ROLLREG - rolling (linear) regressions

[TomDoan]

Attached is a revision (modernization) of the Bank of Canada's ROLLREG procedure, which was written originally for version 4. This does three types of "rolling" linear regressions; the most important of which is done with the MOVE option; that does a moving window of a fixed width. (The ADD option can be done more efficiently and with more options using the RLS instruction).

rollreg.src

(7.67 KiB) Downloaded 213 times

Revised May 2011 to allow saving the history of F-statistics

This is an example of its use (using the same data set as the CONSTANT example)

testrollreg.rpf

(1.42 KiB) Downloaded 312 times

Syntax (except for options, similar to LINREG):

@ROLLREG( options ) depvar first last
# list of regressors

You choose which of the three modes with:

ADD=end of initial period.
Starting with a regression from first to ADD, adds one observation at a time to the end of the range until it does a regression from first to last

DROP=start of last regression range
Starting with a regression from first to last, drops one observation at a time from the beginning of the range until it does a regression from DROP to last

MOVE=width of moving window
Starting with a regression from first to first+MOVE-1 adds observations to the end and drops them from the start one at a time until it does a regression from last-MOVE+1 to last (i.e. there are MOVE observations in each regression)

Other options:

ROBUST/[NOROBUST]Uses the ROBUSTERRORS option to calculate the standard errors.
LAGS = number of lags to use with the ROBUST option [0]This does **not** vary with the degrees of freedom.
LWINDOW=[NEWEY]/BARTLETT/DAMPED/PARZEN/QUADRATICChooses lag window type if LAGS>0.

Output Options:

COHISTORY = VECT[SERIES] for the coefficients
SEHISTORY = VECT[SERIES] for the standard errors of the coefficients
SIGHISTORY = SERIES for the regression standard errors
(These are all similar to the options for the RLS instruction)

FHISTORY = SERIES with regression F's

PRINT]/[NOPRINT] prints the coefficient estimates
WINDOW = title for print window
GRAPH/[NOGRAPH] graphs the coefficients with two standard error bands.

[iloverats]

dear all
in the rolling regression
y=c+a*x1+b*x2
null hypothesis a=b=0
how can i get the" F test" series when i use the "rollreg.src"

thanks

[TomDoan]

The calculations are all being done with LINREG instructions, each of which computes the %FSTAT variable. You just need to save those into a series, similar to what's being done with the SIGHISTORY option.

[challenges]

Dear all,
I have tried to implement, without succeeding, the F-test mentioned by iloverats in the rollreg.src given above. In fact, I have problems with storing the F-statistics for the joint null hypothesis testing (of a and b in the example of iloverats). I used the "exlude" command in the "Moving sample regression" part of the rollreg.src, but I do not know how to save the F-stats for each window. Could you please help me to implement the F-test for this joint null hypothesis in the rollreg.src ? Thank you in advance for your help.

[TomDoan]

Dear all,
I have tried to implement, without succeeding, the F-test mentioned by iloverats in the rollreg.src given above. In fact, I have problems with storing the F-statistics for the joint null hypothesis testing (of a and b in the example of iloverats). I used the "exlude" command in the "Moving sample regression" part of the rollreg.src, but I do not know how to save the F-stats for each window. Could you please help me to implement the F-test for this joint null hypothesis in the rollreg.src ? Thank you in advance for your help.

The procedure posted above now has an option for this:

FHISTORY=series of regression F's

[challenges]

Dear Tom,

Thank you very much for your help and your quick answer. I have tried to use the rollreg.src that you provided but it gives me the following error message "## SX11. Identifier %FSTAT is Not Recognizable. Incorrect Option Field or Parameter Order? >>>>istory(cut)=%fstat<<<<" I use the 6.2 version of Rats, if you need this information. Moreover, I need to implement a specific F-test that would test the null hypothesis b=c=0 in the equation y=a+b*x1+c*x2+d*x3+e*x4. The f-test must concern only the coefficients of the two first regressors and not all the coefficients. I suppose the "exclude" command must be used in the "Moving sample regression" part of the rollreg.src.Thank you again for your help.

[TomDoan]

Dear Tom,

Thank you very much for your help and your quick answer. I have tried to use the rollreg.src that you provided but it gives me the following error message "## SX11. Identifier %FSTAT is Not Recognizable. Incorrect Option Field or Parameter Order? >>>>istory(cut)=%fstat<<<<" I use the 6.2 version of Rats, if you need this information. Moreover, I need to implement a specific F-test that would test the null hypothesis b=c=0 in the equation y=a+b*x1+c*x2+d*x3+e*x4. The f-test must concern only the coefficients of the two first regressors and not all the coefficients. I suppose the "exclude" command must be used in the "Moving sample regression" part of the rollreg.src.Thank you again for your help.

%FSTAT isn't defined by 6.2, and you're correct that it won't give you the result from a partial exclusion. You would need to replace

if %defined(fhistory) ; compute fhistory(cut)=%fstat

with

exclude(noprint)
# X1 X2 (whatever they're actually called)
if %defined(fhistory) ; compute fhistory(cut)=%cdstat

every place the first line occurs in the revised ROLLREG,

[challenges]

Dear Tom,
Thank you very much for your help and your quick reply. I would like to know what are the instructions to be used in order to display in a column only the F-stats for each window (without displaying the marginal significance) stemming from the "Exclude" command. I have tried to define the Fhistory like a vector series with all the implications in the programme, but it doesn't work. Thank you in advance.

[TomDoan]

Dear Tom,
Thank you very much for your help and your quick reply. I would like to know what are the instructions to be used in order to display in a column only the F-stats for each window (without displaying the marginal significance) stemming from the "Exclude" command. I have tried to define the Fhistory like a vector series with all the implications in the programme, but it doesn't work. Thank you in advance.

What I described should work. You would just run rollreg with the fhistory option, then do a PRINT on the series you generated. If you couldn't adapt it to your problem, you'll have to send what you've done to support@estima.com and we'll take a look.

[challenges]

Dear all,

I have to implement in the rollreg.src a normality test of Jarque-Bera on the residuals of my regression at each iteration inside my loop. I do not need to save my residuals, but I need to print the statistics of my test as well as the p-values associated to the test. Could you please help me? I have tried to do it myself but I got errors in my program.
Thank you in advance.

[TomDoan]

Dear all,

I have to implement in the rollreg.src a normality test of Jarque-Bera on the residuals of my regression at each iteration inside my loop. I do not need to save my residuals, but I need to print the statistics of my test as well as the p-values associated to the test. Could you please help me? I have tried to do it myself but I got errors in my program.
Thank you in advance.

If you do a STATISTICS instruction on the residuals, it will produce %JBSTAT and %JBSIGNIF as the J-B statistic and p-value. You would just need to save those into a series the way the F-statistics are saved in the revised procedure that I posted.

[challenges]

Dear Tom,

Thank you very much for your quick reply and your help. It works.
Rats is great !

[basher]

Hi Tom,

I am using your revised rollreg procedure for a regression of "y" on "x". As I also want to perform a hypothesis test on the slope of "x" (i.e. b=1), I replaced

if %defined(fhistory) ; compute fhistory(cut)=%fstat

with

test(noprint)
# 2
# 1
if %defined(fhistory) ; compute fhistory(cut)=%fstat

every place the first line occurs in the revised ROLLREG. However, when I write "print FHISTORY", I get error message. Also, how can I obtain the associated standard errors for the slope coefficients (along with the f-stat)? Thank you.

[moderator]

Are you are writing the PRINT command exactly as

PRINT FHISTORY

If so, that's not the correct syntax. PRINT is expecting "start" and "end" parameters before the list of series, or a / symbol telling it to use the full range of the series. For example, try

print / fhistory

Regards,
Tom Maycock
Estima

[TomDoan]

Also, how can I obtain the associated standard errors for the slope coefficients (along with the f-stat)? Thank you.

Instead of trying to adapt the ROLLREG procedure to do so many other things, you're probably better off just doing the rolling regression as described in the User's Guide and doing whatever additional calculations you want after each regression.