Linear regression with postgres
This is the combination of Joop's statistics and Denis's window functions:
WITH num AS (
SELECT id, idstation
, (udate - '1984-01-01'::date) as idate -- count in dayse since jan 1984
, value AS value
FROM thedata
)
-- id + the ids of the {prev,next} records
-- within the same idstation group
, drag AS (
SELECT id AS center
, LAG(id) OVER www AS prev
, LEAD(id) OVER www AS next
FROM thedata
WINDOW www AS (partition by idstation ORDER BY id)
)
-- junction CTE between ID and its three feeders
, tri AS (
SELECT center AS this, center AS that FROM drag
UNION ALL SELECT center AS this , prev AS that FROM drag
UNION ALL SELECT center AS this , next AS that FROM drag
)
SELECT t.this, n.idstation
, regr_intercept(value,idate) AS intercept
, regr_slope(value,idate) AS slope
, regr_r2(value,idate) AS rsq
, regr_avgx(value,idate) AS avgx
, regr_avgy(value,idate) AS avgy
FROM num n
JOIN tri t ON t.that = n.id
GROUP BY t.this, n.idstation
;
Results:
INSERT 0 7
this | idstation | intercept | slope | rsq | avgx | avgy
------+-----------+-------------------+-------------------+-------------------+------------------+------------------
1 | 12 | -46 | 1 | 1 | 52 | 6
2 | 12 | -24.2105263157895 | 0.578947368421053 | 0.909774436090226 | 53.3333333333333 | 6.66666666666667
3 | 12 | -10.6666666666667 | 0.333333333333333 | 1 | 54.5 | 7.5
4 | 14 | | | | 51 | 9
5 | 15 | | | | 51 | 15
6 | 18 | | | | 51 | 14
7 | 19 | | | | 51 | 200
(7 rows)
The clustering of the group-of-three can probably be done more elegantly using a rank() or row_number() function, which would also allow larger sliding windows to be used.