Complex Query Involving Search for Contiguous Dates (by Month) - sql

Question

I have a table that contains a list of accounts by month along with a field that indicates activity. I want to search through to find when an account has "died", based on the following criteria:
the account had consistent activity for a contiguous period of months
the account had a spike of activity on a final month (spike = 200% or more of average of all previous contiguous months of activity)
the month immediately following the spike of activity and the next 12 months all had 0 activity
So the table might look something like this:
ID | Date | Activity
1 | 1/1/2010 | 2
2 | 1/1/2010 | 3.2
1 | 2/3/2010 | 3
2 | 2/3/2010 | 2.7
1 | 3/2/2010 | 8
2 | 3/2/2010 | 9
1 | 4/6/2010 | 0
2 | 4/6/2010 | 0
1 | 5/2/2010 | 0
2 | 5/2/2010 | 2
So in this case both accounts 1 and 2 have activity in months Jan - Mar. Both accounts exhibit a spike of activity in March. Both accounts have 0 activity in April. Account 2 has activity again in May, but account 1 does not. Therefore, my query should return Account 1, but not Account 2. I would want to see this as my query result:
ID | Last Date
1 | 3/2/2010
I realize this is a complicated question and I'm not expecting anyone to write the whole query for me. The current best approach I can think of is to create a series of sub-queries and join them, but I don't even know what the subqueries would look like. For example: how do I look for a contiguous series of rows for a single ID where activity is all 0 (or all non-zero?).
My fall-back if the SQL is simply too involved is to use a brute-force search using Java where I would first find all unique IDs, and then for each unique ID iterate across the months to determine if and when the ID "died".
Once again: any help to move in the right direction is very much appreciated.

Answer 1

Processing in Java, or partially processing in SQL, and finishing the processing in Java is a good approach.
I'm not going to tackle how to define a spike.
I will suggest that you start with condition 3. It's easy to find the last non-zero value. Then that's the one you want to test for a spike, and consistant data before the spike.
SELECT out.*
FROM monthly_activity out
LEFT OUTER JOIN monthly_activity comp
ON out.ID = comp.ID AND out.Date < comp.Date AND comp.Activity <> 0
WHERE comp.Date IS NULL
Not bad, but you don't want the result if this is because the record is the last for the month, so instead,
SELECT out.*
FROM monthly_activity out
INNER JOIN monthly_activity comp
ON out.ID = comp.ID AND out.Date < comp.Date AND comp.Activity == 0
GROUP BY out.ID

Answer 2

Probably not the world's most efficient code, but I think this does what you're after:
declare #t table (AccountId int, ActivityDate date, Activity float)
insert #t
select 1, '2010-01-01', 2
union select 2, '2010-01-01', 3.2
union select 1, '2010-02-03', 3
union select 2, '2010-02-03', 2.7
union select 1, '2010-03-02', 8
union select 2, '2010-03-02', 9
union select 1, '2010-04-06', 0
union select 2, '2010-04-06', 0
union select 1, '2010-05-02', 0
union select 2, '2010-05-02', 2
select AccountId, ActivityDate LastActivityDate --, Activity
from #t a
where
--Part 2 --select only where the activity is a peak
Activity >= isnull
(
(
select 2 * avg(c.Activity)
from #t c
where c.AccountId = 1
and c.ActivityDate >= isnull
(
(
select max(d.ActivityDate)
from #t d
where d.AccountId = c.AccountId
and d.ActivityDate < c.ActivityDate
and d.Activity = 0
)
,
(
select min(e.ActivityDate)
from #t e
where e.AccountId = c.AccountId
)
)
and c.ActivityDate < a.ActivityDate
)
, Activity + 1 --Part 1 (i.e. if no activity before today don't include the result)
)
--Part 3
and not exists --select only dates which have had no activity for the following 12 months on the same account (assumption: count no record as no activity / also ignore current date in this assumption)
(
select 1
from #t b
where a.AccountId = b.AccountId
and b.Activity > 0
and b.ActivityDate between dateadd(DAY, 1, a.ActivityDate) and dateadd(YEAR, 1, a.ActivityDate)
)