Exponential weighting scheme?

[PCA]

If I am running a time series OLS regression, and I want to place more emphasis on the most recent observations, how do I go about doing that?

I have an exponential weighting series(1.07, 1.15, 1.23, 1.32, 1.41, 1.52, etc...), but I am not sure how to implement this into EViews. I know that there is a weight option in the "estimate equation" pop-up box, but I believe that is for attempting to solve for heteroskedasticity, as opposed to just placing more emphasis on the latter observations, and less on the earlier ones. Am I wrong?

Any insight into this would be greatly appreciated.

[EViews Gareth]

Weights are weights. Using weighted least squares to solve heteroskedasticity works by placing more emphasis on some observations than on others, which is exactly what you want to do.

[PCA]

Which type of type would I be selecting then? Inverse std dev? Inverse variance? Std dev? Variance?

My weight series (which is loaded up as series "WEIGHT" in my workfile) exponentially goes from 1 to 388.02 (82 observations, 20+ years of quarterly data). So in otherwords, I want the last observation to have 388x more weight than the first observation.

Which scaling method would be utilized?


Can you utilize a weighting scheme, as well as the HAC (Newey-West) correction (if AC and/or HD are present)? That seems like double weighing to me.

[EViews Glenn]

There is an issue as to what you want weighting to do which you have correctly identified...

Conceptually, the weights in EViews are designed to return the residuals to homoskedasticity. It sounds as though you want weighting of observations for a different purpose. I will note that applying your exponential weights to homoskedastic data using the EViews options will introduce heteroskedasticity in the residuals.

You say that you want to "place more emphasis on some observations" in estimation. That mandate is pretty vague and there are a number of ways to do it (one being weighted least squares). What mechanism did you have in mind for using your exponential weights to vary the importance of observatinos...

[PCA]

As commonly known, things change over time. If I have time series data (dependent variable and a few explanatory variables) that span 20 years (quarterly data, so 80 or so observations), the observations that occurred 20 years ago will have the same impact on the model as the observations that occurred last quarter. I want to place more emphasis on the most recent observations.

An example of a weighting scheme would be (overly simplified):
1
2
4
8
16
32

So the first observation (all variables in 1990 quarter 1) receive a 1/63 weight, the second observations receives a 2/63 weight, and the last observation (all variables in 2011 quarter 2) receives a 32/63 weight. This is obviously exaggerated, but that is essentially the basic design that I'd like to examine. Is this feasible?

[PCA]

Does something like this even coincide with the basic tenets methodology of OLS?

[EViews Glenn]

Not really. That was my point. I'm not clear as to how you want to adjust the ideas underlying least squares to fit your situation...

There are algorithms for downweighting observations based on residuals that could be adapted for your setting, but until you have a clear idea of what it is you want to do, it's a little difficult to tell you how you might proceed....