A Possible Bug in the SQL Data Mining Time Series Algorithm RRS feed

  • Question

  • Hi, I am returning to this forum after a long time. This time with a serious bug (possible) I observed last week with the Time Series forecasting algorithm. I uses SQL Server 2008 Analysis Services (Enterprise Edition). The anomaly I observed is the following: In the Time Series module there is the option of choosing the underlaying algorithm (parameter FORECAST_METHOD) as either ARTXP, ARIMA or MIXED. If we choose MIXED, which is a combination of ARTXP and ARIMA there is another parameter PREDICTION_SMOOTHING using which one can control the relative weights of ARTXP and ARIMA. The PREDICTION_SMOOTHING parameter takes value between 0 and 1. According to the technical reference choosing 0 means 100% weight is given to ARTXP and choosing 1 means 100% weight is given to ARIMA. However when I choose MIXED option for FORECAST_METHOD and choose PREDICTION_SMOOTHING parameter as 0 I am not getting the same result as using ARTXP for FORECAST_METHOD on the same training data. Similarly the results for PREDICTION_SMOOTHING parameter as 1 is not same as ARIMA result. The difference is very significant, in some data points > 100%. Can some one explain this? This looks to me like some internal implementation bug. I can send you the data if you want to have a look.

    Thank You.


    Dr. Hari Koduvely

    Sunday, May 9, 2010 1:35 PM