Add statistical function to this?

Question

I have this SQL which aggregate quantity per week.
WITH  weeks_vw AS
    ( SELECT  DISTINCT TO_CHAR(fill_year_day,'IYYYIW') AS fill_year_week
            FROM
           ( SELECT  to_date('2012-11-01','yyyy-mm-dd')  + (ROWNUM-1) AS fill_year_day
                    FROM    DUAL
                    CONNECT BY      LEVEL <= TRUNC(to_date('2012-12-10','yyyy-mm-dd') - to_date('2012-11-01','yyyy-mm-dd') ) + 1)
SELECT  a.material, a.plant, sum(NVL(quantity,0)) AS sales_qty_sum
    FROM
SELECT  t.material, t.plant, t.quantity, TO_CHAR(t.posting_date,'IYYYIW') AS posting_week
         FROM    tblTotalCom t
         Where     t.plant = '1900'
         AND     t.Posting_date >= to_date('2012-11-01','yyyy-mm-dd')
         AND     t.Posting_date <= to_date('2012-12-10','yyyy-mm-dd')
         AND     t.material = 'Label1' ) A
    PARTITION BY (a.material,a.plant) RIGHT OUTER JOIN weeks_vw  ON weeks_vw.fill_year_week = A.posting_week
    GROUP BY  a.material, a.plant, fill_year_weekNow I want to add a statistical function as below in this SQL to remove outliers.
WITH     got_analytics     AS
     SELECT       deptno
     ,       ename
     ,       sal
     ,       AVG (sal)    OVER (PARTITION BY  deptno)     AS avg_sal
     ,       STDDEV (sal) OVER (PARTITION BY  deptno)     AS stddev_sal
     FROM       scott.emp
SELECT       a.*
,       CASE
           WHEN  ABS (sal - avg_sal)     > stddev_sal * 1.0   -- Change this number
           THEN  '*** Outlier ***'
       END     AS outlier
FROM       got_analytics  a
ORDER BY  deptno
,            sal
;Do not know how to do it ...?

Answer 1

Why did you answer that as a question? Are you not sure which approach you need? Is my question unclear? Obviously, it's hard to provide something that helps you if we're not sure exactly what you are trying to accomplish...
WITH  weeks_vw AS
    ( SELECT  DISTINCT TO_CHAR(fill_year_day,'IYYYIW') AS fill_year_week
            FROM
           ( SELECT  to_date('2012-11-01','yyyy-mm-dd')  + (ROWNUM-1) AS fill_year_day
                    FROM    DUAL
                    CONNECT BY      LEVEL <= TRUNC(to_date('2012-12-10','yyyy-mm-dd') - to_date('2012-11-01','yyyy-mm-dd') ) + 1)
SELECT material,
      plant,
      sales_qty_sum,
      avg(sales_qty_sum) over () avg_sum,
      stddev(sales_qty_sum) over () std_dev_sum,
      abs((sales_qty_sum - avg(sales_qty_sum) over ())/ stddev(sales_qty_sum) over ()) stddevs_away
  FROM (
      SELECT  a.material, a.plant, sum(NVL(quantity,0)) AS sales_qty_sum
          FROM
            SELECT  t.material, t.plant, t.quantity, TO_CHAR(t.posting_date,'IYYYIW') AS posting_week
              FROM    tblTotalCom t
             Where     t.plant = '1900'
              AND     t.Posting_date >= to_date('2012-11-01','yyyy-mm-dd')
              AND     t.Posting_date <= to_date('2012-12-10','yyyy-mm-dd')
              AND     t.material = 'Label1' ) A
       PARTITION BY (a.material,a.plant) RIGHT OUTER JOIN weeks_vw  ON weeks_vw.fill_year_week = A.posting_week
       GROUP BY  a.material, a.plant, fill_year_week)should compute the average sum, the standard deviation of the sum, and the number of standard deviations a particular row is from the mean. You can then filter out those rows that are more than 1 standard deviation from the mean
WITH  weeks_vw AS
    ( SELECT  DISTINCT TO_CHAR(fill_year_day,'IYYYIW') AS fill_year_week
            FROM
           ( SELECT  to_date('2012-11-01','yyyy-mm-dd')  + (ROWNUM-1) AS fill_year_day
                    FROM    DUAL
                    CONNECT BY      LEVEL <= TRUNC(to_date('2012-12-10','yyyy-mm-dd') - to_date('2012-11-01','yyyy-mm-dd') ) + 1)
SELECT *
  FROM (
SELECT material,
      plant,
      sales_qty_sum,
      avg(sales_qty_sum) over () avg_sum,
      stddev(sales_qty_sum) over () std_dev_sum,
      abs((sales_qty_sum - avg(sales_qty_sum) over ())/ stddev(sales_qty_sum) over ()) stddevs_away
  FROM (
      SELECT  a.material, a.plant, sum(NVL(quantity,0)) AS sales_qty_sum
          FROM
            SELECT  t.material, t.plant, t.quantity, TO_CHAR(t.posting_date,'IYYYIW') AS posting_week
              FROM    tblTotalCom t
             Where     t.plant = '1900'
              AND     t.Posting_date >= to_date('2012-11-01','yyyy-mm-dd')
              AND     t.Posting_date <= to_date('2012-12-10','yyyy-mm-dd')
              AND     t.material = 'Label1' ) A
       PARTITION BY (a.material,a.plant) RIGHT OUTER JOIN weeks_vw  ON weeks_vw.fill_year_week = A.posting_week
       GROUP BY  a.material, a.plant, fill_year_week))
WHERE stddevs_away <= 1Note that this will generally filter out ~33% of your data if you assume a normal distribution so this is a rather odd definition of "outlier". I believe in the earlier question you asked, it was suggested that you would probably want to pick cutoff of more than 1 standard deviation
Also, it would be very helpful if you posted the DDL to create the table, the DML to populate some sample data, and the expected results. Obviously, since I don't have your tables, I can't test the queries I posted which makes it very possible that I introduced an inadvertent syntax error.
Justin
Edited by: Justin Cave on Dec 11, 2012 5:15 PM