Median in the moving window

You are correct that median() won't work. 

From the documentation:

MEDIAN() does not allow the window-order-clause or window-frame-clause.

You'd have to code something manually.  Statistically speaking, I'm not sure a rolling median would differ significantly from a rolling average, especially if you're only looking at 20 rows.  But then, I'm not a statistician.

Thanks for your response. If you think from trading perspective, 1 heavy day can easily skew your average, while median would be still in line with reality. Not sure, if it would be possible for Vertica to start supporting it, as I feel that it is a really useful feature (actually from the statistical perspective).

Check out nth_value - it might be able to give you what you want.

https://my.vertica.com/docs/7.2.x/HTML/index.htm#Authoring/SQLReferenceManual/Functions/Analytic/NTH_VALUEAnalytic.htm?Highlight=nth_value

Hello -
just picking up on that,
nth_value is a good idea but ordering the values in the window frame seems not be possible.
I guess it comes down to writing a function, or similar. There is a postgres equivalent in this post
I tried to translate into vertica syntax but not luck so far.
Anyone would have advice on how to implement a rolling median?
Thanks.

Hi Ralf_S,

You can do the trick involving a complementary self-join.

This is far from perfect I agree but still working:

  select t1.symbol, t1.date, t1.volume
       , approximate_median(t2.volume)
    from MyTable as t1
    join MyTable as t2  on t2.symbol  = t1.symbol
                       and t2.date   >= t1.date - interval '20 days'
                       and t2.date   <= t1.date
group by t1.symbol, t1.date, t1.volume
order by 1, 2;


  select distinct
         t1.symbol, t1.date, t1.volume
       , median(t2.volume) over(partition by t1.symbol, t1.date)
    from MyTable as t1
    join MyTable as t2  on t2.symbol  = t1.symbol
                       and t2.date   >= t1.date - interval '20 days'
                       and t2.date   <= t1.date
order by 1, 2;

Beware of the volume generated by the join.