There are many similar questions and answers already posted but I could not find one with these differences. 1) The count of NULLs starts over, and 2) there is a math function applied to the replaced value.

An event either takes place or not (NULL or 1), by date by customer. Can assume that a customer has one and only one row for every date.

I want to replace the NULLs with a decay function based on number of consecutive NULLs (time from event). A customer can have the event every day, skip a day, skip multiple days. But once the event takes place, the decay starts over. Currently my decay is divide by 2 but that is for example.

I can produce the desired result, but it is very unwieldy. Looking if there is a better way. This will need to be extended for multiple event columns.

create or replace temporary table the_data (
  dt date,
  customer char(10),
  event int,
  desired float)
;
insert into the_data values ('2022-01-01', 'a', 1, 1);
insert into the_data values ('2022-01-02', 'a', 1, 1);
insert into the_data values ('2022-01-03', 'a', 1, 1);
insert into the_data values ('2022-01-04', 'a', 1, 1);
insert into the_data values ('2022-01-05', 'a', 1, 1);

insert into the_data values ('2022-01-01', 'b', 1, 1);
insert into the_data values ('2022-01-02', 'b', NULL, 0.5);
insert into the_data values ('2022-01-03', 'b', NULL, 0.25);
insert into the_data values ('2022-01-04', 'b', 1, 1);
insert into the_data values ('2022-01-05', 'b', NULL, 0.5);

with
    base as (
      select * from the_data
    ),
    find_nan as (
      select *, case when event is null then 1 else 0 end as event_is_nan from base
    ),
    find_nan_diff as (
      select *, event_is_nan - coalesce(lag(event_is_nan) over (partition by customer order by dt), 0) as event_is_nan_diff from find_nan
    ),
    find_nan_group as (
      select *, sum(case when event_is_nan_diff = -1 then 1 else 0 end) over (partition by customer order by dt) as nan_group from find_nan_diff
    ),
    consec_nans as (
      select *, sum(event_is_nan) over (partition by customer, nan_group order by dt) as n_consec_nans from find_nan_group
    ),
    decay as (
      select *, case when n_consec_nans > 0 then 0.5 / n_consec_nans else 1 end as decay_factor from consec_nans
    ),
    ffill as (
      select *, first_value(event) over (partition by customer order by dt) as ffill_value from decay
    ),
    final as (
      select *, ffill_value * decay_factor as the_answer from ffill
    )
select * from final
order by customer, dt
;  

Thanks

CodePudding user response：

The query could be simplified by using