Rolling Aggregation - sql

Question

I am trying to write a program in SQL Server that aggregates based on rolling dates.
Take this below
Acc Dte Amount
1 1/1/20 100
1 1/3/20 200
1 1/8/20 100
1 1/8/20 75
2 1/1/20 50
2 1/2/20 100
2 1/3/20 75
2 1/3/20 125
3 1/3/20 100
3 1/6/20 75
3 1/8/20 75
3 1/10/20 200
3 1/10/20 150
So the goal is I want to find the avg and the count of records and dates for each account PRIOR TO the record being analyzed. I also need to sum the records based on the date So based on the above it would look like this...
Acc Dte Num_of_dates Avg_Amount_per_day Current_Amount
1 1/3/20 1 100 200
1 1/8/20 2 150 175
2 1/2/20 1 50 100
2 1/3/20 2 75 200
3 1/6/20 1 100 75
3 1/8/20 2 83.3 75
3 1/10/20 3 83.3 350
The goal is to create a z-score comparing accounts numbers for the current day to the accounts average per day. But we also need to hit a minimum of 10 days of historical data for each account.
Right now my code looks like this and is not working
select Account,
Dte,
(select sum(case when Cast(EventTimestamp as DATE) < Dte then 1 else 0 end) Num_of_Date,
(select (case when Cast(EventTimestamp as DATE) < Dte then sum(Amount) else 0 end) t_amount
from Data
group by Account, Dte
Any ideas? Thanks

Answer 1

You can use window functions with proper rows clause. For once, distinct comes handy here:
select distinct
acc,
dte,
count(*) over(
partition by acc
order by dte
rows between unbounded preceding and 1 preceding
) num_of_dates,
avg(1.0 * amount) over(
partition by acc
order by dte
rows between unbounded preceding and 1 preceding
) avg_amount_per_day,
sum(amount) over(partition by acc, dte) current_amount
from mytable
If you do want just one record per date and account, as shown in your sample data, you can nest the query and use row_number() - in absence of an obvious column to define the sorting order, I relied on the cumulative count:
select acc, dte, num_of_dates, avg_amount_per_day, current_amount
from (
select
t.*,
row_number() over(partition by acc, dte order by num_of_dates) rn
from (
select
acc,
dte,
count(*) over(
partition by acc
order by dte
rows between unbounded preceding and 1 preceding
) num_of_dates,
avg(1.0 * amount) over(
partition by acc
order by dte
rows between unbounded preceding and 1 preceding
) avg_amount_per_day,
sum(amount) over(partition by acc, dte) current_amount
from mytable
) t
) t
where rn = 1 and avg_amount_per_day is not null
Demo on DB Fiddlde:
acc | dte | num_of_dates | avg_amount_per_day | current_amount
--: | :--------- | -----------: | :----------------- | -------------:
1 | 2020-01-03 | 1 | 100.000000 | 200
1 | 2020-01-08 | 2 | 150.000000 | 175
2 | 2020-01-02 | 1 | 50.000000 | 100
2 | 2020-01-03 | 2 | 75.000000 | 200
3 | 2020-01-06 | 1 | 100.000000 | 75
3 | 2020-01-08 | 2 | 87.500000 | 75
3 | 2020-01-10 | 3 | 83.333333 | 350

Answer 2

Your sample data and description suggest:
select acc, dte,
count(*) as num_on_day,
sum(amount) as sum_on_day,
avg(sum(amount)) over (partition by acc order by date_num range between unbounded preceding and 1 preceding) as avg_previous
from t cross join
(values (datediff(day, '1900-01-01', dte))) v(date_num)
group by acc, dte;
I'm not sure why you don't include the first date for each acc.