GARCH (1,1) results (PLEASE HELP)

[Leif85]

Hey!

I'm estimating volatility for financial time series data, where exists values of dates and returns.

I was just wondering, if the first value that you insert to eviews is the daily index return of date 1.1.2008 for example, and the eviews gives you the volatility time series where the first estimated volatility is for 1.1.2008? Is it true that the first volatility value that Eviews can give is the for 2.1.2008, because it always needs the return for previous day?

my model formula is form index name_c

[trubador]

EViews computes the backcast variance via exponential smoothing using the coefficients from mean equation. So you'll have an estimated initial variance value for the observation 1/1/2008. Rest of the conditional variance values are generated through the estimated variance equation, the initial variance value being the starting point.

[Leif85]

Vuu, thanks, one of my teachers said something else, but my I'm sure that you're right because another teacher answered similarly.

[w2quesang]

Hi,

I am also kind of newbie in the GARCH model. My question is, how will you be able to produce a graph, a forecasted graph with the realized volatility below the estimated volatility. Similarly, how can I compute for the MAPE in Eviews?

Because, I used the forecast key in the GARCH model? and then I clicked on the static, I only came up with an estimated volatility. But I need to present it with the realized volatility so that I can compare it effectively. Also, I realized that my MAPE is around 102. Simply because in think the realized volatility is in a straight blue line. I hope you guys can help me out. Thanks.

Regards,

William

[trubador]

There is no realized volatility in garch modeling, since the whole process of estimation is based on "conditional" volatility, which is unobserved and therefore has to be derived. For realized (or observed) volatility, you should decide or choose a measure in order to make a comparison. For instance, if you are using daily exchange rate returns as the dependent variable, you can calculate moving variance (e.g. for 22 days) as a proxy. Or, better yet, you can calculate intraday variances, if you have access to that type of data.

[w2quesang]

So you mean to say, to get the realized volatility, I should run a MOVING AVERAGE model on the actual returns to have it and then compare it with the conditional volatility? Thanks.

May I know how to run the moving average model? the one your suggesting? Thanks

[trubador]

Suppose that you have a daily return series called y and that you already estimated a GARCH(1,1) model and saved the conditional variance as garch01. For expositional purposes, you can compare the results with a 5-day moving variance series of y. In order to do that, you can enter the following code to the command line:

Code: Select all

plot garch01 @movvar(y,5)

[w2quesang]

Hi Trubador,

Your are soooooo good. Yes so far, I have done a comparative graph between the "conditional" and the "realized - ala moving average". Trubador, i have an inquiry over the number of days to be used for the moving average. You see, I intend to have an analysis based on daily, weekly(5days) and monthly data series. But the thing is, I cannot use 1 day as the number in the movvar. So I was wondering what do you think might be the best average days for these time series?

Also. how can I have the table form of my movvar results? are there special commands to it?

Is the resulting graphs for the GARCH the same as GARCH-M? Thanks Man. Your the BEST !!!

[trubador]

If the original frequency of your weekly and monthly data series is daily, then 5-day and 22-day (non-moving) variances of each week and month will be your realized volatility, respectively. Number of days for the latter depends on the nature of your data and can be less or greater than 22 days, which is chosen to represent the approximate number of working days in a month. For a more precise calculation of the number of (working) days, you should check Chris's Date Arithmetic.

Ideal solution for daily data would be to use intra-day observations to calculate the realized daily volatility. Anyway, if you do not have your data in higher frequencies, then you can use moving variances instead. You can either use backward or centered moving variances. You should keep in mind that while the former creates a phase difference, the latter leads to missing values at both ends. Besides, there is no built-in function for calculating the centered moving variance in EViews, and you'll have to compute it via help of other functions.

The longer the period, the smoother the moving variance will be. So, you should better use a short period. How short is "short" is difficult to answer, though. If you do not have a priori reasoning or information, then you can do an OLS between the conditional and realized variance series to decide the optimum period. Either use the Equation Editor or simply run the code below for different (and integer) values of n, and choose the one that has the highest R-squared value.

Code: Select all

equation eq_n.ls garch01 c @movvar(y,n)

Since you do not need to estimate the above equation for longer periods, you can do it manually by running it several times. If you wish to automate the process, you can create a loop via writing a short programming code.

If you want to tabulate your results, you can use the Table object in EViews. It has many properties, so you should first decide its design and then declare one to suit your own need.

GARCH-M and GARCH models are different, so they'll probably yield different estimated conditional variances.

[w2quesang]

Thanks trubador. : )

[w2quesang]

Hi just wanna ask something.

Is the result of the GARCH model "nearly the same as the "GARCH-M" ? meaning there's only a slight difference between them?

Because it seems, the output of my GARCH-M is just around double of what was the result of my GARCH model. not nearly double....it just simply look like double. : )

Thanks