none
Forecast time series for each key independently? RRS feed

  • Question

  • Hi all,

    I have a really simple data mining structure, it consists only of one key, one date column and one predict.

    So the key is for ClientId and predictable is for USD. The data itself is different for each client, so for example, Client-1 can have data in period between '2015-01-01' and '2016-05-01' by months, and Client-2 with data between '2015-01-01' and '2016-08-01'.

    My problem is that PredictTimeSeries function creates forecast for Client-1 starting from the latest non-empty point, ie '2016-08-01' and replaces all nulls by Mean value (I set it in config) between '2016-05-01' and '2016-08-01'. What I need is to make forecast for each client based on this clients last non-empty point. So correct result appears when I pass data only for a single client or when several clients have the same last point on date axis.
    Not sure, but this can also means that if Client-2 has values in past, sooner than Client-1 starts his orders, forecasting for Client-1 will replace all empty values again in past range. For me its unacceptable because forecasting losts a lot of accuracy. 

    What option or command can solve this problem?

    Thank you

    Tuesday, November 29, 2016 11:50 AM

All replies

  • Sounds like the data set you'll be feeding the SSAS mining structure will need additional manipulation. 

    What I normally do, is create a

    SELECT
    
    A.DATE,
    
    ,B.PK
    
    FROM
    
    (SELECT distinct DATE FROM DATE) AS A
    
    CROSSJOIN
    
    (SELECT distinct PK FROM PKTABLE) AS B


    You can left join back into your query and have a date range for each client.  

    You could build multiple mining structures where the clients end on the same date, such as one mining structure august 2016 etc.  I'm not sure how this will scale for you though, if you have hundreds of clients and tons of dates it probably won't work.  

    If it was a reasonable amount, you would then use an SSIS package to feed your individual results into reporting table where you could then build your dashboard, report.   If it was interleaved it would scale regardless of the date ranges.  

    Just some ideas.....  HTH


    Thursday, December 22, 2016 11:07 PM
  • I can't build multiple structures because I have a lot of clients so it's an unnecessary work and a copy-paste.
    P.S.: Currently we refused to use data mining in SSAS because of this and a few other reasons.
    Friday, December 23, 2016 1:40 PM