1
$\begingroup$

I am still new to Bayesian forecasting, so I am hoping to get some clarification on a simple concept (by the sounds of it).

Suppose that we are interested in forecasting some time series one-step ahead. For simplicity, we can assume the time series itself came from a normal distribution with known parameters. The one-step-ahead predictive likelihood (as I understand) is:

$ PL = p(y^{o}_t|y^{o}_{t-1}) = \int_\theta p(y_{t}^o|y^o_{t-1},\theta)\times p(\theta|y^o_{t-1})d\theta$

Which can be approximated by $M^{-1}p(y^{o}_{t}|y^{o}_{t-1},\theta^{(m)}_{t-1})$, for $m$ MCMC loops.

How is it possible to calculate $p(y^{o}_t|y^{o}_{t-1})$?! It's the probability from a normal distribution (i.e. 0). Unless this has some relationship to the likelihood being evaluated at some point, I have no idea what to do here or evaluate this with a given posterior.

Thanks in advance.

$\endgroup$
1
3
$\begingroup$

Most Bayesian time series models are based on dynamic models. For simplicity, let's consider the random walk with noise model.

$$ y_t = x_{t\phantom{-1}} + v_t \qquad v_t \sim N(0,V) \\ x_t = x_{t-1} + w_t \qquad w_t \sim N(0,W) $$

and $x_0\sim N(m_0,C_0)$. Since we have "known parameters", we know the values for $V$ and $W$. The one-step ahead predictive distribution is

$$ p(y_t|y_{t-1}) = \int \int p(y_t|x_t)p(x_t|x_{t-1})p(x_{t-1}|y_{1:t-1}) dx_t dx_{t-1} $$

Let $x_{t-1}|y_{1:t-1} \sim N(m_{t-1}, C_{t-1})$ where $y_{1:t-1}$ is all the data up to time $t-1$. Then the one-step ahead predictive distribution is $y_t|y_{t-1} \sim N(m_{t-1}, C_{t-1}+W+V)$.

Other dynamic linear models (with known parameters) will also have a normal one-step ahead predictive distribution. Non-linear models or linear models with non-normal errors will generally not have a closed form predictive distribution. If parameters are unknown, you will also have to integrate over your current uncertainty in those parameters, i.e. $p(\theta|y_{1:t-1})$ and this will also typically not have a closed form predictive distribution.

$\endgroup$
2
  • $\begingroup$ This is not the predictive likelihood however? This is a pdf. Or are you suggesting to determine the likelihood of $y_{t}|y_{t-1}$ $\endgroup$
    – akkp
    Feb 22 '15 at 21:56
  • $\begingroup$ I'm not sure what you mean by "predictive likelihood" since you state "known parameters" and you write $p(y_t|y_{t-1})$, so I assumed you were wondering about a one-step ahead predictive distribution for your data. $\endgroup$
    – jaradniemi
    Feb 23 '15 at 16:13

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.