Recursive least square estimation

[auvernois]

Hi,

I'm doing a simple OLS regression, and I'd like to estimate and store coefficients recursively, using either an anchored or a rolling estimation period.
I know I could do this into a loop, but it seems that I could do it faster using the "rls" function.
So far, I'm doing this:
equation eq1.ls log(ser1/ser1(-60)) = c(1) + c(2) * log(ser2/ser2(-60))
eq1.rls(c)
2 questions:
- It seems that this "rls" function is using an anchored estimation period, am I right ? Can we use it with a rolling window ?
- The coefficients are plotted on a graph ,but I can't find the underlying series, how do I retrieve them (I actually do not need the chart, just the series).

Thanks :D

[EViews Gareth]

RLS won't do what you want. However there are numerous examples of doing rolling regression here on the forum, plus there is an EViews add-in that does it.

[auvernois]

Thanks for your quick answer :D !
Actually, I think my explanation wasn't that clear, because I think "rls" would do what I want.
I'd like to simulate doing this regression everyday for the past 50 years, and construct series of coefficients that do not contain forward looking data (coef at time t is computed by doing the regression on a sample that includes only information known up to time t, and so on).
I tried again and replaced rls(c) by rls(c,s) and it seems to be working: I have my series R_C1 and R_C2 with the 2 coefficients over time.
Am I wrong or using this rls function is the equivalent of doing ls regressions in a loop, expanding the estimation sample by a day at each iteration, and storing the coefficients in time series ?
Thanks

[EViews Gareth]

RLS won't do rolling, only anchored.

So your description of what it is doing is correct, but if you want rolling, not anchored, you'll have to use rolling regression.

[auvernois]

OK. I'll just stick with anchored for now.
Many thanks.