I've got a dataset that has id, start date and a claim value (in dollars) in each row - most ids have more than one row - some span over 50 rows. The earliest date for each ID/claim varies, and the claim values are mostly different.
I'd like to do a rolling sum of the value of IDs that have claims within 365 days of each other, to report each ID that has claims that have exceeded a limiting value across each period. So for an ID that had a claim date on 1 January, I'd sum all claims to 31 December (inclusive). Most IDs have several years of data so for the example above, I'd also need to check that if they had a claim on 1 May that they hadn't exceeded the limit by 30 April the following year and so on. I normally see this referred to as a 'rolling sum'. My site has many SAS products including base, stat, ets, and others.
I'm currently testing code on a small mock dataet and so far I've converted a thin file to a fat file with one column for each claim value and each date of the claim. The mock dataset is similar to the client dataset that I'll be using. Here's what I've done so far (noting that the mock data uses days rather than dates - I'm not at the stage where I want to test on real data yet).
data original_data;
input ppt $1. day claim;
datalines;
a 1 7
a 2 12
a 4 12
a 6 18
a 7 11
a 8 10
a 9 14
a 10 17
b 1 27
b 2 12
b 3 14
b 4 12
b 6 18
b 7 11
b 8 10
b 9 14
b 10 17
c 4 2
c 6 4
c 8 8
;
run;
proc sql;
create table ppt_counts as
select ppt, count(*) as ppts
from work.original_data
group by ppt;
select cats('value_', max(ppts) ) into :cats
from work.ppt_counts;
select cats('dates_',max(ppts)) into :cnts
from work.ppt_counts;
quit;
%put &cats;
%put &cnts;
data flipped;
set original_data;
by ppt;
array vars(*) value_1 -&cats.;
array dates(*) dates_1 - &cnts.;
array m_vars value_1 - &cats.;
array m_dates dates_1 - &cnts.;
if first.ppt then do;
i=1;
do over m_vars;
m_vars="";
end;
do over m_dates;
m_dates="";
end;
end;
if first.ppt then do:
i=1;
vars(i) = claim;
dates(i)=day;
if last.ppt then output;
i+1;
retain value_1 - &cats dates_1 - &cnts. 0.;
run;
data output;
set work.flipped;
max_date =max(of dates_1 - &cnts.);
max_value =max(of value_1 - &cats.);
run;
This doesn't give me even close to what I need - not sure how to structure code to make this correct.
What I need to end up with is one row per time that an ID exceeds the yearly limit of claim value (say in the mock data if a claim exceeds 75 across a seven day period), and to include the sum of the claims. So it's likely that there may be multiple lines per ID and the claims from one row may also be included in the claims for the same ID on another row.
type of output:
ID sum of claims
a $85
a $90
b $80
On separate rows.
Any help appreciated.
Thanks
If you need to perform a rolling sum, you can do this with proc expand. The code below will perform a rolling sum of 5 days for each group. First, expand your data to fill in any missing gaps:
proc expand data = original_data
out = original_data_expanded
from = day;
by ppt;
id day;
convert claim / method=none;
run;
Any days with gaps will have missing value of claim. Now we can calculate a moving sum and ignore those missing days when performing the moving sum:
proc expand data = original_data
out = want(where=(NOT missing(claim)));
by ppt;
id day;
convert claim = rolling_sum / transform=(movsum 5) method=none;
run;
Output:
ppt day rolling_sum claim
a 1 7 7
a 2 19 12
a 4 31 12
a 6 42 18
a 7 41 11
...
b 9 53 14
b 10 70 17
c 4 2 2
c 6 6 4
c 8 14 8
The reason we use two proc expand statements is because the rolling sum is calculated before the days are expanded. We need the rolling sum to occur after the expansion. You can test this by running the above code all in a single statement:
/* Performs moving sum, then expands */
proc expand data = original_data
out = test
from = day;
by ppt;
id day;
convert claim = rolling_sum / transform=(movsum 5) method=none;
run;
Use a SQL self join with the dates being within 365 days of itself. This is time/resource intensive if you have a very large data set.
Assuming you have a date variable, the intnx is probably the better way to calculate the date interval than 365 depending on how you want to account for leap years.
If you have a claim id to group on, that would also be better than using the group by clause in this example.
data have;
input ppt $1. day claim;
datalines;
a 1 7
a 2 12
a 4 12
a 6 18
a 7 11
a 8 10
a 9 14
a 10 17
b 1 27
b 2 12
b 3 14
b 4 12
b 6 18
b 7 11
b 8 10
b 9 14
b 10 17
c 4 2
c 6 4
c 8 8
;
run;
proc sql;
create table want as
select a.*, sum(b.claim) as total_claim
from have as a
left join have as b
on a.ppt=b.ppt and
b.day between a.day and a.day+365
group by 1, 2, 3;
/*b.day between a.day and intnx('year', a.day, 1, 's')*/;
quit;
Assuming that you have only one claim per day you could just use a circular array to keep track of the pervious N days of claims to generate the rolling sum. By circular array I mean one where the indexes wrap around back to the beginning when you increment past the end. You can use the MOD() function to convert any integer into an index into the array.
Then to get the running sum just add all of the elements in the array.
Add an extra DO loop to zero out the days skipped when there are days with no claims.
%let N=5;
data want;
set original_data;
by ppt ;
array claims[0:%eval(&n-1)] _temporary_;
lagday=lag(day);
if first.ppt then call missing(of lagday claims[*]);
do index=max(sum(lagday,1),day-&n+1) to day-1;
claims[mod(index,&n)]=0;
end;
claims[mod(day,&n)]=claim;
running_sum=sum(of claims[*]);
drop index lagday ;
run;
Results:
running_
OBS ppt day claim sum
1 a 1 7 7
2 a 2 12 19
3 a 4 12 31
4 a 6 18 42
5 a 7 11 41
6 a 8 10 51
7 a 9 14 53
8 a 10 17 70
9 b 1 27 27
10 b 2 12 39
11 b 3 14 53
12 b 4 12 65
13 b 6 18 56
14 b 7 11 55
15 b 8 10 51
16 b 9 14 53
17 b 10 17 70
18 c 4 2 2
19 c 6 4 6
20 c 8 8 14
Working in a known domain of date integers, you can use a single large array to store the claims at each date and slice out the 365 days to be summed. The bookkeeping needed for the modular approach is not needed.
Example:
data have;
call streaminit(20230202);
do id = 1 to 10;
do date = '01jan2012'd to '02feb2023'd;
date + rand('integer', 25);
claim = rand('integer', 5, 100);
output;
end;
end;
format date yymmdd10.;
run;
options fullstimer;
data want;
set have;
by id;
array claims(100000) _temporary_;
array slice (365) _temporary_;
if first.id then call missing(of claims(*));
claims(date) = claim;
call pokelong(
peekclong(
addrlong (claims(date-365))
, 8*365)
,
addrlong(slice(1))
);
rolling_sum_365 = sum(of slice(*));
if dif1(claim) < 365 then
claims_out_365 = lag(claim) - dif1(rolling_sum_365);
if first.id then claims_out_365 = .;
run;
Note: SAS Date 100,000 is 16OCT2233
I have a table which contains info on customer purchases per year and month respectively. Here is a simplified version.
id
year
month
nb_purch
1
2001
1
1
1
2001
2
4
1
2001
3
7
...
...
...
...
1
2001
12
3
1
2003
1
3
1
2003
2
2
1
2003
3
5
1
2003
4
7
...
...
...
...
1
2003
12
3
2
2001
1
3
2
2001
2
2
2
2001
3
5
2
2001
4
7
Basically there are several constraints. The database contains only the years when the client has made a purchase. If the client has made a purchase within the year X then X will be divided into 12 rows according to months. The months with no purchases have the value 0.
What I am trying to do is to retrieve the number of purchases per certain "windows". Currently its value sits at 3 years. For example i want to retrieve the sum of nb_purch within the last 3 years starting from 2003 march. This means i need to add all values from
march 2001 to march 2003.
SELECT SUM(nb_purch) OVER (PARTITION BY id ORDER BY year, month ASC ROWS BETWEEN 36 PRECEDING AND CURRENT ROW) AS LAST_3_YEARS FROM T
The issue i am facing here is that the table does not contain all years and therefore in my example of purchases between (2001 and 2003) if the year 2002 is missing then i am getting false results. I would like to avoid having to add all missing years and fill them with NULL values for each customer.
I'm trying to figure out if it's possible to transform table rows to columns where the number of rows included changes at the time of the query. Here's a sample of what I'm trying to do:
Characteristics Table
strategy
year
month
aaa
aa
a
InvestmentA
2020
12
5
4
10
InvestmentB
2020
12
8
15
25
Investment(n)
2020
12
x
x
x
Output
year
month
Credit Type
InvestmentA
InvestmentA
Investment(n)
2020
12
aaa
5
8
x
2020
12
aa
4
15
x
2020
12
a
10
25
x
I have a dataset that looks like this:
data have;
input ID P1 P2 P3 P4;
datalines;
ID P1 P2 P3 P4
12 10 15 20 30
12 - 20 5 3
12 - - 25 33
12 - - - 30
19 10 15 20 30
19 - 10 17 30
19 - - 5 30
19 - - - 30
;
run;
I am trying to build in a variable called Year which then can be used to identify that the ID and P1-P4 is an array with each row representing a year. Such that the dataset would look like.
data want;
set have;
input ID P1 P2 P3 P4;
datalines;
ID P1 P2 P3 P4 Year
12 10 15 20 30 2017
12 - 20 5 3 2018
12 - - 25 33 2019
12 - - - 30 2020
19 10 15 20 30 2017
19 - 10 17 30 2018
19 - - 5 30 2019
19 - - - 30 2020
;
run;
I originally used to use this code:
Data Year;
do ID = 1 to 8;
do Year = 2017 to 2020;
output;
end;
end;
run;
data Final;
set have;
Merge Year;
run;
But now that I am working with a different dataset each time and I don't know the structure of the ID, I can't keep changing ID=1 to 8 to suit the dataset each time.
My question: Is there a way to do this through the dataset, possibly a count?
Count ID = 2017;
Year = count + 1;
There is no need to create a second data set that will be merged with the first.
You do need to make assumptions about the grouping in the have data set. The assumptions are the data is already sorted or arranged in a manner that allows a monotonic year value to be assigned to each sequential row in each group.
data want;
set have;
by id;
if first.id
then year = 2017; %* initial year for a group;
else year + 1; %* increment year for subsequent rows of a group;
run;
I have a dataset as follows:
data have;
input;
ID Base Adverse Fixed$ Date RepricingFrequency
1 38 50 FIXED 2016 2
2 40 60 FLOATING 2017 3
3 20 20 FIXED 2016 2
4 ...
5
6
I am looking to build an array such that each ID has four vintage years 2017-2020, where the subsequent years are to be filled out with a piece of array code I have that works
like such
ID Vintage Base Adverse Fixed$ Date RepricingFrequency
1 2017 38 50 FIXED 2016 2
1 2018
1 2019
1 2020
In the beginning I just need to duplicate the dataset with the blanks,
The code I've tried so far is
data want;
set have;
do I=1 to 4;
output;
drop I;
run;
but of course that keeps the repeats of all the observations. So I tried an array.
data want;
set have;
array Base(2017:2020) Base2017-Base2020
array Vintage(2017:2020) Vintage2017-Vintage2020
But I'm not sure where to go from here on either accord.
The question is how do I extrapolate my data set for ID1-8 to a dataset where I have ID 1111-8888 where each ID is repeated 4 times with blanks.
Make a dummy dataset with all of the observations
data frame ;
set have(keep=id);
by id ;
if first.id then do date=2017 to 2020 ;
output;
end;
run;
and merge it back with the original.
data want ;
merge have frame ;
by id date ;
run;