Paramaters in Forecasting Algorithm

Question

Something simple enough I hope, I need to predict customers per day. My time series works, but the prediction doesn't seem right. I have a fairy big variation something like that:

16 October 2009	109
17 October 2009	107
18 October 2009	136
19 October 2009	124
20 October 2009	135
21 October 2009	106
22 October 2009	109

So you can see that the actual customer numbers varies a lot. However the prediction I get for future days doesn't have a big variation.  I get a figure around 128 and there is +3 - 3 for future dates and then stays at around 128 for December. But my real figures vary alot. I've tried to change the parameters and tried the periodicity_hint to to 7 (7 days of the week) as there is a strong correlation between days of the same week and lows or peaks.

 

Is anything I can do with the parameters to make my prediction better? At the moment is not good enough. I am actually not sure that my model understands that my input is a date series, is Key Time enough to denote that my first field is a date? Thank you for your help.

Panos

Answer 1

Hi,

Are you using Analysis Services 2005 or 2008? Are you doing the prediction in the Excel 2007 Data add ins or in BI-Dev Studio?

In AS 2008  you have the option to change the settings for the mix two algorithms(ARTXO and ARIMA) in the Enterprise Edition.

This is from the parameter description in the Excel 2007 data mining add ins.

"Specifies the forecasting algorithm to use. Select MIXED to create models for both ARTXP and ARIMA time series algorithms and combine their results during prediction. In Standard Edition, the models are combined using an automatic ratio that favors ARTXP for near-term and ARIMA for long-term prediction. In Enterprise Edition, the models are combined and weighted according to the value set for PREDICTION_SMOOTHING. If FORECAST_METHOD is set to ARTXP or ARIMA, the value for PREDICTION_SMOOTHING is ignored."

 HTH
Thomas Ivarsson

Answer 2

Hi,

I am only using the Data Mining within Visual Studio for SSAS 2005, we don't have office 2007 (and don't see us getting it  as we will need to retrain the secretaries to Office from scratch, sorry about the office 2007 banter, it's bad bad design)

I can see that there are additional options in 2008 - office 2007 but not an option for me. There are few parameters in the forecasting algorithm to amend, which may fine tune the foreacting, but to be fair this is a very unpredictable set of data we have.

Thanks anyway,

Panos

Answer 3

Are you showing all your data or just a small fraction of it?  If this is all your data, it's likely there isn't enough to do any forecasting and therefore is just predicting a mean value

---

-- Jamie MacLennan -- SQL Server Data Mining -- http://blogs.msdn.com/jamiemac

Answer 4

Hi,

No I have 2 years worth of data, and running it for the year. The graph is displayed for the last 6 months (in preview in VS) but assume it takes all the data as didn't put any restrictions.

Thanks,

Panos.