Re: st: More efficient processing of nested loops?

[Nick Cox <njcoxstata@gmail.com>]

The good news is that I think you are right. This code appears to be
much more complicated than it needs to be.

I can't follow your word description -- doesn't mean it's unclear,
just means that it is too much for me to absorb -- but from looking at
your code there are several major and minor inefficiencies.

The main problems are

1. You have two outer loops tangled together, one Stata's and one
home-made, but neither appears necessary.

2. The inner loop is a loop over one case, and so not needed.

3. -egen- calls are very inefficient to calculate constants that
-summarize- can calculate, except that #1 implies to me that you can
do most of the work in a few -egen- calls.

4. Some copying of values from one place to another to no obvious purpose.

With some guesswork, your code to me boils down to

egen d2_temp_2 = min(datadate_2), by(cum_id_cum)
egen dt_temp_2 = min(date_t_4), by(cum_id_cum)
egen cusip_mode = mode(cusip), by(cum_id_cum)
local mark "date <= d2_temp_2 & date >= dt_temp_2 & cusip_mode==cusip"
egen cum_ab_temp = total(ab_ret) if `mark', by(cum_id_cum)
replace cum_ab = r(sum) if `mark'

Note that your use of "temporary variables" is not the same as Stata's.

Nick

On Wed, Aug 15, 2012 at 4:26 PM, Robson Glasscock <glasscockrc@vcu.edu> wrote:

> I am running Stata 12. I have written code that creates a variable
> that is the sum of abnormal returns for each firm. The abnormal
> returns are accumulated from three days after the firm makes its
> earnings announcement in quarter t-4 to three days after the firm
> makes its earnings announcement in quarter t. My problem is that it
> takes around 2 minutes for my code to execute for each firm/quarter
> announcement, and there are around 150,000 earnings announcements in
> the dataset.
>
> I modified a panel dataset with the earnings announcement dates for
> each firm so that each observation contains both the quarter t
> earnings announcement date and the quarter t-4 earnings announcement
> date (with the three days added to each per above). I then merged this
> dataset with a second panel dataset that contains the daily returns
> for each firm. The merged dataset has around 10.2 million
> observations.
>
> Next, I created a count variable, cum_id_cum, which is a running total
> of each earnings announcement. This variable is truly cumulative and
> does not reset back to 1 when the next firm releases its first
> earnings announcement. The loop contains a variable, runn, that starts
> with a value equal to "1" and increases by 1 each time the loop is
> processed. I'm using that to help identify the particular dates of the
> quarter t and quarter t-4 earnings announcement so that the abnormal
> returns are accumulated over the correct period. Datadate_2 is the
> quarter t earnings announcement and  date_t_4 is the quarter t-4
> earnings announcement. Cusip is the identifier for each firm. Date is
> the date of the firm's return in the stock market for each trading
> day.
>
> The big picture of my approach was to create temporary variables that
> will equal the cusip, datadate_2, and date_t_4 when runn equals
> cum_id_cum. These first-step temporary variables (d2_temp, dt_temp,
> and cusip_temp) have missing values except in the observation where
> runn equals cum_id_cum so I created second-step temporary variables
> (cusip_temp_2, d2_temp_2, and dt_temp_2) which place the cusip,
> datadate_2, and date_t_4 values for each particular run through the
> entire dataset. This allows me to mark the days for each firm that are
> between the dates of the earnings announcements and then sum up the
> abnormal returns in a temporary variable called cum_ab_temp. The final
> variable with the sum of the abnormal returns for each firm is cum_ab
> and is retained in the observation where _merge==3 (from the merge
> mentioned in the second paragraph of this post).
>
> My code is below. Note that I constrain it to the first 25,000
> cum_id_cum values due to macro size constraints:
>
> gen runn= 1
> levelsof cum_id_cum if cum_id_cum !=. & cum_id_cum <25000, local(cum_id_cum)
> foreach 1 of local cum_id_cum{
> gen d2_temp= datadate_2 if runn== cum_id_cum
> gen dt_temp= date_t_4 if runn== cum_id_cum
> gen cusip_temp= cusip if runn== cum_id_cum
> egen cusip_temp_2= mode(cusip_temp)
> egen d2_temp_2= min(d2_temp)
> egen dt_temp_2= min(dt_temp)
> foreach x of varlist date{
> replace mark= 1 if `x' <= d2_temp_2 & `x' >= dt_temp_2 & cusip_temp_2==cusip
> egen cum_ab_temp= total(ab_ret) if mark==1
> replace cum_ab= cum_ab_temp if datadate_2== d2_temp & dt_temp==
> date_t_4 & cusip_temp==cusip & d2_temp !=. & dt_temp !=.
> replace runn= runn+1
> drop d2_temp
> drop dt_temp
> drop d2_temp_2
> drop dt_temp_2
> drop cusip_temp
> drop cusip_temp_2
> drop cum_ab_temp
> replace mark=0
>
> }
> }
>
> I'm wondering if there is a more efficient way to do the above which
> will result in a significantly faster processing time. My fear is that
> the above ignores the functionality of Stata and instead uses
> inefficient brute force.
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/