I am currently working on developing a series of density forecasts and I am encountering some problems. I am working on weekly S&P 500 returns and the returns process is described as

$r_{t} = \mu + \delta r_{t-1} + h_{t}z_{t}$ where $z_{t}$ comes from the Gaussian distribution.

I am forecasting the returns and volatility of the series using the ARMAX-GARCH-K Toolbox in Matlab. Initially I estimate the ARMA(1,0)-GARCH(1,1) model and obtain the one-step ahead forecast of the returns and volatility. I obtain the $\mu$ parameter from the parameters of the GARCH model.

As far as I understand the next step to obtain the density forecast (assuming Gaussian distribution) is to use the pdf of Gaussian distribution, so the

normpdf(x, mu, sigma)

function in Matlab.

And here is the essence of my trouble. To obtain the density forecast should I use the actual observed return as x or the point forecast from the garch model? And should I use the $\mu$ parameter from the GARCH model as input for the mu parameter of normpdf and the forecasted volatility as as sigma in the normpdf function?

I now this may be a basic question but I cannot find any elementary examples on the internet.

Kind regards


1 Answer 1



I read more about it and I get some help with someone else, here is the correct answer :

The density forecast is the predictive likelihood value of the process estimated at the realized value computed in a one step ahead way.

Thus for instance for a standard arma garch process with normal errors:

  1. You forecast the mean $u^{f}_{t|t-1}$ and variance $v^{f}_{t|t-1}$ process at time t-1 for time t
  2. the predictive density forecast for time t of the realized value $u_{t}$ is $N(u^{f}_{t|t-1},v^{f}_{t|t-1})$
  3. the predictive density forecast $u_{t}\sim N(u^{f}_{t|t-1},v^{f}_{t|t-1}) $ is equivalent to the predictive residuals density forecast : $r_{t} = u_{t}-u^{f}_{t|t-1}\sim N(0,v^{f}_{t|t-1}) $
  4. the density forecast is the density of $r_{t}$ with respect to a $N(0,v^{f}_{t|t-1}) $

Note that it is very similar to the "usual" likelihood except you are estimating the model in a one step ahead way (the parameters are re-estimated at every step)

Previous Post :

It is not a "basic question", If I am correct :

First you estimate your model on the return series and obtains parameters. You must estimate your model in such a way you obtain one-step ahead errors (that I will call computed errors in what follow) and associated time-serie of the predictive errors distributions parameters: $\hat{\mu_{t}}$ and $ \hat{\sigma_{t}^{2}}$ (note : these are not the parameters of the mean and variance processes but parameters of the error distribution).

I would take the original return serie minus the fitted returns to obtain the computed errors :

$$\hat{e_{t}} = r_{t} - \hat{r_{t}}$$

These computed errors should behave accordingly your predictive (errors) density which is defined by parameters you obtain in the first step ($\hat{\mu_{t}}, \hat{\sigma_{t}}$) : Note the subscript $_{t}$ for parameters !

Then I would compute the predictive density of these residuals using parameters obtained in the first step estimation, indeed if your forecasts are accurate the density of $\hat{e_{t}}$ should ideally be equal to the predictive density defined as: normal($\hat{\mu_{t}}, \hat{\sigma_{t}}$).

If it fit perfectly (they have the same mean, variance) then the density forecast will returns a high mass . If the model is misspecified, the errors $\hat{e_{t}}$ will fall outside of the range implied by your predictive error density and then it will assign very small probability.

So in your case I will use the following function (again note the time subscript)

Density forecast ($\hat{e_{t}}$) = normpdf($\hat{e_{t}}, \hat{\mu_{t}}, \hat{\sigma_{t}}$)

  • $\begingroup$ Thanks for your answer. I have one more question: how to obtain time-series of the predictive errors distribution parameters $\hat{\mu_{t}}$ and $\hat{\sigma^2_{t}}$? $\endgroup$
    – Masher
    Nov 27, 2015 at 22:50
  • $\begingroup$ @Masher, I think my answer is wrong, I will correct it monday or tuesday and I 'll provide more details about implementation. Meanwhile you can have a look page 19 in the following paper : Comparing Density Forecasts via Weighted Likelihood Ratio Tests Asymptotic and Bootstrap Methods. (January 2002 version ) $\endgroup$
    – Malick
    Nov 27, 2015 at 23:42
  • $\begingroup$ @Masher see also stats.stackexchange.com/questions/164380/density-forecasting $\endgroup$
    – Malick
    Nov 27, 2015 at 23:48
  • 1
    $\begingroup$ I have read the post and the paper you mentioned and my conclusion would be to do as follows: estimate the GARCH model (analogically to the paper let's say with t-distributed innovations), obtain the volatility forecast for time t+1, standardize the actual (observed) return for period t+1 using the mean parameter from the GARCH model and standard deviation forecast for time t+1 and finally use tpdf with the first input being the standardized actual return for period t+1 and the second one corresponding to the degrees of freedom obtained from the GARCH model. $\endgroup$
    – Masher
    Dec 2, 2015 at 14:27
  • $\begingroup$ @Masher yes, this is correct see my edit in the post $\endgroup$
    – Malick
    Dec 3, 2015 at 14:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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