Detecting influential observations

[kpukthua]

Hi,

Does anyone know how we can detect influential observations from time series of data using Eviews?

For instance, I have two time-series daily return data of Argentina and Chile and they are highly correlated. I would like to check which observations causes such high correlation, so how can we detect them?

Your help will be appreciated. Thank you.

Kuntara
kpukthua@mail.sdsu.edu

[trubador]

Some of the outlier detection algorithms can be written in EViews' programming language. You can also try to estimate the true relationship between these variables via building a Quantile Regression model and check the residuals that fall outside the confidence interval. Other than that, you can also use the Robust Regression procedure, which is available in the Program Repository section of the forum. You can either check the weighted residuals (for the significant values) or the weight series itself (for the zero values) generated by the code.

[hetero]

Hi,

Another thing that you may do, if the purpose is to detect those observations which are highly correlated within your sample, is to estimate rolling correlations for a sample window of lets say 20 observations. I don’t know if my suggestion is of any help!

Regards

[kpukthua]

Thank you very much, Trubador and Hetero. I really appreciate your suggestions.