Top N values in window frame - postgresql

Question

I have a table t with 3 fields of interest:
d (date), pid (int), and score (numeric)
I am trying to calculate a 4th field that is an average of each player's top N (3 or 5) scores for the days before the current row.
I tried the following join on a subquery but it is not producing the results I'm looking for:
SELECT t.d, t.pid, t.score, sq.highscores
FROM t, (SELECT *, avg(score) as highscores FROM
(SELECT *, row_number() OVER w AS rnum
FROM t AS t2
WINDOW w AS (PARTITION BY pid ORDER BY score DESC ROWS BETWEEN UNBOUNDED PRECEDING AND 1 PRECEDING)) isq
WHERE rnum <= 3) sq
WHERE t.d = sq.d AND t.pid = sq.pid
Any suggestions would be greatly appreciated! I'm a hobbyist programmer and this is more complex of a query than I'm used to.

Answer 1

You can't select * and avg(score) in the same (inner) query. I.e. which non-aggregated values should be selected for each average? PostgreSQL won't decide this instead of you.
Becasue you PARTITION BY pid in the innermost query, you should use GROUP BY pid in the aggregating subquery. That way, you can SELECT pid, avg(score) as highscores:
SELECT pid, avg(score) as highscores
FROM (SELECT *, row_number() OVER w AS rnum
FROM t AS t2
WINDOW w AS (PARTITION BY pid ORDER BY score DESC)) isq
WHERE rnum <= 3
GROUP BY pid
Note: ROWS BETWEEN UNBOUNDED PRECEDING AND 1 PRECEDING makes no difference for row_number().
But if the top N part is fixed (and N will be few in your real-world use-case too), you can solve this without that much subquery (with the nth_value() window function):
SELECT d, pid, score,
(coalesce(nth_value(score, 1) OVER w, 0) +
coalesce(nth_value(score, 2) OVER w, 0) +
coalesce(nth_value(score, 3) OVER w, 0)) /
((nth_value(score, 1) OVER w IS NOT NULL)::int +
(nth_value(score, 2) OVER w IS NOT NULL)::int +
(nth_value(score, 3) OVER w IS NOT NULL)::int) highscores
FROM t
WINDOW w AS (PARTITION BY pid ORDER BY score DESC ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)
http://rextester.com/GUUPO5148