cancel
Showing results for 
Search instead for 
Did you mean: 

Historical Outlier Correction (IBP Forecast Preprocessing Steps)

brienne_tang
Explorer
0 Kudos

Dear All,

Is there any SAP recommendation on Historical Outlier Correction for IBP Forecast Pre-processing Steps?

Below is the simulation configurations i have tested with, but the confidence level of the output is still quite low:

  • Total product: 280 SKUs
  • Periodicity: Month
  • Historical Period: 12
  • Output analysis: Condidence level is based on acceptance when comparing to 6 Months Average Monthly Sales

Accepted Solutions (0)

Answers (1)

Answers (1)

LauraTozzo
Advisor
Advisor
0 Kudos

Hi Brienne,

you might want to test not only with different multipliers but also with different methods to identify the outliers and correct them. In particular correcting with mean instead of tolerance could change the results significantly.

Here you can find the documentation on all the different options available: https://help.sap.com/docs/SAP_INTEGRATED_BUSINESS_PLANNING/feae3cea3cc549aaa9d9de7d363a83e6/9a458154...

Best regards

Laura

rodney_nadasen
Explorer
0 Kudos

Dear Laura

I am doing similar testing of Outliers to determine which detection method I should select and then which correction method to use. I feel the i need more details regarding the detection methods and the benefits or pros and cons to each detection method i.e. IQR vs Variance tests - Can you provide more details to this question? Follow on question regarding the correction method: Can you provide a brief explanation on how the Tolerance method is working? Does SAP have a preferred detection and correct method and why this combination? This information would really be helpful? Thanks, Regards Rodney Nadasen

LauraTozzo
Advisor
Advisor

Hi Rodney,

here you can find some more details on the IQR and Variance Tests:
https://help.sap.com/docs/SAP_HANA_PLATFORM/2cfbc5cf2bc14f028cfbe2a2bba60a50/3e322ad3ac0a4057adf3674... https://help.sap.com/docs/SAP_HANA_PLATFORM/2cfbc5cf2bc14f028cfbe2a2bba60a50/08cb6731db504e80a4d509b...

In the documentation linked above is already clarified how the tolerance works: the values that fall outside of this tolerance lane are considered as outliers. The tolerance lane is computed based on the two methods above.
We do not have a preferred detection method, as this is very much dependent on the use case. I suggest to try them out and compare the results to see which one achieves better results for you.

Best regards

Laura

rodney_nadasen
Explorer
0 Kudos

Dear Laura

Thanks, this is very helpful.

I understand that there is "no" one size fits al however can you provide guidance on how i could reach a solution to best manage full portfolio of products? Ideally this is the case for me, i will need to determine 1 solution for full portfolio of products. I was thinking that i could backtest and use forecast error as a measure that could lead me to make a decision... do you know any other approach that i could use?

Thanks

Regards

Rodney

LauraTozzo
Advisor
Advisor
0 Kudos

Hi Rodney,

indeed using the forecast error as a measure is a very good approach. You could use a simple forecasting algorithm and compare different models where only the outlier correction has changed to see which yields the best results in terms of forecast error.

Best regards

Laura

rodney_nadasen
Explorer
0 Kudos

Hi Laura

Regarding the SAP documentation about the Time Series Analysis Correction Method for Seasonality and Trend – Can you describe what does it mean by saying “Outliers are changed” What is it changed to? Is there a calculation?

Also, does the Time Series Analysis also take into account the Change points? Meaning does the correction method calculate based on the new change point or does it use the full history that you have specified example 60 months?

Adjustment to Seasonality and Trend

Outliers are changed to fit the moving average calculated for the values in seasonality and trend patterns.

LauraTozzo
Advisor
Advisor

Hi Rodney,

as the documentation explains in the smoothing window section, if the time series shows seasonality or trend patterns and you choose to adjust to these, the algorithm calculates the moving average of the values that are forming these patterns.

If trend and seasonality are both found in the time series, the algorithm uses the length of the seasonal cycle as the smoothing window (calculated by the system, or specified manually in the Manage Forecast Automation Profiles app). If only trends are found in the time series, it uses the number that you define in for this setting. In both cases it computes the moving average based on this window and uses this to fix outliers.

Outlier correction at the moment does not take Change Points into account.

Best regards

Laura