Running average but reset when increasing streak stops - sql

Question

Original Dataset
+---------+--------+------------+
| Product | Date | Units Sold |
+---------+--------+------------+
| Prod A | 1/1/19 | 100 |
| Prod A | 1/2/19 | 200 |
| Prod A | 1/3/19 | 300 |
| Prod A | 1/4/19 | 136 |
| Prod A | 1/5/19 | 116 |
| Prod A | 1/6/19 | 120 |
| Prod A | 1/7/19 | 140 |
| Prod A | 1/8/19 | 160 |
+---------+--------+------------+
Desired Output (Last two columns)
+---------+--------+------------+---------------------+------+--------+--------------------+
| Product | Date | Units Sold | Previous Units Sold | Diff | Streak | Streak Running Avg |
+---------+--------+------------+---------------------+------+--------+--------------------+
| Prod A | 1/1/19 | 100 | | 100 | 1 | 100 |
| Prod A | 1/2/19 | 200 | 100 | 100 | 2 | 150 |
| Prod A | 1/3/19 | 300 | 200 | 100 | 3 | 200 |
| Prod A | 1/4/19 | 100 | 300 | -200 | 0 | 0 |
| Prod A | 1/5/19 | 200 | 100 | 100 | 1 | 200 |
| Prod A | 1/6/19 | 300 | 200 | 100 | 2 | 250 |
| Prod A | 1/7/19 | 200 | 300 | -100 | 0 | 0 |
| Prod A | 1/8/19 | 200 | 200 | 0 | 0 | 0 |
+---------+--------+------------+---------------------+------+--------+--------------------+
My goal is to calculate running average only when the difference between previous day sale and current day sale is positive (i.e. when the streak >0 calculate calculate the running average.)
For example, Day 1, 2, & 3 have diff positive i.e. 100
therefore the running average for
Day 1 is 100 (100/1)
Day 2 is 150 (100+200)/2
Day 3 is 200 (100+200+300)/3
 My Query
select *,
CASE WHEN flag=1 then sum(units) over (partition by item_name order by order_date) else 0 end as running_avg_sum
from
(select *,
lag(units, 1) over (partition by product order by order_date) as previous_day_units
CASE WHEN units > previous_day_units then 1 else 0 end as flag
from (
SELECT product,
order_date,
SUM(units_sold) units
FROM product_table
GROUP BY 1, 2
) tbl
)
But above query throws
Invalid operation: Aggregate window functions with an ORDER BY clause require a frame clause;
I'm aware how to resolve the error by adding rows between unbounded preceding and current row in the window function but that way it would average all preceding rows. I'm now sure how can I achieve the desired output.
If there's a way to define boundaries, that'd be really helpful
Any help is appreciated. Thank you!
Insert commands incase you want to replicate
CREATE TABLE product_table
(
product varchar(200),
order_date timestamp,
units_sold bigint
)
INSERT INTO product values("Prod A","2019-01-01","100");
INSERT INTO product values("Prod A","2019-01-02","200");
INSERT INTO product values("Prod A","2019-01-03","300");
INSERT INTO product values("Prod A","2019-01-04","136");
INSERT INTO product values("Prod A","2019-01-05","116");
INSERT INTO product values("Prod A","2019-01-06","120");
INSERT INTO product values("Prod A","2019-01-07","140");
INSERT INTO product values("Prod A","2019-01-08","160");

Answer 1

This is a somewhat tricky gaps and islands problem. Using the difference in row numbers method, we can try:
WITH cte1 AS (
SELECT *, LAG(units_sold, 1, 0) OVER (PARTITION BY product ORDER BY order_date) AS prev_units_sold,
units_sold - LAG(units_sold, 1, 0) OVER (PARTITION BY product ORDER BY order_date) diff
FROM product_table
),
cte2 AS (
SELECT *, ROW_NUMBER() OVER (PARTITION BY product ORDER BY order_date) rn1,
ROW_NUMBER() OVER (PARTITION BY product, diff > 0 ORDER BY order_date) rn2
FROM cte1
)
SELECT product, order_date, units_sold, prev_units_sold, diff,
CASE WHEN diff > 0
THEN ROW_NUMBER() OVER (PARTITION BY product, rn1 - rn2 ORDER BY order_date)
ELSE 0 END AS streak,
CASE WHEN diff > 0
THEN AVG(units_sold) OVER (PARTITION BY product, rn1 - rn2 ORDER BY order_date)
ELSE 0 END AS running_avg
FROM cte2
ORDER BY product, order_date;
Demo
The first pass CTE is required to compute the lag number of units sold, per product as ordered by order date. The second CTE computes two different row number sequences. The first one is simply ordered by order date. The second computes two separate sequences for positive diff and negative diff records. The difference between the first and second sequence can be used to identify the "islands" over which we want to display the streak and streak average. For the negative diff records, we just report zero for the streak and streak average.