0

Is there Hypothetical-Set Aggregate Function equivalent to ntile (or some other good solution) in Postgres?

I have this query:

select
    frctl
    ,*
from
    (select 
        *
     from
            d_al
     where
        not rtn is null
        and not fund_val is null
     ) dx
    ,lateral(
        select
            round(percent_rank(dx.fund_val) WITHIN GROUP (ORDER BY fund_val)::numeric
              , 6) AS frctl
        from 
            d_al
        where
            gen_qtr_end_dt <= dx.gen_qtr_end_dt
            and not rtn is null
            and not fund_val is null
    ) x
order by gen_qtr_end_dt_id, frctl

The query produces periodic historical percentile ranks. Ranking a value in a certain period/date relative to the current period/date plus all the historical periods/dates (periods before it) time series-wise/chronologically.

It works perfectly, except I want fractiles (i.e. the option to create deciles, quartiles, etc.) like ntile(#) does naturally. Do I have to build a case statement to fit the fractiles I want? For example, if I want ntile(4) (quartiles), do I have to build a case statement based off of 0, 0.25, 0.5,0.75,1. Then if I want ntile(10) (deciles), do I have to build a case statement based off of 0, 0.1, 0.2,0.3,0.4 ... etc? Or is there an ntile type Hypothetical-Set Aggregate Function I am missing?

Helpful links: https://www.postgresql.org/docs/current/static/functions-aggregate.html#FUNCTIONS-HYPOTHETICAL-TABLE

Percentile rank that takes sorted argument (or same functionality) in PostgreSQL 9.3.5 (In the link directly above the problem is a bit different, but very related.)

The data:

  • Big - efficiency is important, but not the focus of my question.
  • Table d_al has three columns, two matter here:
  • gen_qtr_end_dt - not unique, not null, type date
  • fund_val - can be null, type numeric
  • rtn - can be null, type numeric, not important here

I have Postgres 9.6.

PS - this query does all of the history, but my next step is to do a number of days rolling period look back (rather than all of the history).

edit 1: Here is how I am solving it now (with a case statement as mentioned):

I put the above query in an cte then...

   with pl as (
    select
        x.pctl
        ,dx.fund_val
        , dx.rtn
        ,dx.gen_qtr_end_dt
    from
        (select 
            *
         from
                d_al
         where
            not rtn is null
            and not fund_val is null
         ) dx
        ,lateral(
            select
                round(percent_rank(dx.fund_val) WITHIN GROUP (ORDER BY fund_val)::numeric
                  , 6) AS pctl
            from 
                d_al
            where
                gen_qtr_end_dt <= dx.gen_qtr_end_dt
                and not rtn is null
                and not fund_val is null
        ) x
)
-- , f as( 
    select 
        gen_qtr_end_dt_id
        ,case   when pl.pctl < 0.1 then 1
                when pl.pctl < 0.2 then 2
                when pl.pctl < 0.3 then 3
                when pl.pctl < 0.4 then 4
                when pl.pctl < 0.5 then 5
                when pl.pctl < 0.6 then 6
                when pl.pctl < 0.7 then 7
                when pl.pctl < 0.8 then 8
                when pl.pctl < 0.9 then 9
                else 10
         end
            frctl 
        ,rtn
        ,fund_val
        ,*
    from 
        pl
    order by
        gen_qtr_end_dt, frctl

...which is a bit cumbersome/rigid but doable if need be.

edit 2: And here is a sample of the output from edit 1 above:

frctl   fund_val    pctl    gen_qtr_end_dt
1   -14.514 0   3/31/2001
2   -8.618  0.142857    3/31/2001
3   1.707   0.285714    3/31/2001
5   26.162  0.428571    3/31/2001
6   141.873 0.571429    3/31/2001
8   216 0.714286    3/31/2001
9   254 0.857143    3/31/2001
1   -15.237 0.071429    6/30/2001
1   -32 0   6/30/2001
3   -6.949  0.285714    6/30/2001
5   6.307   0.428571    6/30/2001
6   28.542  0.571429    6/30/2001
7   140.816 0.642857    6/30/2001
9   239 0.857143    6/30/2001
1   -47 0.043478    9/30/2001
1   -63.367 0   9/30/2001
2   -16.599 0.130435    9/30/2001
4   -6.087  0.347826    9/30/2001
6   31.425  0.565217    9/30/2001
7   47.137  0.608696    9/30/2001
8   150.678 0.73913 9/30/2001
8   200 0.782609    9/30/2001
10  1902.684    0.956522    9/30/2001
1   -246.545    0   12/31/2001
2   -18.731 0.125   12/31/2001
4   -0.043  0.375   12/31/2001
4   -6  0.34375 12/31/2001
5   9.285   0.46875 12/31/2001
6   43.519  0.59375 12/31/2001
7   111 0.65625 12/31/2001
8   154.573 0.78125 12/31/2001
10  1017.514    0.9375  12/31/2001
1   -23.678 0.095238    3/31/2002
4   2.229   0.357143    3/31/2002
5   14  0.428571    3/31/2002
5   17.689  0.452381    3/31/2002
6   67.245  0.595238    3/31/2002
7   130.604 0.642857    3/31/2002
8   156 0.761905    3/31/2002
8   179.399 0.785714    3/31/2002
9   213.756 0.833333    3/31/2002
10  855.2   0.928571    3/31/2002
1   -26.536 0.076923    6/30/2002
3   1.295   0.288462    6/30/2002
4   9   0.365385    6/30/2002
5   16.714  0.423077    6/30/2002
6   64.547  0.557692    6/30/2002
6   103.539 0.596154    6/30/2002
8   181.284 0.769231    6/30/2002
9   203 0.807692    6/30/2002
10  600.194 0.923077    6/30/2002
10  284.306 0.903846    6/30/2002
1   -85 0.016129    9/30/2002
1   -25.475 0.096774    9/30/2002
2   -20.394 0.129032    9/30/2002
4   2.551   0.33871 9/30/2002
6   102.395 0.564516    9/30/2002
7   113.453 0.612903    9/30/2002
8   168.205 0.725806    9/30/2002
9   248 0.854839    9/30/2002
10  800.551 0.935484    9/30/2002
10  460.067 0.903226    9/30/2002

edit 3: As it stands, the way I am doing it here is so slow it is unusable. The slow part is the query with percent_rank() in it.

6
  • 1
    So what's wrong with the ntile() function?
    – user1822
    Commented Jun 29, 2017 at 20:10
  • I don't see how it could be implemented in this context, but I hope I am wrong. I clearly don't know a ton about this, but I don't know how to use ntile() using two different sets of data (dx.fund_val and fund_val) in the same function like percent_rank() can. And I assume that is necessary... Commented Jun 29, 2017 at 20:27
  • 1
    @mountainclimber your application requirements are not obvious from the question. It would really help if you provided a short sample input and sample output for it
    – filiprem
    Commented Jun 29, 2017 at 21:52
  • @filiprem doesn't edit 2 have all of that? frctl is my desired output and fund_val and gen_qtr_end_dt are my applicable inputs. Is there something else you need? Commented Jun 29, 2017 at 22:01
  • @mountainclimber what we want is basic schema, basic data, and a desired output. We prefer it to be separate from the complexities of your real world example. But, at the very least, we need sample data. Commented Jun 29, 2017 at 22:04

1 Answer 1

1

Try trunc(10 * pl.pctl) + 1, but as percent_rank returns 0 <= n <= 1 the maximum value will be 11 instead of 10.

CUME_DIST is quite similar to PERCENT_RANK but returns 0 < n <= 1, thus you might switch to 1-cume_dist... (ORDER BY fund_val DESC) for the pctl calculation.

1
  • I think in my case since I am always comparing the value to itself, as well as all its current period and previous periods, (gen_qtr_end_dt <= dx.gen_qtr_end_dt, note the =) I won't get a 1 out of percent_rank(); therefore, the 11 issue isn't really an issue for me. At least that is what I am seeing in my results (no 1s). Commented Jun 30, 2017 at 14:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.