EViews.com

Hi everyone!

I have an ARIMA (2,1,0) that, unfortunately, doesn't have a normal distribution. On the boxplot of the DX series I can see several outliers, which I would like to correct by using dummy variables. How can I see which months in the time series are the outliers? The boxplot doesn't show that information when I hover over the data points... Thanks!

You should decide a rule/criterion to detect the outliers. For instance, if you are interested in the observations that fall ±4 standard deviation outside the mean: You can also examine the residuals of your model in that respect. Finally, there is a more formal EViews procedure written by George S. Ford, which you may find very useful: Outlier Statistics using EViews

Thank you!

Hi Trubador, is it possible to perform outlier detection in GARCH type models with Eviews? for example based on Franses (1999) method... Im using Eviews 8, thank you.

Yes, it is. But it may require extensive programming. I also suggest the following approach: http://www.economics.ox.ac.uk/Nuffield- ... rch-models

Yes, it is. But it may require extensive programming. I also suggest the following approach: http://www.economics.ox.ac.uk/Nuffield- ... rch-models

Thanks Trubador, the Influence Statistics that are built in Eviews to detect the outliers are suitable for linear time series, is it true? I would like to know is there codes available for detecting outliers in Non-linear time series by Eviews?

Depends on the type of nonlinearity. Influence statistics will still be of use where you can estimate your model via NLS. In case of Markov-type switching or GARCH-type conditional variance models, however, they will not work. Outliers are specific to the model at hand regardless of the linearity assumption, and therefore, there is more than one way to detect outliers. Doornik and Ooms (2005) approach for GARCH models, for instance, can be implemented in EViews relatively easily.

Depends on the type of nonlinearity. Influence statistics will still be of use where you can estimate your model via NLS. In case of Markov-type switching or GARCH-type conditional variance models, however, they will not work. Outliers are specific to the model at hand regardless of the linearity assumption, and therefore, there is more than one way to detect outliers. Doornik and Ooms (2005) approach for GARCH models, for instance, can be implemented in EViews relatively easily.

Dear Trubador, many thanks for introducing me Doornik and Ooms (2005) paper, it was very interesting and easy to implement in Eviews. I only have some questions I hope you can help me. First, I would like to know that is there any command in Eviews that I can find the largest values in observation? because its not easy to find the highest observation in daily returns visually.

and another question is about the process of correcting data. In the paper they explain that their method detects two kinds of additive outliers: additive level outliers and additive volatility outliers. If the type of outlier is level, then correcting the data is easy just raw data (in the date of outlier) minus coefficient of dummy, but if the outlier is volatility outlier then the process of correcting is more complicated and as the authors said it needs modification in GARCH codes but they dont explain anything about it! I would like to know do you have any idea about it?

Thank you in advance

I mean in general I am not sure how to correct the data, do you think only excluding the outlier can be an appropriate solution?

First, I would like to know that is there any command in Eviews that I can find the largest values in observation?

@max is the function you need. For a detailed list of EViews' functions, please see the help file.

I mean in general I am not sure how to correct the data, do you think only excluding the outlier can be an appropriate solution?

Outlier detection procedures will help determine the number and the location of outliers, if any. How you will proceed after that is just as important. As I mentioned in my previous post, outliers are specific to the model. You may not have enough information to capture the behavior of a specific observation, which is why it turns out to be an outlier. In such cases, using dummy variables instead of excluding the outliers might be a better choice. Or adding exogenous variables (other than dummies) to the model may also help explain the so called "anomaly".