Time Series regression

[RDellaVilla]

Hello, I have to perform this following task in a replication step aimed at a research. 

"To compute FFC-adjusted returns for day t, I first estimate individual stock factor loadings by regressing delisting-adjusted excess returns on the FFC four factors (including an intercept) on a 120-trading day rolling window from t –150 to t –31 trading days for each stock".

I think I have properly sorted out my dataset so as to have the data in the desired form. Still, I have troubles in performing this time series regression. Moreover, my sample is really big, so I would require something which does not takes ages to run. 

Any suggestions on how to carry on? 

Thanks,

Riccardo

[sbxkoenk]

Hello,

If you have your input data set right (I guess you shifted the target column such that the x-vector for day [t-31] is in the same observation as the target y for day [t]?), then use THE BRUTE FORCE APPROACH as explained in this paper:

Rolling Regressions with PROC FCMP and PROC REG
Mark Keintz, Wharton Research Data Services, University of Pennsylvania

https://www.lexjansen.com/nesug/nesug12/fi/fi08.pdf

After creating the data set RWIN (TYPE=VIEW), like done on page 1, call your procedure (for example proc autoreg) with the by-variable "w".

This BRUTE FORCE APPROACH is of course greedy. Much more efficient techniques can be programmed.

Let us know if the real-time (time-to-completion) of your PROC by w takes too long.

By the way, for time series regression it's better to post your question in the SAS Forecasting and Econometrics board.

Good luck,

Koen

[RDellaVilla]

Thank you Koen for the help.

I have tried and it is not actually working properly. Indeed, it takes quite a long time as my dataset is big, not to say enormous.

I will try to look for something a little bit more efficient. Still, thank you for this BRUTE FORCE APPROACH.

[RDellaVilla]

To compute FFC-adjusted returns for day t, I first estimate individual stock factor loadings by regressing delisting-adjusted excess returns on the FFC four factors (including an intercept) on a 120-trading day rolling window from t –150 to t –31 trading days for each stock

this is exactly my issue. 

I feel I have sorted out data properly, setting the date of interest t, and the estimation window for each date t. I now have predict the value in t with the estimation period data. 

[mkeintz]

You want to make rolling window datasets (or even rolling window SSCP's - sum-of-squares-and-cross-products) to perform your regressions, correct?

You can take a look at these papers/presentations/comments I made on this subject:

1. Philadelphia SAS User's Group (2019) - slide deck.
   

   Rapid Rolling Window Regressions via Home Made Sum of Squares and Cross Products (philasug.org)
   
   

2. Northeast SAS Users Group (2012) - paper presentation
   

   Rolling Regressions with PROC FCMP and PROC REG (lexjansen.com)

3. SAS Communities forum
   

   How to create rolling windows with different numbe... - SAS Support Communities