2
$\begingroup$

I have a question regarding some time series concepts:

Suppose I have some "time series" data with cross correlation. Suppose I am able to fit a copula, say to capture dependencies between data of $\{t,t-1,...,t-n\}$ - i.e. for some time-lag. Now, if I want to draw some random numbers of this time series, I could do the following (ignoring the marginal distributions):

If I have the sample of the time-lag, e.g. $X_{t-1},...,X_{t-n}$, then I could draw $X_t|X_{t-1},...,X_{t-n}$, if I know the conditional distribution (the easiest case is a multivariate normal).

What is the closest time series concept to this? Is there any buzzword I have to look for? I used this concept to do a quick analysis of the data, but I'd like to know if this is a common and reasonable approach.

$\endgroup$
  • $\begingroup$ You probably want to look at state-space models, which provide a way of doing what you want. $\endgroup$ – F. Tusell May 5 '14 at 10:41
  • $\begingroup$ Could you elaborate on this a little bit more? Thank you. $\endgroup$ – user13655 May 5 '14 at 11:28
  • $\begingroup$ I cannot elaborate much more in the space of a comment; see answer below. Note I describe the simplest case; neither linearity nor gaussianity are required (although life is simpler with those two assumptions). $\endgroup$ – F. Tusell May 5 '14 at 14:36
2
$\begingroup$

In the simplest linear case, a state-space model is composed of two equations, the state equation (or transition equation), $$ \boldsymbol{\theta_t} = \boldsymbol{T_t}\boldsymbol{\theta_{t-1}} + \boldsymbol{R_t}\boldsymbol{\nu_t}$$ and the observation equation, $$\boldsymbol{Y_t} = \boldsymbol{F_t}\boldsymbol{\theta_t} + \boldsymbol{\epsilon_t}.$$ Vector $\boldsymbol{\theta_t}$ is termed the state-vector, and embodies all information that describes the state of the system at moment $t$. It evolves over time following the dynamics given by the state-equation.

The state vector is in principle made of latent, unobservable variables; what you do observe is $\boldsymbol{Y_t}$, which are linear combinations of elements of the state vector corrupted by noise $\boldsymbol{\epsilon_t}$. The specification is completed with the distribution of the state and observation noises ($\boldsymbol{\nu_t}$ and $\boldsymbol{\epsilon_t}$), usually gaussian distributions in the linear case.

Now, if you consider the linear time-invariant case ($\boldsymbol{T_t} =\boldsymbol{T}, \boldsymbol{F_t} = \boldsymbol{F}$ for all $t$), you have something very close to what you want; you would estimate the needed parameters, in case there are some. The best estimator of $\boldsymbol{Y_t}$ would be $\boldsymbol{F}\boldsymbol{\hat\theta_t}$ and $\boldsymbol{\hat\theta_t}$ can be estimated in terms of $\boldsymbol{Y_{t-1},\ldots}$ using the Kalman filter.

This is only a telegraphic account. For a good discussion see for instance Durbin and Koopman or Anderson and Moore among many books on state-space time series methods.

$\endgroup$

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.