it is only popular in neuroscience community. I was wondering if I can find analogues of this model in regular statistical literature. $\endgroup$ – Aleksejs Fomins Oct 20 at 11:27
  • $\begingroup$ Also, could you suggest a good read for learning about different existing time-series statistical models like ARMA. $\endgroup$ – Aleksejs Fomins Oct 20 at 11:32
  • $\begingroup$ @AleksejsFomins the DCM assumes continuous time (I'm a neuroscientist). It has a stochastic variant with neuronal noise as well (Stochastic DCM). Having said that, in DCM $x$ is latent, but indeed partially observable (due to the assumption that it's zero at the onset of sessions). $\endgroup$ – Firebug Oct 20 at 12:38

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.