Advanced forecasting models and techniques

[Former Member]

Is anyone familiar or have experience with advance forecasting models in APO (5.0) such as ARIMA, causal forecasting, Dynamic Regression, Multivariate ARIMA, Pooled and Panel Regression, and/or Fixed or Random Effects model? I think that an user exit can be used to employ some of these techniques. Does any have any experience implement external forecasting BADI? What are the drawbacks to using the BADI? How does these techniques fix with standard APO functionality? The definitions are below:

Dynamic Regression u2013 similar to typical regression analysis, however lagged values of independent variables are used as regressands (independent, or explanatory variables used to explain the variation in the dependent variable). A dynamic regression can predict what will happen to the dependent variable if the value of the explanatory variable changes, i.e. a change in price today will have this effect on volume of beer consumed next week.

ARIMA u2013 stand for AutoRegressive Integrated Moving Average models. These models are also sometimes referred to as the Box-Jenkins methodology. In ARIMA models the past values for a series of data are used to predict future values, in other words the only explanatory variables used in the model are past values of the dependent variable.

Multivariate ARIMA u2013 A combination of ARIMA and Dynamic Regression in which lagged values of the dependent variable as well as other independent variables are used to explain fluctuations in the dependent variable. One of the main benefits of this type of model is that contrary to other types of regression and econometric techniques where the maxim u201Ccorrelation does not imply causationu201D holds, this model allows for a specific test that can prove causation, e.g. Vector AutoRegression (VAR).

Pooled Regression u2013 Pooled regression is regression analysis applied to pooled data, which is the combination of data elements that are both time series data and cross-sectional data. This is the type of data that we typically use at MillerCoors, i.e. observations for multiple outlets (or distributors or Mgmt Units) or time (weekly or monthly).

Panel Regression u2013 Panel regression is regression analysis applied to panel data, which is a special type of pooled data in which the same cross-sectional unit is surveyed over time. Note: sometimes pooled and panel are used interchangeably and other names include micropanel data and longitudinal data.

Fixed Effects Model u2013 A method in which the slope coefficients for independent variables are constant across different cross-sectional units (i.e. distributors), but the intercept term varies for each of the cross-sectional units while the intercepts are constant over time (for each individual cross-sectional unit).

Random Effects Model u2013 Also referred to as the error components model, this technique is similar to the fixed effects model (actually the fixed effects model is a specific type of random effects model) where rather than each cross-sectional unit having a specific intercept they all are assumed to be drawn from a broader population that has a mean value of an intercept and each units intercept is expressed as the deviation from this mean intercept. The random effects model will be more efficient than the fixed effects model because it will have more degrees of freedom (resulting in greater confidence in the estimated parameters) and fewer regressands (a slight improvement in processing time).

Regards,

James

[Former Member]

James,

Bear with me it's going to be a long response. First, I have already asked this questions specifically at a meeting of the SAP Users group. The answer I received from SAP was that this "advanced" functionality was included on the road map for future releases of the 'merged' Business Objects/SAP product solutions.

Now, as a bit of background, I have exactly the same problem. At the moment, I use SPSS (not sure if you are familiar with the product) to run all those analysis on SAP and BW extracted data. Once the numbers are crunched outside SAP I then re-import the main parameters back into SAP. By the way, the response from SAP actually said that they were working with SPSS so the software would run side by side with SAP to provide this very functionality. Since them a few things have happened so I don't know what are the updated situation is.

Good news: SAP can handle most models - ARIMA is a problem, but depending on your particular circunstances either a Winters or Holt models will aprroximate close enough and those two (2) models are available on both APO and ECC 6.0. So, IF, and it's a BIG IF), you do have the parameters to run a particular model SAP will be able to handle your forecast within 5-10% variance on a 95% confidence level (that's pretty much my results).

Bad news - If you let SAP generate it's own parameters, in particular in situations such as yours, it goes horribly wrong, i.e., the parameter values are all over the place, mostly revert back to default values with does not provide an accurate forecast in any shape or form.

So, in short, SAP can handle 90% of the advanced forecast models and techniques you've described IF you have a detailed knowledge of each particular model. I understand this can be a particular problem (been there, suffered that). Once you have this particular piece of knowledge to implemetn the models is quite stratigh forward and I can show you how.

For disclaimer purposes, I'm a big fan of SPSS and I was quite please to learn that both (SPSS and SAP) were planning to work together.

[Former Member]

James,

Bear with me it's going to be a long response. First, I have already asked this questions specifically at a meeting of the SAP Users group. The answer I received from SAP was that this "advanced" functionality was included on the road map for future releases of the 'merged' Business Objects/SAP product solutions.

Now, as a bit of background, I have exactly the same problem. At the moment, I use SPSS (not sure if you are familiar with the product) to run all those analysis on SAP and BW extracted data. Once the numbers are crunched outside SAP I then re-import the main parameters back into SAP. By the way, the response from SAP actually said that they were working with SPSS so the software would run side by side with SAP to provide this very functionality. Since them a few things have happened so I don't know what are the updated situation is.

Good news: SAP can handle most models - ARIMA is a problem, but depending on your particular circunstances either a Winters or Holt models will aprroximate close enough and those two (2) models are available on both APO and ECC 6.0. So, IF, and it's a BIG IF), you do have the parameters to run a particular model SAP will be able to handle your forecast within 5-10% variance on a 95% confidence level (that's pretty much my results).

Bad news - If you let SAP generate it's own parameters, in particular in situations such as yours, it goes horribly wrong, i.e., the parameter values are all over the place, mostly revert back to default values with does not provide an accurate forecast in any shape or form.

So, in short, SAP can handle 90% of the advanced forecast models and techniques you've described IF you have a detailed knowledge of each particular model. I understand this can be a particular problem (been there, suffered that). Once you have this particular piece of knowledge to implemetn the models is quite stratigh forward and I can show you how.

For disclaimer purposes, I'm a big fan of SPSS and I was quite please to learn that both (SPSS and SAP) were planning to work together.

[william_getz]

Did you end up using the BADI or did you find another method. We too are finding greater needs in the causal analysis.

-Bill